v1

activations.py

+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+
+import torch
+import torch.nn.functional as F
+from packaging import version
+
+from utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+def _gelu_python(x):
+    """
+    Original Implementation of the GELU activation function in Google BERT repo when initially created. For
+    information: OpenAI GPT's GELU is slightly different (and gives slightly different results): 0.5 * x * (1 +
+    torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3)))) This is now written in C in
+    torch.nn.functional Also see the Gaussian Error Linear Units paper: https://arxiv.org/abs/1606.08415
+    """
+    return x * 0.5 * (1.0 + torch.erf(x / math.sqrt(2.0)))
+
+
+def gelu_new(x):
+    """
+    Implementation of the GELU activation function currently in Google BERT repo (identical to OpenAI GPT). Also see
+    the Gaussian Error Linear Units paper: https://arxiv.org/abs/1606.08415
+    """
+    return 0.5 * x * (1.0 + torch.tanh(math.sqrt(2.0 / math.pi) * (x + 0.044715 * torch.pow(x, 3.0))))
+
+
+if version.parse(torch.__version__) < version.parse("1.4"):
+    gelu = _gelu_python
+else:
+    gelu = F.gelu
+
+
+def gelu_fast(x):
+    return 0.5 * x * (1.0 + torch.tanh(x * 0.7978845608 * (1.0 + 0.044715 * x * x)))
+
+
+def _silu_python(x):
+    """
+    See Gaussian Error Linear Units (Hendrycks et al., https://arxiv.org/abs/1606.08415) where the SiLU (Sigmoid Linear
+    Unit) was originally introduced and coined, and see Sigmoid-Weighted Linear Units for Neural Network Function
+    Approximation in Reinforcement Learning (Elfwing et al., https://arxiv.org/abs/1702.03118) and Swish: a Self-Gated
+    Activation Function (Ramachandran et al., https://arxiv.org/abs/1710.05941v1) where the SiLU was experimented with
+    later.
+    """
+    return x * torch.sigmoid(x)
+
+
+if version.parse(torch.__version__) < version.parse("1.7"):
+    silu = _silu_python
+else:
+    silu = F.silu
+
+
+def mish(x):
+    return x * torch.tanh(torch.nn.functional.softplus(x))
+
+
+def linear_act(x):
+    return x
+
+
+ACT2FN = {
+    "relu": F.relu,
+    "silu": silu,
+    "swish": silu,
+    "gelu": gelu,
+    "tanh": torch.tanh,
+    "gelu_new": gelu_new,
+    "gelu_fast": gelu_fast,
+    "mish": mish,
+    "linear": linear_act,
+    "sigmoid": torch.sigmoid,
+}
+
+
+def get_activation(activation_string):
+    if activation_string in ACT2FN:
+        return ACT2FN[activation_string]
+    else:
+        raise KeyError("function {} not found in ACT2FN mapping {}".format(activation_string, list(ACT2FN.keys())))

config/config.json

+{
+  "architectures": [
+    "BertForMaskedLM"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-12,
+  "max_position_embeddings": 512,
+  "model_type": "bert",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "pad_token_id": 0,
+  "type_vocab_size": 2,
+  "vocab_size": 23615
+}

configuration_bert.py

+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" BERT model configuration """
+
+from configuration_utils import PretrainedConfig
+from file_utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+BERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "bert-base-uncased": "https://huggingface.co/bert-base-uncased/resolve/main/config.json",
+    "bert-large-uncased": "https://huggingface.co/bert-large-uncased/resolve/main/config.json",
+    "bert-base-cased": "https://huggingface.co/bert-base-cased/resolve/main/config.json",
+    "bert-large-cased": "https://huggingface.co/bert-large-cased/resolve/main/config.json",
+    "bert-base-multilingual-uncased": "https://huggingface.co/bert-base-multilingual-uncased/resolve/main/config.json",
+    "bert-base-multilingual-cased": "https://huggingface.co/bert-base-multilingual-cased/resolve/main/config.json",
+    "bert-base-chinese": "https://huggingface.co/bert-base-chinese/resolve/main/config.json",
+    "bert-base-german-cased": "https://huggingface.co/bert-base-german-cased/resolve/main/config.json",
+    "bert-large-uncased-whole-word-masking": "https://huggingface.co/bert-large-uncased-whole-word-masking/resolve/main/config.json",
+    "bert-large-cased-whole-word-masking": "https://huggingface.co/bert-large-cased-whole-word-masking/resolve/main/config.json",
+    "bert-large-uncased-whole-word-masking-finetuned-squad": "https://huggingface.co/bert-large-uncased-whole-word-masking-finetuned-squad/resolve/main/config.json",
+    "bert-large-cased-whole-word-masking-finetuned-squad": "https://huggingface.co/bert-large-cased-whole-word-masking-finetuned-squad/resolve/main/config.json",
+    "bert-base-cased-finetuned-mrpc": "https://huggingface.co/bert-base-cased-finetuned-mrpc/resolve/main/config.json",
+    "bert-base-german-dbmdz-cased": "https://huggingface.co/bert-base-german-dbmdz-cased/resolve/main/config.json",
+    "bert-base-german-dbmdz-uncased": "https://huggingface.co/bert-base-german-dbmdz-uncased/resolve/main/config.json",
+    "cl-tohoku/bert-base-japanese": "https://huggingface.co/cl-tohoku/bert-base-japanese/resolve/main/config.json",
+    "cl-tohoku/bert-base-japanese-whole-word-masking": "https://huggingface.co/cl-tohoku/bert-base-japanese-whole-word-masking/resolve/main/config.json",
+    "cl-tohoku/bert-base-japanese-char": "https://huggingface.co/cl-tohoku/bert-base-japanese-char/resolve/main/config.json",
+    "cl-tohoku/bert-base-japanese-char-whole-word-masking": "https://huggingface.co/cl-tohoku/bert-base-japanese-char-whole-word-masking/resolve/main/config.json",
+    "TurkuNLP/bert-base-finnish-cased-v1": "https://huggingface.co/TurkuNLP/bert-base-finnish-cased-v1/resolve/main/config.json",
+    "TurkuNLP/bert-base-finnish-uncased-v1": "https://huggingface.co/TurkuNLP/bert-base-finnish-uncased-v1/resolve/main/config.json",
+    "wietsedv/bert-base-dutch-cased": "https://huggingface.co/wietsedv/bert-base-dutch-cased/resolve/main/config.json",
+    # See all BERT models at https://huggingface.co/models?filter=bert
+}
+
+
+class BertConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a :class:`~transformers.BertModel` or a
+    :class:`~transformers.TFBertModel`. It is used to instantiate a BERT model according to the specified arguments,
+    defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration
+    to that of the BERT `bert-base-uncased <https://huggingface.co/bert-base-uncased>`__ architecture.
+
+    Configuration objects inherit from :class:`~transformers.PretrainedConfig` and can be used to control the model
+    outputs. Read the documentation from :class:`~transformers.PretrainedConfig` for more information.
+
+
+    Args:
+        vocab_size (:obj:`int`, `optional`, defaults to 30522):
+            Vocabulary size of the BERT model. Defines the number of different tokens that can be represented by the
+            :obj:`inputs_ids` passed when calling :class:`~transformers.BertModel` or
+            :class:`~transformers.TFBertModel`.
+        hidden_size (:obj:`int`, `optional`, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (:obj:`int`, `optional`, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (:obj:`int`, `optional`, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (:obj:`int`, `optional`, defaults to 3072):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
+        hidden_act (:obj:`str` or :obj:`Callable`, `optional`, defaults to :obj:`"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string,
+            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
+        hidden_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):
+            The dropout ratio for the attention probabilities.
+        max_position_embeddings (:obj:`int`, `optional`, defaults to 512):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        type_vocab_size (:obj:`int`, `optional`, defaults to 2):
+            The vocabulary size of the :obj:`token_type_ids` passed when calling :class:`~transformers.BertModel` or
+            :class:`~transformers.TFBertModel`.
+        initializer_range (:obj:`float`, `optional`, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (:obj:`float`, `optional`, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+        gradient_checkpointing (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            If True, use gradient checkpointing to save memory at the expense of slower backward pass.
+        position_embedding_type (:obj:`str`, `optional`, defaults to :obj:`"absolute"`):
+            Type of position embedding. Choose one of :obj:`"absolute"`, :obj:`"relative_key"`,
+            :obj:`"relative_key_query"`. For positional embeddings use :obj:`"absolute"`. For more information on
+            :obj:`"relative_key"`, please refer to `Self-Attention with Relative Position Representations (Shaw et al.)
+            <https://arxiv.org/abs/1803.02155>`__. For more information on :obj:`"relative_key_query"`, please refer to
+            `Method 4` in `Improve Transformer Models with Better Relative Position Embeddings (Huang et al.)
+            <https://arxiv.org/abs/2009.13658>`__.
+        use_cache (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if ``config.is_decoder=True``.
+
+    Examples::
+
+        >>> from transformers import BertModel, BertConfig
+
+        >>> # Initializing a BERT bert-base-uncased style configuration
+        >>> configuration = BertConfig()
+
+        >>> # Initializing a model from the bert-base-uncased style configuration
+        >>> model = BertModel(configuration)
+
+        >>> # Accessing the model configuration
+        >>> configuration = model.config
+    """
+    model_type = "bert"
+
+    def __init__(
+        self,
+        vocab_size=30522,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        intermediate_size=3072,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        type_vocab_size=2,
+        initializer_range=0.02,
+        layer_norm_eps=1e-12,
+        pad_token_id=0,
+        gradient_checkpointing=False,
+        position_embedding_type="absolute",
+        use_cache=True,
+        **kwargs
+    ):
+        super().__init__(pad_token_id=pad_token_id, **kwargs)
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.gradient_checkpointing = gradient_checkpointing
+        self.position_embedding_type = position_embedding_type
+        self.use_cache = use_cache

configuration_utils.py

+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Configuration base class and utilities."""
+
+
+import copy
+import json
+import os
+from typing import Any, Dict, Tuple, Union
+
+from file_utils import CONFIG_NAME, cached_path, hf_bucket_url, is_offline_mode, is_remote_url
+from utils import logging
+
+__version__ = "4.4.2"
+logger = logging.get_logger(__name__)
+
+
+class PretrainedConfig(object):
+    r"""
+    Base class for all configuration classes. Handles a few parameters common to all models' configurations as well as
+    methods for loading/downloading/saving configurations.
+
+    Note: A configuration file can be loaded and saved to disk. Loading the configuration file and using this file to
+    initialize a model does **not** load the model weights. It only affects the model's configuration.
+
+    Class attributes (overridden by derived classes)
+
+        - **model_type** (:obj:`str`): An identifier for the model type, serialized into the JSON file, and used to
+          recreate the correct object in :class:`~transformers.AutoConfig`.
+        - **is_composition** (:obj:`bool`): Whether the config class is composed of multiple sub-configs. In this case
+          the config has to be initialized from two or more configs of type :class:`~transformers.PretrainedConfig`
+          like: :class:`~transformers.EncoderDecoderConfig` or :class:`~RagConfig`.
+        - **keys_to_ignore_at_inference** (:obj:`List[str]`): A list of keys to ignore by default when looking at
+          dictionary outputs of the model during inference.
+
+    Args:
+        name_or_path (:obj:`str`, `optional`, defaults to :obj:`""`):
+            Store the string that was passed to :func:`~transformers.PreTrainedModel.from_pretrained` or
+            :func:`~transformers.TFPreTrainedModel.from_pretrained` as ``pretrained_model_name_or_path`` if the
+            configuration was created with such a method.
+        output_hidden_states (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether or not the model should return all hidden-states.
+        output_attentions (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether or not the model should returns all attentions.
+        return_dict (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether or not the model should return a :class:`~transformers.file_utils.ModelOutput` instead of a plain
+            tuple.
+        is_encoder_decoder (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether the model is used as an encoder/decoder or not.
+        is_decoder (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether the model is used as decoder or not (in which case it's used as an encoder).
+        add_cross_attention (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether cross-attention layers should be added to the model. Note, this option is only relevant for models
+            that can be used as decoder models within the `:class:~transformers.EncoderDecoderModel` class, which
+            consists of all models in ``AUTO_MODELS_FOR_CAUSAL_LM``.
+        tie_encoder_decoder (:obj:`bool`, `optional`, defaults to :obj:`False`)
+            Whether all encoder weights should be tied to their equivalent decoder weights. This requires the encoder
+            and decoder model to have the exact same parameter names.
+        prune_heads (:obj:`Dict[int, List[int]]`, `optional`, defaults to :obj:`{}`):
+            Pruned heads of the model. The keys are the selected layer indices and the associated values, the list of
+            heads to prune in said layer.
+
+            For instance ``{1: [0, 2], 2: [2, 3]}`` will prune heads 0 and 2 on layer 1 and heads 2 and 3 on layer 2.
+        chunk_size_feed_forward (:obj:`int`, `optional`, defaults to :obj:`0`):
+            The chunk size of all feed forward layers in the residual attention blocks. A chunk size of :obj:`0` means
+            that the feed forward layer is not chunked. A chunk size of n means that the feed forward layer processes
+            :obj:`n` < sequence_length embeddings at a time. For more information on feed forward chunking, see `How
+            does Feed Forward Chunking work? <../glossary.html#feed-forward-chunking>`__ .
+
+    Parameters for sequence generation
+
+        - **max_length** (:obj:`int`, `optional`, defaults to 20) -- Maximum length that will be used by default in the
+          :obj:`generate` method of the model.
+        - **min_length** (:obj:`int`, `optional`, defaults to 10) -- Minimum length that will be used by default in the
+          :obj:`generate` method of the model.
+        - **do_sample** (:obj:`bool`, `optional`, defaults to :obj:`False`) -- Flag that will be used by default in the
+          :obj:`generate` method of the model. Whether or not to use sampling ; use greedy decoding otherwise.
+        - **early_stopping** (:obj:`bool`, `optional`, defaults to :obj:`False`) -- Flag that will be used by default
+          in the :obj:`generate` method of the model. Whether to stop the beam search when at least ``num_beams``
+          sentences are finished per batch or not.
+        - **num_beams** (:obj:`int`, `optional`, defaults to 1) -- Number of beams for beam search that will be used by
+          default in the :obj:`generate` method of the model. 1 means no beam search.
+        - **num_beam_groups** (:obj:`int`, `optional`, defaults to 1) -- Number of groups to divide :obj:`num_beams`
+          into in order to ensure diversity among different groups of beams that will be used by default in the
+          :obj:`generate` method of the model. 1 means no group beam search.
+        - **diversity_penalty** (:obj:`float`, `optional`, defaults to 0.0) -- Value to control diversity for group
+          beam search. that will be used by default in the :obj:`generate` method of the model. 0 means no diversity
+          penalty. The higher the penalty, the more diverse are the outputs.
+        - **temperature** (:obj:`float`, `optional`, defaults to 1) -- The value used to module the next token
+          probabilities that will be used by default in the :obj:`generate` method of the model. Must be strictly
+          positive.
+        - **top_k** (:obj:`int`, `optional`, defaults to 50) -- Number of highest probability vocabulary tokens to keep
+          for top-k-filtering that will be used by default in the :obj:`generate` method of the model.
+        - **top_p** (:obj:`float`, `optional`, defaults to 1) -- Value that will be used by default in the
+          :obj:`generate` method of the model for ``top_p``. If set to float < 1, only the most probable tokens with
+          probabilities that add up to ``top_p`` or higher are kept for generation.
+        - **repetition_penalty** (:obj:`float`, `optional`, defaults to 1) -- Parameter for repetition penalty that
+          will be used by default in the :obj:`generate` method of the model. 1.0 means no penalty.
+        - **length_penalty** (:obj:`float`, `optional`, defaults to 1) -- Exponential penalty to the length that will
+          be used by default in the :obj:`generate` method of the model.
+        - **no_repeat_ngram_size** (:obj:`int`, `optional`, defaults to 0) -- Value that will be used by default in the
+          :obj:`generate` method of the model for ``no_repeat_ngram_size``. If set to int > 0, all ngrams of that size
+          can only occur once.
+        - **encoder_no_repeat_ngram_size** (:obj:`int`, `optional`, defaults to 0) -- Value that will be used by
+          default in the :obj:`generate` method of the model for ``encoder_no_repeat_ngram_size``. If set to int > 0,
+          all ngrams of that size that occur in the ``encoder_input_ids`` cannot occur in the ``decoder_input_ids``.
+        - **bad_words_ids** (:obj:`List[int]`, `optional`) -- List of token ids that are not allowed to be generated
+          that will be used by default in the :obj:`generate` method of the model. In order to get the tokens of the
+          words that should not appear in the generated text, use :obj:`tokenizer.encode(bad_word,
+          add_prefix_space=True)`.
+        - **num_return_sequences** (:obj:`int`, `optional`, defaults to 1) -- Number of independently computed returned
+          sequences for each element in the batch that will be used by default in the :obj:`generate` method of the
+          model.
+        - **output_scores** (:obj:`bool`, `optional`, defaults to :obj:`False`) -- Whether the model should return the
+          logits when used for generation
+        - **return_dict_in_generate** (:obj:`bool`, `optional`, defaults to :obj:`False`) -- Whether the model should
+          return a :class:`~transformers.file_utils.ModelOutput` instead of a :obj:`torch.LongTensor`
+        - **forced_bos_token_id** (:obj:`int`, `optional`) -- The id of the token to force as the first generated token
+          after the :obj:`decoder_start_token_id`. Useful for multilingual models like :doc:`mBART
+          <../model_doc/mbart>` where the first generated token needs to be the target language token.
+        - **forced_eos_token_id** (:obj:`int`, `optional`) -- The id of the token to force as the last generated token
+          when :obj:`max_length` is reached.
+
+
+    Parameters for fine-tuning tasks
+
+        - **architectures** (:obj:`List[str]`, `optional`) -- Model architectures that can be used with the model
+          pretrained weights.
+        - **finetuning_task** (:obj:`str`, `optional`) -- Name of the task used to fine-tune the model. This can be
+          used when converting from an original (TensorFlow or PyTorch) checkpoint.
+        - **id2label** (:obj:`Dict[int, str]`, `optional`) -- A map from index (for instance prediction index, or
+          target index) to label.
+        - **label2id** (:obj:`Dict[str, int]`, `optional`) -- A map from label to index for the model.
+        - **num_labels** (:obj:`int`, `optional`) -- Number of labels to use in the last layer added to the model,
+          typically for a classification task.
+        - **task_specific_params** (:obj:`Dict[str, Any]`, `optional`) -- Additional keyword arguments to store for the
+          current task.
+
+    Parameters linked to the tokenizer
+
+        - **tokenizer_class** (:obj:`str`, `optional`) -- The name of the associated tokenizer class to use (if none is
+          set, will use the tokenizer associated to the model by default).
+        - **prefix** (:obj:`str`, `optional`) -- A specific prompt that should be added at the beginning of each text
+          before calling the model.
+        - **bos_token_id** (:obj:`int`, `optional`)) -- The id of the `beginning-of-stream` token.
+        - **pad_token_id** (:obj:`int`, `optional`)) -- The id of the `padding` token.
+        - **eos_token_id** (:obj:`int`, `optional`)) -- The id of the `end-of-stream` token.
+        - **decoder_start_token_id** (:obj:`int`, `optional`)) -- If an encoder-decoder model starts decoding with a
+          different token than `bos`, the id of that token.
+        - **sep_token_id** (:obj:`int`, `optional`)) -- The id of the `separation` token.
+
+    PyTorch specific parameters
+
+        - **torchscript** (:obj:`bool`, `optional`, defaults to :obj:`False`) -- Whether or not the model should be
+          used with Torchscript.
+        - **tie_word_embeddings** (:obj:`bool`, `optional`, defaults to :obj:`True`) -- Whether the model's input and
+          output word embeddings should be tied. Note that this is only relevant if the model has a output word
+          embedding layer.
+
+    TensorFlow specific parameters
+
+        - **use_bfloat16** (:obj:`bool`, `optional`, defaults to :obj:`False`) -- Whether or not the model should use
+          BFloat16 scalars (only used by some TensorFlow models).
+    """
+    model_type: str = ""
+    is_composition: bool = False
+
+    def __init__(self, **kwargs):
+        # Attributes with defaults
+        self.return_dict = kwargs.pop("return_dict", True)
+        self.output_hidden_states = kwargs.pop("output_hidden_states", False)
+        self.output_attentions = kwargs.pop("output_attentions", False)
+        self.torchscript = kwargs.pop("torchscript", False)  # Only used by PyTorch models
+        self.use_bfloat16 = kwargs.pop("use_bfloat16", False)
+        self.pruned_heads = kwargs.pop("pruned_heads", {})
+        self.tie_word_embeddings = kwargs.pop(
+            "tie_word_embeddings", True
+        )  # Whether input and output word embeddings should be tied for all MLM, LM and Seq2Seq models.
+
+        # Is decoder is used in encoder-decoder models to differentiate encoder from decoder
+        self.is_encoder_decoder = kwargs.pop("is_encoder_decoder", False)
+        self.is_decoder = kwargs.pop("is_decoder", False)
+        self.add_cross_attention = kwargs.pop("add_cross_attention", False)
+        self.tie_encoder_decoder = kwargs.pop("tie_encoder_decoder", False)
+
+        # Parameters for sequence generation
+        self.max_length = kwargs.pop("max_length", 20)
+        self.min_length = kwargs.pop("min_length", 0)
+        self.do_sample = kwargs.pop("do_sample", False)
+        self.early_stopping = kwargs.pop("early_stopping", False)
+        self.num_beams = kwargs.pop("num_beams", 1)
+        self.num_beam_groups = kwargs.pop("num_beam_groups", 1)
+        self.diversity_penalty = kwargs.pop("diversity_penalty", 0.0)
+        self.temperature = kwargs.pop("temperature", 1.0)
+        self.top_k = kwargs.pop("top_k", 50)
+        self.top_p = kwargs.pop("top_p", 1.0)
+        self.repetition_penalty = kwargs.pop("repetition_penalty", 1.0)
+        self.length_penalty = kwargs.pop("length_penalty", 1.0)
+        self.no_repeat_ngram_size = kwargs.pop("no_repeat_ngram_size", 0)
+        self.encoder_no_repeat_ngram_size = kwargs.pop("encoder_no_repeat_ngram_size", 0)
+        self.bad_words_ids = kwargs.pop("bad_words_ids", None)
+        self.num_return_sequences = kwargs.pop("num_return_sequences", 1)
+        self.chunk_size_feed_forward = kwargs.pop("chunk_size_feed_forward", 0)
+        self.output_scores = kwargs.pop("output_scores", False)
+        self.return_dict_in_generate = kwargs.pop("return_dict_in_generate", False)
+        self.forced_bos_token_id = kwargs.pop("forced_bos_token_id", None)
+        self.forced_eos_token_id = kwargs.pop("forced_eos_token_id", None)
+
+        # Fine-tuning task arguments
+        self.architectures = kwargs.pop("architectures", None)
+        self.finetuning_task = kwargs.pop("finetuning_task", None)
+        self.id2label = kwargs.pop("id2label", None)
+        self.label2id = kwargs.pop("label2id", None)
+        if self.id2label is not None:
+            kwargs.pop("num_labels", None)
+            self.id2label = dict((int(key), value) for key, value in self.id2label.items())
+            # Keys are always strings in JSON so convert ids to int here.
+        else:
+            self.num_labels = kwargs.pop("num_labels", 2)
+
+        # Tokenizer arguments TODO: eventually tokenizer and models should share the same config
+        self.tokenizer_class = kwargs.pop("tokenizer_class", None)
+        self.prefix = kwargs.pop("prefix", None)
+        self.bos_token_id = kwargs.pop("bos_token_id", None)
+        self.pad_token_id = kwargs.pop("pad_token_id", None)
+        self.eos_token_id = kwargs.pop("eos_token_id", None)
+        self.sep_token_id = kwargs.pop("sep_token_id", None)
+
+        self.decoder_start_token_id = kwargs.pop("decoder_start_token_id", None)
+
+        # task specific arguments
+        self.task_specific_params = kwargs.pop("task_specific_params", None)
+
+        # TPU arguments
+        if kwargs.pop("xla_device", None) is not None:
+            logger.warn(
+                "The `xla_device` argument has been deprecated in v4.4.0 of Transformers. It is ignored and you can "
+                "safely remove it from your `config.json` file."
+            )
+
+        # Name or path to the pretrained checkpoint
+        self._name_or_path = str(kwargs.pop("name_or_path", ""))
+
+        # Drop the transformers version info
+        kwargs.pop("transformers_version", None)
+
+        # Additional attributes without default values
+        for key, value in kwargs.items():
+            try:
+                setattr(self, key, value)
+            except AttributeError as err:
+                logger.error("Can't set {} with value {} for {}".format(key, value, self))
+                raise err
+
+    @property
+    def name_or_path(self) -> str:
+        return self._name_or_path
+
+    @name_or_path.setter
+    def name_or_path(self, value):
+        self._name_or_path = str(value)  # Make sure that name_or_path is a string (for JSON encoding)
+
+    @property
+    def use_return_dict(self) -> bool:
+        """
+        :obj:`bool`: Whether or not return :class:`~transformers.file_utils.ModelOutput` instead of tuples.
+        """
+        # If torchscript is set, force `return_dict=False` to avoid jit errors
+        return self.return_dict and not self.torchscript
+
+    @property
+    def num_labels(self) -> int:
+        """
+        :obj:`int`: The number of labels for classification models.
+        """
+        return len(self.id2label)
+
+    @num_labels.setter
+    def num_labels(self, num_labels: int):
+        if self.id2label is None or len(self.id2label) != num_labels:
+            self.id2label = {i: "LABEL_{}".format(i) for i in range(num_labels)}
+            self.label2id = dict(zip(self.id2label.values(), self.id2label.keys()))
+
+    def save_pretrained(self, save_directory: Union[str, os.PathLike]):
+        """
+        Save a configuration object to the directory ``save_directory``, so that it can be re-loaded using the
+        :func:`~transformers.PretrainedConfig.from_pretrained` class method.
+
+        Args:
+            save_directory (:obj:`str` or :obj:`os.PathLike`):
+                Directory where the configuration JSON file will be saved (will be created if it does not exist).
+        """
+        if os.path.isfile(save_directory):
+            raise AssertionError("Provided path ({}) should be a directory, not a file".format(save_directory))
+        os.makedirs(save_directory, exist_ok=True)
+        # If we save using the predefined names, we can load using `from_pretrained`
+        output_config_file = os.path.join(save_directory, CONFIG_NAME)
+
+        self.to_json_file(output_config_file, use_diff=True)
+        logger.info(f"Configuration saved in {output_config_file}")
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        r"""
+        Instantiate a :class:`~transformers.PretrainedConfig` (or a derived class) from a pretrained model
+        configuration.
+
+        Args:
+            pretrained_model_name_or_path (:obj:`str` or :obj:`os.PathLike`):
+                This can be either:
+
+                - a string, the `model id` of a pretrained model configuration hosted inside a model repo on
+                  huggingface.co. Valid model ids can be located at the root-level, like ``bert-base-uncased``, or
+                  namespaced under a user or organization name, like ``dbmdz/bert-base-german-cased``.
+                - a path to a `directory` containing a configuration file saved using the
+                  :func:`~transformers.PretrainedConfig.save_pretrained` method, e.g., ``./my_model_directory/``.
+                - a path or url to a saved configuration JSON `file`, e.g.,
+                  ``./my_model_directory/configuration.json``.
+            cache_dir (:obj:`str` or :obj:`os.PathLike`, `optional`):
+                Path to a directory in which a downloaded pretrained model configuration should be cached if the
+                standard cache should not be used.
+            force_download (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Whether or not to force to (re-)download the configuration files and override the cached versions if
+                they exist.
+            resume_download (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Whether or not to delete incompletely received file. Attempts to resume the download if such a file
+                exists.
+            proxies (:obj:`Dict[str, str]`, `optional`):
+                A dictionary of proxy servers to use by protocol or endpoint, e.g., :obj:`{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}.` The proxies are used on each request.
+            use_auth_token (:obj:`str` or `bool`, `optional`):
+                The token to use as HTTP bearer authorization for remote files. If :obj:`True`, will use the token
+                generated when running :obj:`transformers-cli login` (stored in :obj:`~/.huggingface`).
+            revision(:obj:`str`, `optional`, defaults to :obj:`"main"`):
+                The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
+                git-based system for storing models and other artifacts on huggingface.co, so ``revision`` can be any
+                identifier allowed by git.
+            return_unused_kwargs (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                If :obj:`False`, then this function returns just the final configuration object.
+
+                If :obj:`True`, then this functions returns a :obj:`Tuple(config, unused_kwargs)` where `unused_kwargs`
+                is a dictionary consisting of the key/value pairs whose keys are not configuration attributes: i.e.,
+                the part of ``kwargs`` which has not been used to update ``config`` and is otherwise ignored.
+            kwargs (:obj:`Dict[str, Any]`, `optional`):
+                The values in kwargs of any keys which are configuration attributes will be used to override the loaded
+                values. Behavior concerning key/value pairs whose keys are *not* configuration attributes is controlled
+                by the ``return_unused_kwargs`` keyword parameter.
+
+        .. note::
+
+            Passing :obj:`use_auth_token=True` is required when you want to use a private model.
+
+
+        Returns:
+            :class:`PretrainedConfig`: The configuration object instantiated from this pretrained model.
+
+        Examples::
+
+            # We can't instantiate directly the base class `PretrainedConfig` so let's show the examples on a
+            # derived class: BertConfig
+            config = BertConfig.from_pretrained('bert-base-uncased')    # Download configuration from huggingface.co and cache.
+            config = BertConfig.from_pretrained('./test/saved_model/')  # E.g. config (or model) was saved using `save_pretrained('./test/saved_model/')`
+            config = BertConfig.from_pretrained('./test/saved_model/my_configuration.json')
+            config = BertConfig.from_pretrained('bert-base-uncased', output_attentions=True, foo=False)
+            assert config.output_attentions == True
+            config, unused_kwargs = BertConfig.from_pretrained('bert-base-uncased', output_attentions=True,
+                                                               foo=False, return_unused_kwargs=True)
+            assert config.output_attentions == True
+            assert unused_kwargs == {'foo': False}
+
+        """
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+        return cls.from_dict(config_dict, **kwargs)
+
+    @classmethod
+    def get_config_dict(
+        cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs
+    ) -> Tuple[Dict[str, Any], Dict[str, Any]]:
+        """
+        From a ``pretrained_model_name_or_path``, resolve to a dictionary of parameters, to be used for instantiating a
+        :class:`~transformers.PretrainedConfig` using ``from_dict``.
+
+
+
+        Parameters:
+            pretrained_model_name_or_path (:obj:`str` or :obj:`os.PathLike`):
+                The identifier of the pre-trained checkpoint from which we want the dictionary of parameters.
+
+        Returns:
+            :obj:`Tuple[Dict, Dict]`: The dictionary(ies) that will be used to instantiate the configuration object.
+
+        """
+        cache_dir = kwargs.pop("cache_dir", None)
+        force_download = kwargs.pop("force_download", False)
+        resume_download = kwargs.pop("resume_download", False)
+        proxies = kwargs.pop("proxies", None)
+        use_auth_token = kwargs.pop("use_auth_token", None)
+        local_files_only = kwargs.pop("local_files_only", False)
+        revision = kwargs.pop("revision", None)
+
+        if is_offline_mode() and not local_files_only:
+            logger.info("Offline mode: forcing local_files_only=True")
+            local_files_only = True
+
+        pretrained_model_name_or_path = str(pretrained_model_name_or_path)
+        if os.path.isdir(pretrained_model_name_or_path):
+            config_file = os.path.join(pretrained_model_name_or_path, CONFIG_NAME)
+        elif os.path.isfile(pretrained_model_name_or_path) or is_remote_url(pretrained_model_name_or_path):
+            config_file = pretrained_model_name_or_path
+        else:
+            config_file = hf_bucket_url(
+                pretrained_model_name_or_path, filename=CONFIG_NAME, revision=revision, mirror=None
+            )
+
+        try:
+            # Load from URL or cache if already cached
+            resolved_config_file = cached_path(
+                config_file,
+                cache_dir=cache_dir,
+                force_download=force_download,
+                proxies=proxies,
+                resume_download=resume_download,
+                local_files_only=local_files_only,
+                use_auth_token=use_auth_token,
+            )
+            # Load config dict
+            config_dict = cls._dict_from_json_file(resolved_config_file)
+
+        except EnvironmentError as err:
+            logger.error(err)
+            msg = (
+                f"Can't load config for '{pretrained_model_name_or_path}'. Make sure that:\n\n"
+                f"- '{pretrained_model_name_or_path}' is a correct model identifier listed on 'https://huggingface.co/models'\n\n"
+                f"- or '{pretrained_model_name_or_path}' is the correct path to a directory containing a {CONFIG_NAME} file\n\n"
+            )
+            raise EnvironmentError(msg)
+
+        except json.JSONDecodeError:
+            msg = (
+                "Couldn't reach server at '{}' to download configuration file or "
+                "configuration file is not a valid JSON file. "
+                "Please check network or file content here: {}.".format(config_file, resolved_config_file)
+            )
+            raise EnvironmentError(msg)
+
+        if resolved_config_file == config_file:
+            logger.info("loading configuration file {}".format(config_file))
+        else:
+            logger.info("loading configuration file {} from cache at {}".format(config_file, resolved_config_file))
+
+        return config_dict, kwargs
+
+    @classmethod
+    def from_dict(cls, config_dict: Dict[str, Any], **kwargs) -> "PretrainedConfig":
+        """
+        Instantiates a :class:`~transformers.PretrainedConfig` from a Python dictionary of parameters.
+
+        Args:
+            config_dict (:obj:`Dict[str, Any]`):
+                Dictionary that will be used to instantiate the configuration object. Such a dictionary can be
+                retrieved from a pretrained checkpoint by leveraging the
+                :func:`~transformers.PretrainedConfig.get_config_dict` method.
+            kwargs (:obj:`Dict[str, Any]`):
+                Additional parameters from which to initialize the configuration object.
+
+        Returns:
+            :class:`PretrainedConfig`: The configuration object instantiated from those parameters.
+        """
+        return_unused_kwargs = kwargs.pop("return_unused_kwargs", False)
+
+        config = cls(**config_dict)
+
+        if hasattr(config, "pruned_heads"):
+            config.pruned_heads = dict((int(key), value) for key, value in config.pruned_heads.items())
+
+        # Update config with kwargs if needed
+        to_remove = []
+        for key, value in kwargs.items():
+            if hasattr(config, key):
+                setattr(config, key, value)
+                to_remove.append(key)
+        for key in to_remove:
+            kwargs.pop(key, None)
+
+        logger.info("Model config %s", str(config))
+        if return_unused_kwargs:
+            return config, kwargs
+        else:
+            return config
+
+    @classmethod
+    def from_json_file(cls, json_file: Union[str, os.PathLike]) -> "PretrainedConfig":
+        """
+        Instantiates a :class:`~transformers.PretrainedConfig` from the path to a JSON file of parameters.
+
+        Args:
+            json_file (:obj:`str` or :obj:`os.PathLike`):
+                Path to the JSON file containing the parameters.
+
+        Returns:
+            :class:`PretrainedConfig`: The configuration object instantiated from that JSON file.
+
+        """
+        config_dict = cls._dict_from_json_file(json_file)
+        return cls(**config_dict)
+
+    @classmethod
+    def _dict_from_json_file(cls, json_file: Union[str, os.PathLike]):
+        with open(json_file, "r", encoding="utf-8") as reader:
+            text = reader.read()
+        return json.loads(text)
+
+    def __eq__(self, other):
+        return self.__dict__ == other.__dict__
+
+    def __repr__(self):
+        return "{} {}".format(self.__class__.__name__, self.to_json_string())
+
+    def to_diff_dict(self) -> Dict[str, Any]:
+        """
+        Removes all attributes from config which correspond to the default config attributes for better readability and
+        serializes to a Python dictionary.
+
+        Returns:
+            :obj:`Dict[str, Any]`: Dictionary of all the attributes that make up this configuration instance,
+        """
+        config_dict = self.to_dict()
+
+        # get the default config dict
+        default_config_dict = PretrainedConfig().to_dict()
+
+        # get class specific config dict
+        class_config_dict = self.__class__().to_dict() if not self.is_composition else {}
+
+        serializable_config_dict = {}
+
+        # only serialize values that differ from the default config
+        for key, value in config_dict.items():
+            if (
+                key not in default_config_dict
+                or key == "transformers_version"
+                or value != default_config_dict[key]
+                or (key in class_config_dict and value != class_config_dict[key])
+            ):
+                serializable_config_dict[key] = value
+
+        return serializable_config_dict
+
+    def to_dict(self) -> Dict[str, Any]:
+        """
+        Serializes this instance to a Python dictionary.
+
+        Returns:
+            :obj:`Dict[str, Any]`: Dictionary of all the attributes that make up this configuration instance.
+        """
+        output = copy.deepcopy(self.__dict__)
+        if hasattr(self.__class__, "model_type"):
+            output["model_type"] = self.__class__.model_type
+
+        # Transformers version when serializing the model
+        output["transformers_version"] = __version__
+
+        return output
+
+    def to_json_string(self, use_diff: bool = True) -> str:
+        """
+        Serializes this instance to a JSON string.
+
+        Args:
+            use_diff (:obj:`bool`, `optional`, defaults to :obj:`True`):
+                If set to ``True``, only the difference between the config instance and the default
+                ``PretrainedConfig()`` is serialized to JSON string.
+
+        Returns:
+            :obj:`str`: String containing all the attributes that make up this configuration instance in JSON format.
+        """
+        if use_diff is True:
+            config_dict = self.to_diff_dict()
+        else:
+            config_dict = self.to_dict()
+        return json.dumps(config_dict, indent=2, sort_keys=True) + "\n"
+
+    def to_json_file(self, json_file_path: Union[str, os.PathLike], use_diff: bool = True):
+        """
+        Save this instance to a JSON file.
+
+        Args:
+            json_file_path (:obj:`str` or :obj:`os.PathLike`):
+                Path to the JSON file in which this configuration instance's parameters will be saved.
+            use_diff (:obj:`bool`, `optional`, defaults to :obj:`True`):
+                If set to ``True``, only the difference between the config instance and the default
+                ``PretrainedConfig()`` is serialized to JSON file.
+        """
+        with open(json_file_path, "w", encoding="utf-8") as writer:
+            writer.write(self.to_json_string(use_diff=use_diff))
+
+    def update(self, config_dict: Dict[str, Any]):
+        """
+        Updates attributes of this class with attributes from ``config_dict``.
+
+        Args:
+            config_dict (:obj:`Dict[str, Any]`): Dictionary of attributes that shall be updated for this class.
+        """
+        for key, value in config_dict.items():
+            setattr(self, key, value)

dataset.py

@@ -110,15 +110,17 @@ def seq2emb_encoder(input_seq, max_len, vocab):
 class Data_Encoder(data.Dataset):
     def __init__(self, train_file, tokenizer_config):
         'Initialization'
-        self.begin_id = train_file["begin_id"]
-        self.sep_id = train_file["separate_id"]
-        self.max_len = train_file["max_len"]
+        # load data
         with open(train_file["sps"], 'r') as f:
             self.sps = f.readlines()
         with open(train_file["smile"], 'r') as f:
             self.smile = f.readlines()
         with open(train_file["affinity"], 'r') as f:
             self.affinity = f.readlines()
+        # define tokenizer
+        self.begin_id = tokenizer_config["begin_id"]
+        self.sep_id = tokenizer_config["separate_id"]
+        self.max_len = tokenizer_config["max_len"]
         self.vocab = load_vocab(tokenizer_config["vocab_file"])
         bpe_codes_drug = codecs.open(tokenizer_config["vocab_pair"])
         self.dbpe = BPE(bpe_codes_drug, merges=-1, separator='')
@@ -158,17 +160,16 @@ if __name__ == "__main__":
     df_train = {"sps": './data/train_sps',
                 "smile": './data/train_smile',
                 "affinity": './data/train_ic50',
-                "vocab_file": './config/vocab.txt',
-                "begin_id": '[CLS]',
-                "separate_id": "[SEP]",
-                "max_len": 256
                 }
 
     tokenizer_config = {"vocab_file": './config/vocab.txt',
-                        "vocab_pair": './config/drug_codes_chembl.txt'
+                        "vocab_pair": './config/drug_codes_chembl.txt',
+                        "begin_id": '[CLS]',
+                        "separate_id": "[SEP]",
+                        "max_len": 256
                         }
     params = {'batch_size': 5,
-              'shuffle': False,
+              'shuffle': True,
               'num_workers': 0,
               'drop_last': True}
     trainset = Data_Encoder(df_train, tokenizer_config)

file_utils.py

+# Copyright 2020 The HuggingFace Team, the AllenNLP library authors. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Utilities for working with the local dataset cache. Parts of this file is adapted from the AllenNLP library at
+https://github.com/allenai/allennlp.
+"""
+
+import copy
+import fnmatch
+import importlib.util
+import io
+import json
+import os
+import re
+import shutil
+import sys
+import tarfile
+import tempfile
+from collections import OrderedDict, UserDict
+from contextlib import contextmanager
+from dataclasses import fields
+from enum import Enum
+from functools import partial, wraps
+from hashlib import sha256
+from pathlib import Path
+from types import ModuleType
+from typing import Any, BinaryIO, Dict, List, Optional, Tuple, Union
+from urllib.parse import urlparse
+from zipfile import ZipFile, is_zipfile
+
+import numpy as np
+from packaging import version
+from tqdm.auto import tqdm
+
+import requests
+from filelock import FileLock
+
+__version__ = '4.4.2'
+from utils.hf_api import HfFolder
+from utils import logging
+
+
+# The package importlib_metadata is in a different place, depending on the python version.
+if sys.version_info < (3, 8):
+    import importlib_metadata
+else:
+    import importlib.metadata as importlib_metadata
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+ENV_VARS_TRUE_VALUES = {"1", "ON", "YES", "TRUE"}
+ENV_VARS_TRUE_AND_AUTO_VALUES = ENV_VARS_TRUE_VALUES.union({"AUTO"})
+
+USE_TF = os.environ.get("USE_TF", "AUTO").upper()
+USE_TORCH = os.environ.get("USE_TORCH", "AUTO").upper()
+USE_JAX = os.environ.get("USE_FLAX", "AUTO").upper()
+
+if USE_TORCH in ENV_VARS_TRUE_AND_AUTO_VALUES and USE_TF not in ENV_VARS_TRUE_VALUES:
+    _torch_available = importlib.util.find_spec("torch") is not None
+    if _torch_available:
+        try:
+            _torch_version = importlib_metadata.version("torch")
+            logger.info(f"PyTorch version {_torch_version} available.")
+        except importlib_metadata.PackageNotFoundError:
+            _torch_available = False
+else:
+    logger.info("Disabling PyTorch because USE_TF is set")
+    _torch_available = False
+
+
+if USE_TF in ENV_VARS_TRUE_AND_AUTO_VALUES and USE_TORCH not in ENV_VARS_TRUE_VALUES:
+    _tf_available = importlib.util.find_spec("tensorflow") is not None
+    if _tf_available:
+        # For the metadata, we have to look for both tensorflow and tensorflow-cpu
+        try:
+            _tf_version = importlib_metadata.version("tensorflow")
+        except importlib_metadata.PackageNotFoundError:
+            try:
+                _tf_version = importlib_metadata.version("tensorflow-cpu")
+            except importlib_metadata.PackageNotFoundError:
+                try:
+                    _tf_version = importlib_metadata.version("tensorflow-gpu")
+                except importlib_metadata.PackageNotFoundError:
+                    try:
+                        _tf_version = importlib_metadata.version("tf-nightly")
+                    except importlib_metadata.PackageNotFoundError:
+                        try:
+                            _tf_version = importlib_metadata.version("tf-nightly-cpu")
+                        except importlib_metadata.PackageNotFoundError:
+                            try:
+                                _tf_version = importlib_metadata.version("tf-nightly-gpu")
+                            except importlib_metadata.PackageNotFoundError:
+                                # Support for intel-tensorflow version
+                                try:
+                                    _tf_version = importlib_metadata.version("intel-tensorflow")
+                                except importlib_metadata.PackageNotFoundError:
+                                    _tf_version = None
+                                    _tf_available = False
+    if _tf_available:
+        if version.parse(_tf_version) < version.parse("2"):
+            logger.info(f"TensorFlow found but with version {_tf_version}. Transformers requires version 2 minimum.")
+            _tf_available = False
+        else:
+            logger.info(f"TensorFlow version {_tf_version} available.")
+else:
+    logger.info("Disabling Tensorflow because USE_TORCH is set")
+    _tf_available = False
+
+
+if USE_JAX in ENV_VARS_TRUE_AND_AUTO_VALUES:
+    _flax_available = importlib.util.find_spec("jax") is not None and importlib.util.find_spec("flax") is not None
+    if _flax_available:
+        try:
+            _jax_version = importlib_metadata.version("jax")
+            _flax_version = importlib_metadata.version("flax")
+            logger.info(f"JAX version {_jax_version}, Flax version {_flax_version} available.")
+        except importlib_metadata.PackageNotFoundError:
+            _flax_available = False
+else:
+    _flax_available = False
+
+
+_datasets_available = importlib.util.find_spec("datasets") is not None
+try:
+    # Check we're not importing a "datasets" directory somewhere but the actual library by trying to grab the version
+    # AND checking it has an author field in the metadata that is HuggingFace.
+    _ = importlib_metadata.version("datasets")
+    _datasets_metadata = importlib_metadata.metadata("datasets")
+    if _datasets_metadata.get("author", "") != "HuggingFace Inc.":
+        _datasets_available = False
+except importlib_metadata.PackageNotFoundError:
+    _datasets_available = False
+
+
+_faiss_available = importlib.util.find_spec("faiss") is not None
+try:
+    _faiss_version = importlib_metadata.version("faiss")
+    logger.debug(f"Successfully imported faiss version {_faiss_version}")
+except importlib_metadata.PackageNotFoundError:
+    try:
+        _faiss_version = importlib_metadata.version("faiss-cpu")
+        logger.debug(f"Successfully imported faiss version {_faiss_version}")
+    except importlib_metadata.PackageNotFoundError:
+        _faiss_available = False
+
+
+_onnx_available = (
+    importlib.util.find_spec("keras2onnx") is not None and importlib.util.find_spec("onnxruntime") is not None
+)
+try:
+    _onxx_version = importlib_metadata.version("onnx")
+    logger.debug(f"Successfully imported onnx version {_onxx_version}")
+except importlib_metadata.PackageNotFoundError:
+    _onnx_available = False
+
+
+_scatter_available = importlib.util.find_spec("torch_scatter") is not None
+try:
+    _scatter_version = importlib_metadata.version("torch_scatter")
+    logger.debug(f"Successfully imported torch-scatter version {_scatter_version}")
+except importlib_metadata.PackageNotFoundError:
+    _scatter_available = False
+
+
+_soundfile_available = importlib.util.find_spec("soundfile") is not None
+try:
+    _soundfile_version = importlib_metadata.version("soundfile")
+    logger.debug(f"Successfully imported soundfile version {_soundfile_version}")
+except importlib_metadata.PackageNotFoundError:
+    _soundfile_available = False
+
+_torchaudio_available = importlib.util.find_spec("torchaudio")
+try:
+    _torchaudio_version = importlib_metadata.version("torchaudio")
+    logger.debug(f"Successfully imported soundfile version {_torchaudio_version}")
+except importlib_metadata.PackageNotFoundError:
+    _torchaudio_available = False
+
+
+torch_cache_home = os.getenv("TORCH_HOME", os.path.join(os.getenv("XDG_CACHE_HOME", "~/.cache"), "torch"))
+old_default_cache_path = os.path.join(torch_cache_home, "transformers")
+# New default cache, shared with the Datasets library
+hf_cache_home = os.path.expanduser(
+    os.getenv("HF_HOME", os.path.join(os.getenv("XDG_CACHE_HOME", "~/.cache"), "huggingface"))
+)
+default_cache_path = os.path.join(hf_cache_home, "transformers")
+
+# Onetime move from the old location to the new one if no ENV variable has been set.
+if (
+    os.path.isdir(old_default_cache_path)
+    and not os.path.isdir(default_cache_path)
+    and "PYTORCH_PRETRAINED_BERT_CACHE" not in os.environ
+    and "PYTORCH_TRANSFORMERS_CACHE" not in os.environ
+    and "TRANSFORMERS_CACHE" not in os.environ
+):
+    logger.warn(
+        "In Transformers v4.0.0, the default path to cache downloaded models changed from "
+        "'~/.cache/torch/transformers' to '~/.cache/huggingface/transformers'. Since you don't seem to have overridden "
+        "and '~/.cache/torch/transformers' is a directory that exists, we're moving it to "
+        "'~/.cache/huggingface/transformers' to avoid redownloading models you have already in the cache. You should "
+        "only see this message once."
+    )
+    shutil.move(old_default_cache_path, default_cache_path)
+
+PYTORCH_PRETRAINED_BERT_CACHE = os.getenv("PYTORCH_PRETRAINED_BERT_CACHE", default_cache_path)
+PYTORCH_TRANSFORMERS_CACHE = os.getenv("PYTORCH_TRANSFORMERS_CACHE", PYTORCH_PRETRAINED_BERT_CACHE)
+TRANSFORMERS_CACHE = os.getenv("TRANSFORMERS_CACHE", PYTORCH_TRANSFORMERS_CACHE)
+DISABLE_TELEMETRY = os.getenv("DISABLE_TELEMETRY", False)
+
+WEIGHTS_NAME = "pytorch_model.bin"
+TF2_WEIGHTS_NAME = "tf_model.h5"
+TF_WEIGHTS_NAME = "model.ckpt"
+FLAX_WEIGHTS_NAME = "flax_model.msgpack"
+CONFIG_NAME = "config.json"
+FEATURE_EXTRACTOR_NAME = "preprocessor_config.json"
+MODEL_CARD_NAME = "modelcard.json"
+
+SENTENCEPIECE_UNDERLINE = "▁"
+SPIECE_UNDERLINE = SENTENCEPIECE_UNDERLINE  # Kept for backward compatibility
+
+MULTIPLE_CHOICE_DUMMY_INPUTS = [
+    [[0, 1, 0, 1], [1, 0, 0, 1]]
+] * 2  # Needs to have 0s and 1s only since XLM uses it for langs too.
+DUMMY_INPUTS = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
+DUMMY_MASK = [[1, 1, 1, 1, 1], [1, 1, 1, 0, 0], [0, 0, 0, 1, 1]]
+
+S3_BUCKET_PREFIX = "https://s3.amazonaws.com/models.huggingface.co/bert"
+CLOUDFRONT_DISTRIB_PREFIX = "https://cdn.huggingface.co"
+HUGGINGFACE_CO_PREFIX = "https://huggingface.co/{model_id}/resolve/{revision}/{filename}"
+
+PRESET_MIRROR_DICT = {
+    "tuna": "https://mirrors.tuna.tsinghua.edu.cn/hugging-face-models",
+    "bfsu": "https://mirrors.bfsu.edu.cn/hugging-face-models",
+}
+
+
+_is_offline_mode = True if os.environ.get("TRANSFORMERS_OFFLINE", "0").upper() in ENV_VARS_TRUE_VALUES else False
+
+
+def is_offline_mode():
+    return _is_offline_mode
+
+
+def is_torch_available():
+    return _torch_available
+
+
+def is_torch_cuda_available():
+    if is_torch_available():
+        import torch
+
+        return torch.cuda.is_available()
+    else:
+        return False
+
+
+def is_tf_available():
+    return _tf_available
+
+
+def is_onnx_available():
+    return _onnx_available
+
+
+def is_flax_available():
+    return _flax_available
+
+
+def is_torch_tpu_available():
+    if not _torch_available:
+        return False
+    # This test is probably enough, but just in case, we unpack a bit.
+    if importlib.util.find_spec("torch_xla") is None:
+        return False
+    if importlib.util.find_spec("torch_xla.core") is None:
+        return False
+    return importlib.util.find_spec("torch_xla.core.xla_model") is not None
+
+
+def is_datasets_available():
+    return _datasets_available
+
+
+def is_psutil_available():
+    return importlib.util.find_spec("psutil") is not None
+
+
+def is_py3nvml_available():
+    return importlib.util.find_spec("py3nvml") is not None
+
+
+def is_apex_available():
+    return importlib.util.find_spec("apex") is not None
+
+
+def is_faiss_available():
+    return _faiss_available
+
+
+def is_sklearn_available():
+    if importlib.util.find_spec("sklearn") is None:
+        return False
+    if importlib.util.find_spec("scipy") is None:
+        return False
+    return importlib.util.find_spec("sklearn.metrics") and importlib.util.find_spec("scipy.stats")
+
+
+def is_sentencepiece_available():
+    return importlib.util.find_spec("sentencepiece") is not None
+
+
+def is_protobuf_available():
+    if importlib.util.find_spec("google") is None:
+        return False
+    return importlib.util.find_spec("google.protobuf") is not None
+
+
+def is_tokenizers_available():
+    return importlib.util.find_spec("tokenizers") is not None
+
+
+def is_in_notebook():
+    try:
+        # Test adapted from tqdm.autonotebook: https://github.com/tqdm/tqdm/blob/master/tqdm/autonotebook.py
+        get_ipython = sys.modules["IPython"].get_ipython
+        if "IPKernelApp" not in get_ipython().config:
+            raise ImportError("console")
+        if "VSCODE_PID" in os.environ:
+            raise ImportError("vscode")
+
+        return importlib.util.find_spec("IPython") is not None
+    except (AttributeError, ImportError, KeyError):
+        return False
+
+
+def is_scatter_available():
+    return _scatter_available
+
+
+def is_pandas_available():
+    return importlib.util.find_spec("pandas") is not None
+
+
+def is_sagemaker_distributed_available():
+    # Get the sagemaker specific env variable.
+    sagemaker_params = os.getenv("SM_FRAMEWORK_PARAMS", "{}")
+    try:
+        # Parse it and check the field "sagemaker_distributed_dataparallel_enabled".
+        sagemaker_params = json.loads(sagemaker_params)
+        if not sagemaker_params.get("sagemaker_distributed_dataparallel_enabled", False):
+            return False
+    except json.JSONDecodeError:
+        return False
+    # Lastly, check if the `smdistributed` module is present.
+    return importlib.util.find_spec("smdistributed") is not None
+
+
+def is_training_run_on_sagemaker():
+    return "SAGEMAKER_JOB_NAME" in os.environ and not DISABLE_TELEMETRY
+
+
+def is_soundfile_availble():
+    return _soundfile_available
+
+
+def is_torchaudio_available():
+    return _torchaudio_available
+
+
+def torch_only_method(fn):
+    def wrapper(*args, **kwargs):
+        if not _torch_available:
+            raise ImportError(
+                "You need to install pytorch to use this method or class, "
+                "or activate it with environment variables USE_TORCH=1 and USE_TF=0."
+            )
+        else:
+            return fn(*args, **kwargs)
+
+    return wrapper
+
+
+# docstyle-ignore
+DATASETS_IMPORT_ERROR = """
+{0} requires the 🤗 Datasets library but it was not found in your environment. You can install it with:
+```
+pip install datasets
+```
+In a notebook or a colab, you can install it by executing a cell with
+```
+!pip install datasets
+```
+then restarting your kernel.
+
+Note that if you have a local folder named `datasets` or a local python file named `datasets.py` in your current
+working directory, python may try to import this instead of the 🤗 Datasets library. You should rename this folder or
+that python file if that's the case.
+"""
+
+
+# docstyle-ignore
+TOKENIZERS_IMPORT_ERROR = """
+{0} requires the 🤗 Tokenizers library but it was not found in your environment. You can install it with:
+```
+pip install tokenizers
+```
+In a notebook or a colab, you can install it by executing a cell with
+```
+!pip install tokenizers
+```
+"""
+
+
+# docstyle-ignore
+SENTENCEPIECE_IMPORT_ERROR = """
+{0} requires the SentencePiece library but it was not found in your environment. Checkout the instructions on the
+installation page of its repo: https://github.com/google/sentencepiece#installation and follow the ones
+that match your environment.
+"""
+
+
+# docstyle-ignore
+PROTOBUF_IMPORT_ERROR = """
+{0} requires the protobuf library but it was not found in your environment. Checkout the instructions on the
+installation page of its repo: https://github.com/protocolbuffers/protobuf/tree/master/python#installation and follow the ones
+that match your environment.
+"""
+
+
+# docstyle-ignore
+FAISS_IMPORT_ERROR = """
+{0} requires the faiss library but it was not found in your environment. Checkout the instructions on the
+installation page of its repo: https://github.com/facebookresearch/faiss/blob/master/INSTALL.md and follow the ones
+that match your environment.
+"""
+
+
+# docstyle-ignore
+PYTORCH_IMPORT_ERROR = """
+{0} requires the PyTorch library but it was not found in your environment. Checkout the instructions on the
+installation page: https://pytorch.org/get-started/locally/ and follow the ones that match your environment.
+"""
+
+
+# docstyle-ignore
+SKLEARN_IMPORT_ERROR = """
+{0} requires the scikit-learn library but it was not found in your environment. You can install it with:
+```
+pip install -U scikit-learn
+```
+In a notebook or a colab, you can install it by executing a cell with
+```
+!pip install -U scikit-learn
+```
+"""
+
+
+# docstyle-ignore
+TENSORFLOW_IMPORT_ERROR = """
+{0} requires the TensorFlow library but it was not found in your environment. Checkout the instructions on the
+installation page: https://www.tensorflow.org/install and follow the ones that match your environment.
+"""
+
+
+# docstyle-ignore
+FLAX_IMPORT_ERROR = """
+{0} requires the FLAX library but it was not found in your environment. Checkout the instructions on the
+installation page: https://github.com/google/flax and follow the ones that match your environment.
+"""
+
+
+# docstyle-ignore
+SCATTER_IMPORT_ERROR = """
+{0} requires the torch-scatter library but it was not found in your environment. You can install it with pip as
+explained here: https://github.com/rusty1s/pytorch_scatter.
+"""
+
+
+# docstyle-ignore
+PANDAS_IMPORT_ERROR = """
+{0} requires the pandas library but it was not found in your environment. You can install it with pip as
+explained here: https://pandas.pydata.org/pandas-docs/stable/getting_started/install.html.
+"""
+
+
+def requires_datasets(obj):
+    name = obj.__name__ if hasattr(obj, "__name__") else obj.__class__.__name__
+    if not is_datasets_available():
+        raise ImportError(DATASETS_IMPORT_ERROR.format(name))
+
+
+def requires_faiss(obj):
+    name = obj.__name__ if hasattr(obj, "__name__") else obj.__class__.__name__
+    if not is_faiss_available():
+        raise ImportError(FAISS_IMPORT_ERROR.format(name))
+
+
+def requires_pytorch(obj):
+    name = obj.__name__ if hasattr(obj, "__name__") else obj.__class__.__name__
+    if not is_torch_available():
+        raise ImportError(PYTORCH_IMPORT_ERROR.format(name))
+
+
+def requires_sklearn(obj):
+    name = obj.__name__ if hasattr(obj, "__name__") else obj.__class__.__name__
+    if not is_sklearn_available():
+        raise ImportError(SKLEARN_IMPORT_ERROR.format(name))
+
+
+def requires_tf(obj):
+    name = obj.__name__ if hasattr(obj, "__name__") else obj.__class__.__name__
+    if not is_tf_available():
+        raise ImportError(TENSORFLOW_IMPORT_ERROR.format(name))
+
+
+def requires_flax(obj):
+    name = obj.__name__ if hasattr(obj, "__name__") else obj.__class__.__name__
+    if not is_flax_available():
+        raise ImportError(FLAX_IMPORT_ERROR.format(name))
+
+
+def requires_tokenizers(obj):
+    name = obj.__name__ if hasattr(obj, "__name__") else obj.__class__.__name__
+    if not is_tokenizers_available():
+        raise ImportError(TOKENIZERS_IMPORT_ERROR.format(name))
+
+
+def requires_sentencepiece(obj):
+    name = obj.__name__ if hasattr(obj, "__name__") else obj.__class__.__name__
+    if not is_sentencepiece_available():
+        raise ImportError(SENTENCEPIECE_IMPORT_ERROR.format(name))
+
+
+def requires_protobuf(obj):
+    name = obj.__name__ if hasattr(obj, "__name__") else obj.__class__.__name__
+    if not is_protobuf_available():
+        raise ImportError(PROTOBUF_IMPORT_ERROR.format(name))
+
+
+def requires_pandas(obj):
+    name = obj.__name__ if hasattr(obj, "__name__") else obj.__class__.__name__
+    if not is_pandas_available():
+        raise ImportError(PANDAS_IMPORT_ERROR.format(name))
+
+
+def requires_scatter(obj):
+    name = obj.__name__ if hasattr(obj, "__name__") else obj.__class__.__name__
+    if not is_scatter_available():
+        raise ImportError(SCATTER_IMPORT_ERROR.format(name))
+
+
+def add_start_docstrings(*docstr):
+    def docstring_decorator(fn):
+        fn.__doc__ = "".join(docstr) + (fn.__doc__ if fn.__doc__ is not None else "")
+        return fn
+
+    return docstring_decorator
+
+
+def add_start_docstrings_to_model_forward(*docstr):
+    def docstring_decorator(fn):
+        class_name = ":class:`~transformers.{}`".format(fn.__qualname__.split(".")[0])
+        intro = "   The {} forward method, overrides the :func:`__call__` special method.".format(class_name)
+        note = r"""
+
+    .. note::
+        Although the recipe for forward pass needs to be defined within this function, one should call the
+        :class:`Module` instance afterwards instead of this since the former takes care of running the pre and post
+        processing steps while the latter silently ignores them.
+        """
+        fn.__doc__ = intro + note + "".join(docstr) + (fn.__doc__ if fn.__doc__ is not None else "")
+        return fn
+
+    return docstring_decorator
+
+
+def add_end_docstrings(*docstr):
+    def docstring_decorator(fn):
+        fn.__doc__ = fn.__doc__ + "".join(docstr)
+        return fn
+
+    return docstring_decorator
+
+
+PT_RETURN_INTRODUCTION = r"""
+    Returns:
+        :class:`~{full_output_type}` or :obj:`tuple(torch.FloatTensor)`: A :class:`~{full_output_type}` (if
+        ``return_dict=True`` is passed or when ``config.return_dict=True``) or a tuple of :obj:`torch.FloatTensor`
+        comprising various elements depending on the configuration (:class:`~transformers.{config_class}`) and inputs.
+
+"""
+
+
+TF_RETURN_INTRODUCTION = r"""
+    Returns:
+        :class:`~{full_output_type}` or :obj:`tuple(tf.Tensor)`: A :class:`~{full_output_type}` (if
+        ``return_dict=True`` is passed or when ``config.return_dict=True``) or a tuple of :obj:`tf.Tensor` comprising
+        various elements depending on the configuration (:class:`~transformers.{config_class}`) and inputs.
+
+"""
+
+
+def _get_indent(t):
+    """Returns the indentation in the first line of t"""
+    search = re.search(r"^(\s*)\S", t)
+    return "" if search is None else search.groups()[0]
+
+
+def _convert_output_args_doc(output_args_doc):
+    """Convert output_args_doc to display properly."""
+    # Split output_arg_doc in blocks argument/description
+    indent = _get_indent(output_args_doc)
+    blocks = []
+    current_block = ""
+    for line in output_args_doc.split("\n"):
+        # If the indent is the same as the beginning, the line is the name of new arg.
+        if _get_indent(line) == indent:
+            if len(current_block) > 0:
+                blocks.append(current_block[:-1])
+            current_block = f"{line}\n"
+        else:
+            # Otherwise it's part of the description of the current arg.
+            # We need to remove 2 spaces to the indentation.
+            current_block += f"{line[2:]}\n"
+    blocks.append(current_block[:-1])
+
+    # Format each block for proper rendering
+    for i in range(len(blocks)):
+        blocks[i] = re.sub(r"^(\s+)(\S+)(\s+)", r"\1- **\2**\3", blocks[i])
+        blocks[i] = re.sub(r":\s*\n\s*(\S)", r" -- \1", blocks[i])
+
+    return "\n".join(blocks)
+
+
+def _prepare_output_docstrings(output_type, config_class):
+    """
+    Prepares the return part of the docstring using `output_type`.
+    """
+    docstrings = output_type.__doc__
+
+    # Remove the head of the docstring to keep the list of args only
+    lines = docstrings.split("\n")
+    i = 0
+    while i < len(lines) and re.search(r"^\s*(Args|Parameters):\s*$", lines[i]) is None:
+        i += 1
+    if i < len(lines):
+        docstrings = "\n".join(lines[(i + 1) :])
+        docstrings = _convert_output_args_doc(docstrings)
+
+    # Add the return introduction
+    full_output_type = f"{output_type.__module__}.{output_type.__name__}"
+    intro = TF_RETURN_INTRODUCTION if output_type.__name__.startswith("TF") else PT_RETURN_INTRODUCTION
+    intro = intro.format(full_output_type=full_output_type, config_class=config_class)
+    return intro + docstrings
+
+
+PT_TOKEN_CLASSIFICATION_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import torch
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
+        >>> labels = torch.tensor([1] * inputs["input_ids"].size(1)).unsqueeze(0)  # Batch size 1
+
+        >>> outputs = model(**inputs, labels=labels)
+        >>> loss = outputs.loss
+        >>> logits = outputs.logits
+"""
+
+PT_QUESTION_ANSWERING_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import torch
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
+        >>> inputs = tokenizer(question, text, return_tensors='pt')
+        >>> start_positions = torch.tensor([1])
+        >>> end_positions = torch.tensor([3])
+
+        >>> outputs = model(**inputs, start_positions=start_positions, end_positions=end_positions)
+        >>> loss = outputs.loss
+        >>> start_scores = outputs.start_logits
+        >>> end_scores = outputs.end_logits
+"""
+
+PT_SEQUENCE_CLASSIFICATION_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import torch
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
+        >>> labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
+        >>> outputs = model(**inputs, labels=labels)
+        >>> loss = outputs.loss
+        >>> logits = outputs.logits
+"""
+
+PT_MASKED_LM_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import torch
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> inputs = tokenizer("The capital of France is {mask}.", return_tensors="pt")
+        >>> labels = tokenizer("The capital of France is Paris.", return_tensors="pt")["input_ids"]
+
+        >>> outputs = model(**inputs, labels=labels)
+        >>> loss = outputs.loss
+        >>> logits = outputs.logits
+"""
+
+PT_BASE_MODEL_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import torch
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
+        >>> outputs = model(**inputs)
+
+        >>> last_hidden_states = outputs.last_hidden_state
+"""
+
+PT_MULTIPLE_CHOICE_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import torch
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> prompt = "In Italy, pizza served in formal settings, such as at a restaurant, is presented unsliced."
+        >>> choice0 = "It is eaten with a fork and a knife."
+        >>> choice1 = "It is eaten while held in the hand."
+        >>> labels = torch.tensor(0).unsqueeze(0)  # choice0 is correct (according to Wikipedia ;)), batch size 1
+
+        >>> encoding = tokenizer([[prompt, prompt], [choice0, choice1]], return_tensors='pt', padding=True)
+        >>> outputs = model(**{{k: v.unsqueeze(0) for k,v in encoding.items()}}, labels=labels)  # batch size is 1
+
+        >>> # the linear classifier still needs to be trained
+        >>> loss = outputs.loss
+        >>> logits = outputs.logits
+"""
+
+PT_CAUSAL_LM_SAMPLE = r"""
+    Example::
+
+        >>> import torch
+        >>> from transformers import {tokenizer_class}, {model_class}
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
+        >>> outputs = model(**inputs, labels=inputs["input_ids"])
+        >>> loss = outputs.loss
+        >>> logits = outputs.logits
+"""
+
+TF_TOKEN_CLASSIFICATION_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import tensorflow as tf
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="tf")
+        >>> input_ids = inputs["input_ids"]
+        >>> inputs["labels"] = tf.reshape(tf.constant([1] * tf.size(input_ids).numpy()), (-1, tf.size(input_ids))) # Batch size 1
+
+        >>> outputs = model(inputs)
+        >>> loss = outputs.loss
+        >>> logits = outputs.logits
+"""
+
+TF_QUESTION_ANSWERING_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import tensorflow as tf
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
+        >>> input_dict = tokenizer(question, text, return_tensors='tf')
+        >>> outputs = model(input_dict)
+        >>> start_logits = outputs.start_logits
+        >>> end_logits = outputs.end_logits
+
+        >>> all_tokens = tokenizer.convert_ids_to_tokens(input_dict["input_ids"].numpy()[0])
+        >>> answer = ' '.join(all_tokens[tf.math.argmax(start_logits, 1)[0] : tf.math.argmax(end_logits, 1)[0]+1])
+"""
+
+TF_SEQUENCE_CLASSIFICATION_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import tensorflow as tf
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="tf")
+        >>> inputs["labels"] = tf.reshape(tf.constant(1), (-1, 1)) # Batch size 1
+
+        >>> outputs = model(inputs)
+        >>> loss = outputs.loss
+        >>> logits = outputs.logits
+"""
+
+TF_MASKED_LM_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import tensorflow as tf
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> inputs = tokenizer("The capital of France is {mask}.", return_tensors="tf")
+        >>> inputs["labels"] = tokenizer("The capital of France is Paris.", return_tensors="tf")["input_ids"]
+
+        >>> outputs = model(inputs)
+        >>> loss = outputs.loss
+        >>> logits = outputs.logits
+"""
+
+TF_BASE_MODEL_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import tensorflow as tf
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="tf")
+        >>> outputs = model(inputs)
+
+        >>> last_hidden_states = outputs.last_hidden_state
+"""
+
+TF_MULTIPLE_CHOICE_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import tensorflow as tf
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> prompt = "In Italy, pizza served in formal settings, such as at a restaurant, is presented unsliced."
+        >>> choice0 = "It is eaten with a fork and a knife."
+        >>> choice1 = "It is eaten while held in the hand."
+
+        >>> encoding = tokenizer([[prompt, prompt], [choice0, choice1]], return_tensors='tf', padding=True)
+        >>> inputs = {{k: tf.expand_dims(v, 0) for k, v in encoding.items()}}
+        >>> outputs = model(inputs)  # batch size is 1
+
+        >>> # the linear classifier still needs to be trained
+        >>> logits = outputs.logits
+"""
+
+TF_CAUSAL_LM_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import tensorflow as tf
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="tf")
+        >>> outputs = model(inputs)
+        >>> logits = outputs.logits
+"""
+
+
+def add_code_sample_docstrings(
+    *docstr, tokenizer_class=None, checkpoint=None, output_type=None, config_class=None, mask=None
+):
+    def docstring_decorator(fn):
+        model_class = fn.__qualname__.split(".")[0]
+        is_tf_class = model_class[:2] == "TF"
+        doc_kwargs = dict(model_class=model_class, tokenizer_class=tokenizer_class, checkpoint=checkpoint)
+
+        if "SequenceClassification" in model_class:
+            code_sample = TF_SEQUENCE_CLASSIFICATION_SAMPLE if is_tf_class else PT_SEQUENCE_CLASSIFICATION_SAMPLE
+        elif "QuestionAnswering" in model_class:
+            code_sample = TF_QUESTION_ANSWERING_SAMPLE if is_tf_class else PT_QUESTION_ANSWERING_SAMPLE
+        elif "TokenClassification" in model_class:
+            code_sample = TF_TOKEN_CLASSIFICATION_SAMPLE if is_tf_class else PT_TOKEN_CLASSIFICATION_SAMPLE
+        elif "MultipleChoice" in model_class:
+            code_sample = TF_MULTIPLE_CHOICE_SAMPLE if is_tf_class else PT_MULTIPLE_CHOICE_SAMPLE
+        elif "MaskedLM" in model_class or model_class in ["FlaubertWithLMHeadModel", "XLMWithLMHeadModel"]:
+            doc_kwargs["mask"] = "[MASK]" if mask is None else mask
+            code_sample = TF_MASKED_LM_SAMPLE if is_tf_class else PT_MASKED_LM_SAMPLE
+        elif "LMHead" in model_class or "CausalLM" in model_class:
+            code_sample = TF_CAUSAL_LM_SAMPLE if is_tf_class else PT_CAUSAL_LM_SAMPLE
+        elif "Model" in model_class or "Encoder" in model_class:
+            code_sample = TF_BASE_MODEL_SAMPLE if is_tf_class else PT_BASE_MODEL_SAMPLE
+        else:
+            raise ValueError(f"Docstring can't be built for model {model_class}")
+
+        output_doc = _prepare_output_docstrings(output_type, config_class) if output_type is not None else ""
+        built_doc = code_sample.format(**doc_kwargs)
+        fn.__doc__ = (fn.__doc__ or "") + "".join(docstr) + output_doc + built_doc
+        return fn
+
+    return docstring_decorator
+
+
+def replace_return_docstrings(output_type=None, config_class=None):
+    def docstring_decorator(fn):
+        docstrings = fn.__doc__
+        lines = docstrings.split("\n")
+        i = 0
+        while i < len(lines) and re.search(r"^\s*Returns?:\s*$", lines[i]) is None:
+            i += 1
+        if i < len(lines):
+            lines[i] = _prepare_output_docstrings(output_type, config_class)
+            docstrings = "\n".join(lines)
+        else:
+            raise ValueError(
+                f"The function {fn} should have an empty 'Return:' or 'Returns:' in its docstring as placeholder, current docstring is:\n{docstrings}"
+            )
+        fn.__doc__ = docstrings
+        return fn
+
+    return docstring_decorator
+
+
+def is_remote_url(url_or_filename):
+    parsed = urlparse(url_or_filename)
+    return parsed.scheme in ("http", "https")
+
+
+def hf_bucket_url(
+    model_id: str, filename: str, subfolder: Optional[str] = None, revision: Optional[str] = None, mirror=None
+) -> str:
+    """
+    Resolve a model identifier, a file name, and an optional revision id, to a huggingface.co-hosted url, redirecting
+    to Cloudfront (a Content Delivery Network, or CDN) for large files.
+
+    Cloudfront is replicated over the globe so downloads are way faster for the end user (and it also lowers our
+    bandwidth costs).
+
+    Cloudfront aggressively caches files by default (default TTL is 24 hours), however this is not an issue here
+    because we migrated to a git-based versioning system on huggingface.co, so we now store the files on S3/Cloudfront
+    in a content-addressable way (i.e., the file name is its hash). Using content-addressable filenames means cache
+    can't ever be stale.
+
+    In terms of client-side caching from this library, we base our caching on the objects' ETag. An object' ETag is:
+    its sha1 if stored in git, or its sha256 if stored in git-lfs. Files cached locally from transformers before v3.5.0
+    are not shared with those new files, because the cached file's name contains a hash of the url (which changed).
+    """
+    if subfolder is not None:
+        filename = f"{subfolder}/{filename}"
+
+    if mirror:
+        endpoint = PRESET_MIRROR_DICT.get(mirror, mirror)
+        legacy_format = "/" not in model_id
+        if legacy_format:
+            return f"{endpoint}/{model_id}-{filename}"
+        else:
+            return f"{endpoint}/{model_id}/{filename}"
+
+    if revision is None:
+        revision = "main"
+    return HUGGINGFACE_CO_PREFIX.format(model_id=model_id, revision=revision, filename=filename)
+
+
+def url_to_filename(url: str, etag: Optional[str] = None) -> str:
+    """
+    Convert `url` into a hashed filename in a repeatable way. If `etag` is specified, append its hash to the url's,
+    delimited by a period. If the url ends with .h5 (Keras HDF5 weights) adds '.h5' to the name so that TF 2.0 can
+    identify it as a HDF5 file (see
+    https://github.com/tensorflow/tensorflow/blob/00fad90125b18b80fe054de1055770cfb8fe4ba3/tensorflow/python/keras/engine/network.py#L1380)
+    """
+    url_bytes = url.encode("utf-8")
+    filename = sha256(url_bytes).hexdigest()
+
+    if etag:
+        etag_bytes = etag.encode("utf-8")
+        filename += "." + sha256(etag_bytes).hexdigest()
+
+    if url.endswith(".h5"):
+        filename += ".h5"
+
+    return filename
+
+
+def filename_to_url(filename, cache_dir=None):
+    """
+    Return the url and etag (which may be ``None``) stored for `filename`. Raise ``EnvironmentError`` if `filename` or
+    its stored metadata do not exist.
+    """
+    if cache_dir is None:
+        cache_dir = TRANSFORMERS_CACHE
+    if isinstance(cache_dir, Path):
+        cache_dir = str(cache_dir)
+
+    cache_path = os.path.join(cache_dir, filename)
+    if not os.path.exists(cache_path):
+        raise EnvironmentError("file {} not found".format(cache_path))
+
+    meta_path = cache_path + ".json"
+    if not os.path.exists(meta_path):
+        raise EnvironmentError("file {} not found".format(meta_path))
+
+    with open(meta_path, encoding="utf-8") as meta_file:
+        metadata = json.load(meta_file)
+    url = metadata["url"]
+    etag = metadata["etag"]
+
+    return url, etag
+
+
+def get_cached_models(cache_dir: Union[str, Path] = None) -> List[Tuple]:
+    """
+    Returns a list of tuples representing model binaries that are cached locally. Each tuple has shape
+    :obj:`(model_url, etag, size_MB)`. Filenames in :obj:`cache_dir` are use to get the metadata for each model, only
+    urls ending with `.bin` are added.
+
+    Args:
+        cache_dir (:obj:`Union[str, Path]`, `optional`):
+            The cache directory to search for models within. Will default to the transformers cache if unset.
+
+    Returns:
+        List[Tuple]: List of tuples each with shape :obj:`(model_url, etag, size_MB)`
+    """
+    if cache_dir is None:
+        cache_dir = TRANSFORMERS_CACHE
+    elif isinstance(cache_dir, Path):
+        cache_dir = str(cache_dir)
+
+    cached_models = []
+    for file in os.listdir(cache_dir):
+        if file.endswith(".json"):
+            meta_path = os.path.join(cache_dir, file)
+            with open(meta_path, encoding="utf-8") as meta_file:
+                metadata = json.load(meta_file)
+                url = metadata["url"]
+                etag = metadata["etag"]
+                if url.endswith(".bin"):
+                    size_MB = os.path.getsize(meta_path.strip(".json")) / 1e6
+                    cached_models.append((url, etag, size_MB))
+
+    return cached_models
+
+
+def cached_path(
+    url_or_filename,
+    cache_dir=None,
+    force_download=False,
+    proxies=None,
+    resume_download=False,
+    user_agent: Union[Dict, str, None] = None,
+    extract_compressed_file=False,
+    force_extract=False,
+    use_auth_token: Union[bool, str, None] = None,
+    local_files_only=False,
+) -> Optional[str]:
+    """
+    Given something that might be a URL (or might be a local path), determine which. If it's a URL, download the file
+    and cache it, and return the path to the cached file. If it's already a local path, make sure the file exists and
+    then return the path
+
+    Args:
+        cache_dir: specify a cache directory to save the file to (overwrite the default cache dir).
+        force_download: if True, re-download the file even if it's already cached in the cache dir.
+        resume_download: if True, resume the download if incompletely received file is found.
+        user_agent: Optional string or dict that will be appended to the user-agent on remote requests.
+        use_auth_token: Optional string or boolean to use as Bearer token for remote files. If True,
+            will get token from ~/.huggingface.
+        extract_compressed_file: if True and the path point to a zip or tar file, extract the compressed
+            file in a folder along the archive.
+        force_extract: if True when extract_compressed_file is True and the archive was already extracted,
+            re-extract the archive and override the folder where it was extracted.
+
+    Return:
+        Local path (string) of file or if networking is off, last version of file cached on disk.
+
+    Raises:
+        In case of non-recoverable file (non-existent or inaccessible url + no cache on disk).
+    """
+    if cache_dir is None:
+        cache_dir = TRANSFORMERS_CACHE
+    if isinstance(url_or_filename, Path):
+        url_or_filename = str(url_or_filename)
+    if isinstance(cache_dir, Path):
+        cache_dir = str(cache_dir)
+
+    if is_offline_mode() and not local_files_only:
+        logger.info("Offline mode: forcing local_files_only=True")
+        local_files_only = True
+
+    if is_remote_url(url_or_filename):
+        # URL, so get it from the cache (downloading if necessary)
+        output_path = get_from_cache(
+            url_or_filename,
+            cache_dir=cache_dir,
+            force_download=force_download,
+            proxies=proxies,
+            resume_download=resume_download,
+            user_agent=user_agent,
+            use_auth_token=use_auth_token,
+            local_files_only=local_files_only,
+        )
+    elif os.path.exists(url_or_filename):
+        # File, and it exists.
+        output_path = url_or_filename
+    elif urlparse(url_or_filename).scheme == "":
+        # File, but it doesn't exist.
+        raise EnvironmentError("file {} not found".format(url_or_filename))
+    else:
+        # Something unknown
+        raise ValueError("unable to parse {} as a URL or as a local path".format(url_or_filename))
+
+    if extract_compressed_file:
+        if not is_zipfile(output_path) and not tarfile.is_tarfile(output_path):
+            return output_path
+
+        # Path where we extract compressed archives
+        # We avoid '.' in dir name and add "-extracted" at the end: "./model.zip" => "./model-zip-extracted/"
+        output_dir, output_file = os.path.split(output_path)
+        output_extract_dir_name = output_file.replace(".", "-") + "-extracted"
+        output_path_extracted = os.path.join(output_dir, output_extract_dir_name)
+
+        if os.path.isdir(output_path_extracted) and os.listdir(output_path_extracted) and not force_extract:
+            return output_path_extracted
+
+        # Prevent parallel extractions
+        lock_path = output_path + ".lock"
+        with FileLock(lock_path):
+            shutil.rmtree(output_path_extracted, ignore_errors=True)
+            os.makedirs(output_path_extracted)
+            if is_zipfile(output_path):
+                with ZipFile(output_path, "r") as zip_file:
+                    zip_file.extractall(output_path_extracted)
+                    zip_file.close()
+            elif tarfile.is_tarfile(output_path):
+                tar_file = tarfile.open(output_path)
+                tar_file.extractall(output_path_extracted)
+                tar_file.close()
+            else:
+                raise EnvironmentError("Archive format of {} could not be identified".format(output_path))
+
+        return output_path_extracted
+
+    return output_path
+
+
+def define_sagemaker_information():
+    try:
+        instance_data = requests.get(os.environ["ECS_CONTAINER_METADATA_URI"]).json()
+        dlc_container_used = instance_data["Image"]
+        dlc_tag = instance_data["Image"].split(":")[1]
+    except Exception:
+        dlc_container_used = None
+        dlc_tag = None
+
+    sagemaker_params = json.loads(os.getenv("SM_FRAMEWORK_PARAMS", "{}"))
+    runs_distributed_training = True if "sagemaker_distributed_dataparallel_enabled" in sagemaker_params else False
+    account_id = os.getenv("TRAINING_JOB_ARN").split(":")[4] if "TRAINING_JOB_ARN" in os.environ else None
+
+    sagemaker_object = {
+        "sm_framework": os.getenv("SM_FRAMEWORK_MODULE", None),
+        "sm_region": os.getenv("AWS_REGION", None),
+        "sm_number_gpu": os.getenv("SM_NUM_GPUS", 0),
+        "sm_number_cpu": os.getenv("SM_NUM_CPUS", 0),
+        "sm_distributed_training": runs_distributed_training,
+        "sm_deep_learning_container": dlc_container_used,
+        "sm_deep_learning_container_tag": dlc_tag,
+        "sm_account_id": account_id,
+    }
+    return sagemaker_object
+
+
+def http_user_agent(user_agent: Union[Dict, str, None] = None) -> str:
+    """
+    Formats a user-agent string with basic info about a request.
+    """
+    ua = "transformers/{}; python/{}".format(__version__, sys.version.split()[0])
+    if is_torch_available():
+        ua += f"; torch/{_torch_version}"
+    if is_tf_available():
+        ua += f"; tensorflow/{_tf_version}"
+    if is_training_run_on_sagemaker():
+        ua += "; " + "; ".join(f"{k}/{v}" for k, v in define_sagemaker_information().items())
+    if isinstance(user_agent, dict):
+        ua += "; " + "; ".join(f"{k}/{v}" for k, v in user_agent.items())
+    elif isinstance(user_agent, str):
+        ua += "; " + user_agent
+    return ua
+
+
+def http_get(url: str, temp_file: BinaryIO, proxies=None, resume_size=0, headers: Optional[Dict[str, str]] = None):
+    """
+    Donwload remote file. Do not gobble up errors.
+    """
+    headers = copy.deepcopy(headers)
+    if resume_size > 0:
+        headers["Range"] = "bytes=%d-" % (resume_size,)
+    r = requests.get(url, stream=True, proxies=proxies, headers=headers)
+    r.raise_for_status()
+    content_length = r.headers.get("Content-Length")
+    total = resume_size + int(content_length) if content_length is not None else None
+    progress = tqdm(
+        unit="B",
+        unit_scale=True,
+        total=total,
+        initial=resume_size,
+        desc="Downloading",
+        disable=bool(logging.get_verbosity() == logging.NOTSET),
+    )
+    for chunk in r.iter_content(chunk_size=1024):
+        if chunk:  # filter out keep-alive new chunks
+            progress.update(len(chunk))
+            temp_file.write(chunk)
+    progress.close()
+
+
+def get_from_cache(
+    url: str,
+    cache_dir=None,
+    force_download=False,
+    proxies=None,
+    etag_timeout=10,
+    resume_download=False,
+    user_agent: Union[Dict, str, None] = None,
+    use_auth_token: Union[bool, str, None] = None,
+    local_files_only=False,
+) -> Optional[str]:
+    """
+    Given a URL, look for the corresponding file in the local cache. If it's not there, download it. Then return the
+    path to the cached file.
+
+    Return:
+        Local path (string) of file or if networking is off, last version of file cached on disk.
+
+    Raises:
+        In case of non-recoverable file (non-existent or inaccessible url + no cache on disk).
+    """
+    if cache_dir is None:
+        cache_dir = TRANSFORMERS_CACHE
+    if isinstance(cache_dir, Path):
+        cache_dir = str(cache_dir)
+
+    os.makedirs(cache_dir, exist_ok=True)
+
+    headers = {"user-agent": http_user_agent(user_agent)}
+    if isinstance(use_auth_token, str):
+        headers["authorization"] = "Bearer {}".format(use_auth_token)
+    elif use_auth_token:
+        token = HfFolder.get_token()
+        if token is None:
+            raise EnvironmentError("You specified use_auth_token=True, but a huggingface token was not found.")
+        headers["authorization"] = "Bearer {}".format(token)
+
+    url_to_download = url
+    etag = None
+    if not local_files_only:
+        try:
+            r = requests.head(url, headers=headers, allow_redirects=False, proxies=proxies, timeout=etag_timeout)
+            r.raise_for_status()
+            etag = r.headers.get("X-Linked-Etag") or r.headers.get("ETag")
+            # We favor a custom header indicating the etag of the linked resource, and
+            # we fallback to the regular etag header.
+            # If we don't have any of those, raise an error.
+            if etag is None:
+                raise OSError(
+                    "Distant resource does not have an ETag, we won't be able to reliably ensure reproducibility."
+                )
+            # In case of a redirect,
+            # save an extra redirect on the request.get call,
+            # and ensure we download the exact atomic version even if it changed
+            # between the HEAD and the GET (unlikely, but hey).
+            if 300 <= r.status_code <= 399:
+                url_to_download = r.headers["Location"]
+        except (requests.exceptions.ConnectionError, requests.exceptions.Timeout):
+            # etag is already None
+            pass
+
+    filename = url_to_filename(url, etag)
+
+    # get cache path to put the file
+    cache_path = os.path.join(cache_dir, filename)
+
+    # etag is None == we don't have a connection or we passed local_files_only.
+    # try to get the last downloaded one
+    if etag is None:
+        if os.path.exists(cache_path):
+            return cache_path
+        else:
+            matching_files = [
+                file
+                for file in fnmatch.filter(os.listdir(cache_dir), filename.split(".")[0] + ".*")
+                if not file.endswith(".json") and not file.endswith(".lock")
+            ]
+            if len(matching_files) > 0:
+                return os.path.join(cache_dir, matching_files[-1])
+            else:
+                # If files cannot be found and local_files_only=True,
+                # the models might've been found if local_files_only=False
+                # Notify the user about that
+                if local_files_only:
+                    raise FileNotFoundError(
+                        "Cannot find the requested files in the cached path and outgoing traffic has been"
+                        " disabled. To enable model look-ups and downloads online, set 'local_files_only'"
+                        " to False."
+                    )
+                else:
+                    raise ValueError(
+                        "Connection error, and we cannot find the requested files in the cached path."
+                        " Please try again or make sure your Internet connection is on."
+                    )
+
+    # From now on, etag is not None.
+    if os.path.exists(cache_path) and not force_download:
+        return cache_path
+
+    # Prevent parallel downloads of the same file with a lock.
+    lock_path = cache_path + ".lock"
+    with FileLock(lock_path):
+
+        # If the download just completed while the lock was activated.
+        if os.path.exists(cache_path) and not force_download:
+            # Even if returning early like here, the lock will be released.
+            return cache_path
+
+        if resume_download:
+            incomplete_path = cache_path + ".incomplete"
+
+            @contextmanager
+            def _resumable_file_manager() -> "io.BufferedWriter":
+                with open(incomplete_path, "ab") as f:
+                    yield f
+
+            temp_file_manager = _resumable_file_manager
+            if os.path.exists(incomplete_path):
+                resume_size = os.stat(incomplete_path).st_size
+            else:
+                resume_size = 0
+        else:
+            temp_file_manager = partial(tempfile.NamedTemporaryFile, mode="wb", dir=cache_dir, delete=False)
+            resume_size = 0
+
+        # Download to temporary file, then copy to cache dir once finished.
+        # Otherwise you get corrupt cache entries if the download gets interrupted.
+        with temp_file_manager() as temp_file:
+            logger.info("%s not found in cache or force_download set to True, downloading to %s", url, temp_file.name)
+
+            http_get(url_to_download, temp_file, proxies=proxies, resume_size=resume_size, headers=headers)
+
+        logger.info("storing %s in cache at %s", url, cache_path)
+        os.replace(temp_file.name, cache_path)
+
+        logger.info("creating metadata file for %s", cache_path)
+        meta = {"url": url, "etag": etag}
+        meta_path = cache_path + ".json"
+        with open(meta_path, "w") as meta_file:
+            json.dump(meta, meta_file)
+
+    return cache_path
+
+
+class cached_property(property):
+    """
+    Descriptor that mimics @property but caches output in member variable.
+
+    From tensorflow_datasets
+
+    Built-in in functools from Python 3.8.
+    """
+
+    def __get__(self, obj, objtype=None):
+        # See docs.python.org/3/howto/descriptor.html#properties
+        if obj is None:
+            return self
+        if self.fget is None:
+            raise AttributeError("unreadable attribute")
+        attr = "__cached_" + self.fget.__name__
+        cached = getattr(obj, attr, None)
+        if cached is None:
+            cached = self.fget(obj)
+            setattr(obj, attr, cached)
+        return cached
+
+
+def torch_required(func):
+    # Chose a different decorator name than in tests so it's clear they are not the same.
+    @wraps(func)
+    def wrapper(*args, **kwargs):
+        if is_torch_available():
+            return func(*args, **kwargs)
+        else:
+            raise ImportError(f"Method `{func.__name__}` requires PyTorch.")
+
+    return wrapper
+
+
+def tf_required(func):
+    # Chose a different decorator name than in tests so it's clear they are not the same.
+    @wraps(func)
+    def wrapper(*args, **kwargs):
+        if is_tf_available():
+            return func(*args, **kwargs)
+        else:
+            raise ImportError(f"Method `{func.__name__}` requires TF.")
+
+    return wrapper
+
+
+def is_tensor(x):
+    """ Tests if ``x`` is a :obj:`torch.Tensor`, :obj:`tf.Tensor` or :obj:`np.ndarray`. """
+    if is_torch_available():
+        import torch
+
+        if isinstance(x, torch.Tensor):
+            return True
+    if is_tf_available():
+        import tensorflow as tf
+
+        if isinstance(x, tf.Tensor):
+            return True
+    return isinstance(x, np.ndarray)
+
+
+def _is_numpy(x):
+    return isinstance(x, np.ndarray)
+
+
+def _is_torch(x):
+    import torch
+
+    return isinstance(x, torch.Tensor)
+
+
+def _is_torch_device(x):
+    import torch
+
+    return isinstance(x, torch.device)
+
+
+def _is_tensorflow(x):
+    import tensorflow as tf
+
+    return isinstance(x, tf.Tensor)
+
+
+def _is_jax(x):
+    import jax.numpy as jnp  # noqa: F811
+
+    return isinstance(x, jnp.ndarray)
+
+
+def to_py_obj(obj):
+    """
+    Convert a TensorFlow tensor, PyTorch tensor, Numpy array or python list to a python list.
+    """
+    if isinstance(obj, (dict, UserDict)):
+        return {k: to_py_obj(v) for k, v in obj.items()}
+    elif isinstance(obj, (list, tuple)):
+        return [to_py_obj(o) for o in obj]
+    elif is_tf_available() and _is_tensorflow(obj):
+        return obj.numpy().tolist()
+    elif is_torch_available() and _is_torch(obj):
+        return obj.detach().cpu().tolist()
+    elif isinstance(obj, np.ndarray):
+        return obj.tolist()
+    else:
+        return obj
+
+
+class ModelOutput(OrderedDict):
+    """
+    Base class for all model outputs as dataclass. Has a ``__getitem__`` that allows indexing by integer or slice (like
+    a tuple) or strings (like a dictionary) that will ignore the ``None`` attributes. Otherwise behaves like a regular
+    python dictionary.
+
+    .. warning::
+        You can't unpack a :obj:`ModelOutput` directly. Use the :meth:`~transformers.file_utils.ModelOutput.to_tuple`
+        method to convert it to a tuple before.
+    """
+
+    def __post_init__(self):
+        class_fields = fields(self)
+
+        # Safety and consistency checks
+        assert len(class_fields), f"{self.__class__.__name__} has no fields."
+        assert all(
+            field.default is None for field in class_fields[1:]
+        ), f"{self.__class__.__name__} should not have more than one required field."
+
+        first_field = getattr(self, class_fields[0].name)
+        other_fields_are_none = all(getattr(self, field.name) is None for field in class_fields[1:])
+
+        if other_fields_are_none and not is_tensor(first_field):
+            try:
+                iterator = iter(first_field)
+                first_field_iterator = True
+            except TypeError:
+                first_field_iterator = False
+
+            # if we provided an iterator as first field and the iterator is a (key, value) iterator
+            # set the associated fields
+            if first_field_iterator:
+                for element in iterator:
+                    if (
+                        not isinstance(element, (list, tuple))
+                        or not len(element) == 2
+                        or not isinstance(element[0], str)
+                    ):
+                        break
+                    setattr(self, element[0], element[1])
+                    if element[1] is not None:
+                        self[element[0]] = element[1]
+            elif first_field is not None:
+                self[class_fields[0].name] = first_field
+        else:
+            for field in class_fields:
+                v = getattr(self, field.name)
+                if v is not None:
+                    self[field.name] = v
+
+    def __delitem__(self, *args, **kwargs):
+        raise Exception(f"You cannot use ``__delitem__`` on a {self.__class__.__name__} instance.")
+
+    def setdefault(self, *args, **kwargs):
+        raise Exception(f"You cannot use ``setdefault`` on a {self.__class__.__name__} instance.")
+
+    def pop(self, *args, **kwargs):
+        raise Exception(f"You cannot use ``pop`` on a {self.__class__.__name__} instance.")
+
+    def update(self, *args, **kwargs):
+        raise Exception(f"You cannot use ``update`` on a {self.__class__.__name__} instance.")
+
+    def __getitem__(self, k):
+        if isinstance(k, str):
+            inner_dict = {k: v for (k, v) in self.items()}
+            return inner_dict[k]
+        else:
+            return self.to_tuple()[k]
+
+    def __setattr__(self, name, value):
+        if name in self.keys() and value is not None:
+            # Don't call self.__setitem__ to avoid recursion errors
+            super().__setitem__(name, value)
+        super().__setattr__(name, value)
+
+    def __setitem__(self, key, value):
+        # Will raise a KeyException if needed
+        super().__setitem__(key, value)
+        # Don't call self.__setattr__ to avoid recursion errors
+        super().__setattr__(key, value)
+
+    def to_tuple(self) -> Tuple[Any]:
+        """
+        Convert self to a tuple containing all the attributes/keys that are not ``None``.
+        """
+        return tuple(self[k] for k in self.keys())
+
+
+class ExplicitEnum(Enum):
+    """
+    Enum with more explicit error message for missing values.
+    """
+
+    @classmethod
+    def _missing_(cls, value):
+        raise ValueError(
+            "%r is not a valid %s, please select one of %s"
+            % (value, cls.__name__, str(list(cls._value2member_map_.keys())))
+        )
+
+
+class PaddingStrategy(ExplicitEnum):
+    """
+    Possible values for the ``padding`` argument in :meth:`PreTrainedTokenizerBase.__call__`. Useful for tab-completion
+    in an IDE.
+    """
+
+    LONGEST = "longest"
+    MAX_LENGTH = "max_length"
+    DO_NOT_PAD = "do_not_pad"
+
+
+class TensorType(ExplicitEnum):
+    """
+    Possible values for the ``return_tensors`` argument in :meth:`PreTrainedTokenizerBase.__call__`. Useful for
+    tab-completion in an IDE.
+    """
+
+    PYTORCH = "pt"
+    TENSORFLOW = "tf"
+    NUMPY = "np"
+    JAX = "jax"
+
+
+class _BaseLazyModule(ModuleType):
+    """
+    Module class that surfaces all objects but only performs associated imports when the objects are requested.
+    """
+
+    # Very heavily inspired by optuna.integration._IntegrationModule
+    # https://github.com/optuna/optuna/blob/master/optuna/integration/__init__.py
+    def __init__(self, name, import_structure):
+        super().__init__(name)
+        self._modules = set(import_structure.keys())
+        self._class_to_module = {}
+        for key, values in import_structure.items():
+            for value in values:
+                self._class_to_module[value] = key
+        # Needed for autocompletion in an IDE
+        self.__all__ = list(import_structure.keys()) + sum(import_structure.values(), [])
+
+    # Needed for autocompletion in an IDE
+    def __dir__(self):
+        return super().__dir__() + self.__all__
+
+    def __getattr__(self, name: str) -> Any:
+        if name in self._modules:
+            value = self._get_module(name)
+        elif name in self._class_to_module.keys():
+            module = self._get_module(self._class_to_module[name])
+            value = getattr(module, name)
+        else:
+            raise AttributeError(f"module {self.__name__} has no attribute {name}")
+
+        setattr(self, name, value)
+        return value
+
+    def _get_module(self, module_name: str) -> ModuleType:
+        raise NotImplementedError

generation_beam_search.py

+# coding=utf-8
+# Copyright 2020 The HuggingFace Inc. team
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from abc import ABC, abstractmethod
+from collections import UserDict
+from typing import Optional, Tuple
+
+import torch
+
+from file_utils import add_start_docstrings
+
+
+PROCESS_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size * num_beams, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using any class inheriting from :class:`~transformers.PretrainedTokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        next_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, 2 * num_beams)`):
+            Current scores of the top :obj:`2 * num_beams` non-finished beam hypotheses.
+        next_tokens (:obj:`torch.LongTensor` of shape :obj:`(batch_size, 2 * num_beams)`):
+            :obj:`input_ids` of the tokens corresponding to the top :obj:`2 * num_beams` non-finished beam hypotheses.
+        next_indices (:obj:`torch.LongTensor` of shape :obj:`(batch_size, 2 * num_beams)`):
+            Beam indices indicating to which beam hypothesis the :obj:`next_tokens` correspond.
+        pad_token_id (:obj:`int`, `optional`):
+            The id of the `padding` token.
+        eos_token_id (:obj:`int`, `optional`):
+            The id of the `end-of-sequence` token.
+
+    Return:
+        :obj:`UserDict`: A dictionary composed of the fields as defined above:
+
+            - **next_beam_scores** (:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_beams)`) -- Updated
+              scores of all non-finished beams.
+            - **next_beam_tokens** (:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_beams)`) -- Next tokens
+              to be added to the non-finished beam_hypotheses.
+            - **next_beam_indices** (:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_beams)`) -- Beam indices
+              indicating to which beam the next tokens shall be added.
+
+"""
+
+FINALIZE_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size * num_beams, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using any class inheriting from :class:`~transformers.PretrainedTokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        final_beam_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_beams)`):
+            The final scores of all non-finished beams.
+        final_beam_tokens (:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_beams)`):
+            The last tokens to be added to the non-finished beam_hypotheses.
+        final_beam_indices (:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_beams)`):
+            The beam indices indicating to which beam the :obj:`final_beam_tokens` shall be added.
+        pad_token_id (:obj:`int`, `optional`):
+            The id of the `padding` token.
+        eos_token_id (:obj:`int`, `optional`):
+            The id of the `end-of-sequence` token.
+
+    Return:
+        :obj:`torch.LongTensor` of shape :obj:`(batch_size * num_return_sequences, sequence_length)`: The generated
+        sequences. The second dimension (sequence_length) is either equal to :obj:`max_length` or shorter if all
+        batches finished early due to the :obj:`eos_token_id`.
+
+"""
+
+
+class BeamScorer(ABC):
+    """
+    Abstract base class for all beam scorers that are used for :meth:`~transformers.PretrainedModel.beam_search` and
+    :meth:`~transformers.PretrainedModel.beam_sample`.
+    """
+
+    @abstractmethod
+    @add_start_docstrings(PROCESS_INPUTS_DOCSTRING)
+    def process(
+        self,
+        input_ids: torch.LongTensor,
+        next_scores: torch.FloatTensor,
+        next_tokens: torch.LongTensor,
+        next_indices: torch.LongTensor,
+        **kwargs
+    ) -> Tuple[torch.Tensor]:
+        raise NotImplementedError("This is an abstract method.")
+
+    @abstractmethod
+    @add_start_docstrings(FINALIZE_INPUTS_DOCSTRING)
+    def finalize(
+        self,
+        input_ids: torch.LongTensor,
+        next_scores: torch.FloatTensor,
+        next_tokens: torch.LongTensor,
+        next_indices: torch.LongTensor,
+        **kwargs
+    ) -> torch.LongTensor:
+        raise NotImplementedError("This is an abstract method.")
+
+
+class BeamSearchScorer(BeamScorer):
+    r"""
+    :class:`transformers.BeamScorer` implementing standard beam search decoding.
+
+    Adapted in part from `Facebook's XLM beam search code
+    <https://github.com/facebookresearch/XLM/blob/9e6f6814d17be4fe5b15f2e6c43eb2b2d76daeb4/src/model/transformer.py#L529>`__.
+
+    Reference for the diverse beam search algorithm and implementation `Ashwin Kalyan's DBS implementation
+    <https://github.com/ashwinkalyan/dbs/blob/master/dbs/beam_utils.lua>`__
+
+    Args:
+        batch_size (:obj:`int`):
+            Batch Size of :obj:`input_ids` for which standard beam search decoding is run in parallel.
+        max_length (:obj:`int`):
+            The maximum length of the sequence to be generated.
+        num_beams (:obj:`int`):
+            Number of beams for beam search.
+        device (:obj:`torch.device`):
+            Defines the device type (*e.g.*, :obj:`"cpu"` or :obj:`"cuda"`) on which this instance of
+            :obj:`BeamSearchScorer` will be allocated.
+        length_penalty (:obj:`float`, `optional`, defaults to 1.0):
+            Exponential penalty to the length. 1.0 means no penalty. Set to values < 1.0 in order to encourage the
+            model to generate shorter sequences, to a value > 1.0 in order to encourage the model to produce longer
+            sequences.
+        do_early_stopping (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether to stop the beam search when at least ``num_beams`` sentences are finished per batch or not.
+        num_beam_hyps_to_keep (:obj:`int`, `optional`, defaults to 1):
+            The number of beam hypotheses that shall be returned upon calling
+            :meth:`~transformer.BeamSearchScorer.finalize`.
+        num_beam_groups (:obj:`int`):
+            Number of groups to divide :obj:`num_beams` into in order to ensure diversity among different groups of
+            beams. See `this paper <https://arxiv.org/pdf/1610.02424.pdf>`__ for more details.
+    """
+
+    def __init__(
+        self,
+        batch_size: int,
+        max_length: int,
+        num_beams: int,
+        device: torch.device,
+        length_penalty: Optional[float] = 1.0,
+        do_early_stopping: Optional[bool] = False,
+        num_beam_hyps_to_keep: Optional[int] = 1,
+        num_beam_groups: Optional[int] = 1,
+    ):
+        self.max_length = max_length
+        self.num_beams = num_beams
+        self.device = device
+        self.length_penalty = length_penalty
+        self.do_early_stopping = do_early_stopping
+        self.num_beam_hyps_to_keep = num_beam_hyps_to_keep
+        self.num_beam_groups = num_beam_groups
+        self.group_size = self.num_beams // self.num_beam_groups
+
+        self._is_init = False
+        self._beam_hyps = [
+            BeamHypotheses(
+                num_beams=self.num_beams,
+                max_length=self.max_length,
+                length_penalty=self.length_penalty,
+                early_stopping=self.do_early_stopping,
+            )
+            for _ in range(batch_size)
+        ]
+        self._done = torch.tensor([False for _ in range(batch_size)], dtype=torch.bool, device=self.device)
+
+        if not isinstance(num_beams, int) or num_beams <= 1:
+            raise ValueError(
+                f"`num_beams` has to be an integer strictly greater than 1, but is {num_beams}. For `num_beams` == 1, one should make use of `greedy_search` instead."
+            )
+
+        if not isinstance(num_beam_groups, int) or (num_beam_groups > num_beams) or (num_beams % num_beam_groups != 0):
+            raise ValueError(
+                f"`num_beam_groups` has to be an integer smaller or equal than `num_beams` and `num_beams` "
+                f"has to be divisible by `num_beam_groups`, but is {num_beam_groups} with `num_beams` being {num_beams}."
+            )
+
+    @property
+    def is_done(self) -> bool:
+        return self._done.all()
+
+    def process(
+        self,
+        input_ids: torch.LongTensor,
+        next_scores: torch.FloatTensor,
+        next_tokens: torch.LongTensor,
+        next_indices: torch.LongTensor,
+        pad_token_id: Optional[int] = None,
+        eos_token_id: Optional[int] = None,
+    ) -> Tuple[torch.Tensor]:
+        cur_len = input_ids.shape[-1]
+        batch_size = len(self._beam_hyps)
+        assert batch_size == (input_ids.shape[0] // self.group_size)
+
+        device = input_ids.device
+        next_beam_scores = torch.zeros((batch_size, self.group_size), dtype=next_scores.dtype, device=device)
+        next_beam_tokens = torch.zeros((batch_size, self.group_size), dtype=next_tokens.dtype, device=device)
+        next_beam_indices = torch.zeros((batch_size, self.group_size), dtype=next_indices.dtype, device=device)
+
+        for batch_idx, beam_hyp in enumerate(self._beam_hyps):
+            if self._done[batch_idx]:
+                assert (
+                    len(beam_hyp) >= self.num_beams
+                ), "Batch can only be done if at least {} beams have been generated".format(self.num_beams)
+                assert (
+                    eos_token_id is not None and pad_token_id is not None
+                ), "generated beams >= num_beams -> eos_token_id and pad_token have to be defined"
+                # pad the batch
+                next_beam_scores[batch_idx, :] = 0
+                next_beam_tokens[batch_idx, :] = pad_token_id
+                next_beam_indices[batch_idx, :] = 0
+                continue
+
+            # next tokens for this sentence
+            beam_idx = 0
+            for beam_token_rank, (next_token, next_score, next_index) in enumerate(
+                zip(next_tokens[batch_idx], next_scores[batch_idx], next_indices[batch_idx])
+            ):
+                batch_beam_idx = batch_idx * self.group_size + next_index
+                # add to generated hypotheses if end of sentence
+                if (eos_token_id is not None) and (next_token.item() == eos_token_id):
+                    # if beam_token does not belong to top num_beams tokens, it should not be added
+                    is_beam_token_worse_than_top_num_beams = beam_token_rank >= self.group_size
+                    if is_beam_token_worse_than_top_num_beams:
+                        continue
+                    beam_hyp.add(
+                        input_ids[batch_beam_idx].clone(),
+                        next_score.item(),
+                    )
+                else:
+                    # add next predicted token since it is not eos_token
+                    next_beam_scores[batch_idx, beam_idx] = next_score
+                    next_beam_tokens[batch_idx, beam_idx] = next_token
+                    next_beam_indices[batch_idx, beam_idx] = batch_beam_idx
+                    beam_idx += 1
+
+                # once the beam for next step is full, don't add more tokens to it.
+                if beam_idx == self.group_size:
+                    break
+
+            if beam_idx < self.group_size:
+                raise ValueError(
+                    f"At most {self.group_size} tokens in {next_tokens[batch_idx]} can be equal to `eos_token_id: {eos_token_id}`. Make sure {next_tokens[batch_idx]} are corrected."
+                )
+
+            # Check if we are done so that we can save a pad step if all(done)
+            self._done[batch_idx] = self._done[batch_idx] or beam_hyp.is_done(
+                next_scores[batch_idx].max().item(), cur_len
+            )
+
+        return UserDict(
+            {
+                "next_beam_scores": next_beam_scores.view(-1),
+                "next_beam_tokens": next_beam_tokens.view(-1),
+                "next_beam_indices": next_beam_indices.view(-1),
+            }
+        )
+
+    def finalize(
+        self,
+        input_ids: torch.LongTensor,
+        final_beam_scores: torch.FloatTensor,
+        final_beam_tokens: torch.LongTensor,
+        final_beam_indices: torch.LongTensor,
+        pad_token_id: Optional[int] = None,
+        eos_token_id: Optional[int] = None,
+    ) -> Tuple[torch.LongTensor]:
+        batch_size = len(self._beam_hyps)
+
+        # finalize all open beam hypotheses and add to generated hypotheses
+        for batch_idx, beam_hyp in enumerate(self._beam_hyps):
+            if self._done[batch_idx]:
+                continue
+
+            # all open beam hypotheses are added to the beam hypothesis
+            # beam hypothesis class automatically keeps the best beams
+            for beam_id in range(self.num_beams):
+                batch_beam_idx = batch_idx * self.num_beams + beam_id
+                final_score = final_beam_scores[batch_beam_idx].item()
+                final_tokens = input_ids[batch_beam_idx]
+                beam_hyp.add(final_tokens, final_score)
+
+        # select the best hypotheses
+        sent_lengths = input_ids.new(batch_size * self.num_beam_hyps_to_keep)
+        best = []
+        best_scores = torch.zeros(batch_size * self.num_beam_hyps_to_keep, device=self.device, dtype=torch.float32)
+
+        # retrieve best hypotheses
+        for i, beam_hyp in enumerate(self._beam_hyps):
+            sorted_hyps = sorted(beam_hyp.beams, key=lambda x: x[0])
+            for j in range(self.num_beam_hyps_to_keep):
+                best_hyp_tuple = sorted_hyps.pop()
+                best_score = best_hyp_tuple[0]
+                best_hyp = best_hyp_tuple[1]
+                sent_lengths[self.num_beam_hyps_to_keep * i + j] = len(best_hyp)
+
+                # append to lists
+                best.append(best_hyp)
+                best_scores[i * self.num_beam_hyps_to_keep + j] = best_score
+
+        # prepare for adding eos
+        sent_max_len = min(sent_lengths.max().item() + 1, self.max_length)
+        decoded: torch.LongTensor = input_ids.new(batch_size * self.num_beam_hyps_to_keep, sent_max_len)
+        # shorter batches are padded if needed
+        if sent_lengths.min().item() != sent_lengths.max().item():
+            assert pad_token_id is not None, "`pad_token_id` has to be defined"
+            decoded.fill_(pad_token_id)
+
+        # fill with hypotheses and eos_token_id if the latter fits in
+        for i, hypo in enumerate(best):
+            decoded[i, : sent_lengths[i]] = hypo
+            if sent_lengths[i] < self.max_length:
+                decoded[i, sent_lengths[i]] = eos_token_id
+        return UserDict(
+            {
+                "sequences": decoded,
+                "sequence_scores": best_scores,
+            }
+        )
+
+
+class BeamHypotheses:
+    def __init__(self, num_beams: int, max_length: int, length_penalty: float, early_stopping: bool):
+        """
+        Initialize n-best list of hypotheses.
+        """
+        self.max_length = max_length - 1  # ignoring bos_token
+        self.length_penalty = length_penalty
+        self.early_stopping = early_stopping
+        self.num_beams = num_beams
+        self.beams = []
+        self.worst_score = 1e9
+
+    def __len__(self):
+        """
+        Number of hypotheses in the list.
+        """
+        return len(self.beams)
+
+    def add(self, hyp: torch.LongTensor, sum_logprobs: float):
+        """
+        Add a new hypothesis to the list.
+        """
+        score = sum_logprobs / (hyp.shape[-1] ** self.length_penalty)
+        if len(self) < self.num_beams or score > self.worst_score:
+            self.beams.append((score, hyp))
+            if len(self) > self.num_beams:
+                sorted_next_scores = sorted([(s, idx) for idx, (s, _) in enumerate(self.beams)])
+                del self.beams[sorted_next_scores[0][1]]
+                self.worst_score = sorted_next_scores[1][0]
+            else:
+                self.worst_score = min(score, self.worst_score)
+
+    def is_done(self, best_sum_logprobs: float, cur_len: int) -> bool:
+        """
+        If there are enough hypotheses and that none of the hypotheses being generated can become better than the worst
+        one in the heap, then we are done with this sentence.
+        """
+
+        if len(self) < self.num_beams:
+            return False
+        elif self.early_stopping:
+            return True
+        else:
+            cur_score = best_sum_logprobs / cur_len ** self.length_penalty
+            ret = self.worst_score >= cur_score
+            return ret

generation_logits_process.py

+# coding=utf-8
+# Copyright 2020 The HuggingFace Inc. team
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import math
+from abc import ABC
+from typing import Callable, Iterable, List
+
+import numpy as np
+import torch
+
+from file_utils import add_start_docstrings
+
+
+LOGITS_PROCESSOR_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using :class:`~transformers.BertTokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.vocab_size)`):
+            Prediction scores of a language modeling head. These can be scores for each vocabulary token before SoftMax
+            or scores for each vocabulary token after SoftMax.
+        kwargs:
+            Additional logits processor specific kwargs.
+
+    Return:
+        :obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.vocab_size)`: The processed prediction scores.
+
+"""
+
+
+class LogitsProcessor(ABC):
+    """Abstract base class for all logit processors that can be applied during generation."""
+
+    @add_start_docstrings(LOGITS_PROCESSOR_INPUTS_DOCSTRING)
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        """Torch method for processing logits."""
+        raise NotImplementedError(
+            f"{self.__class__} is an abstract class. Only classes inheriting this class can be called."
+        )
+
+
+class LogitsWarper(ABC):
+    """Abstract base class for all logit warpers that can be applied during generation with multinomial sampling."""
+
+    @add_start_docstrings(LOGITS_PROCESSOR_INPUTS_DOCSTRING)
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        """Torch method for warping logits."""
+        raise NotImplementedError(
+            f"{self.__class__} is an abstract class. Only classes inheriting this class can be called."
+        )
+
+
+class LogitsProcessorList(list):
+    """
+    This class can be used to create a list of :class:`~transformers.LogitsProcessor` or
+    :class:`~transformers.LogitsWarper` to subsequently process a :obj:`scores` input tensor. This class inherits from
+    list and adds a specific `__call__` method to apply each :class:`~transformers.LogitsProcessor` or
+    :class:`~transformers.LogitsProcessor` to the inputs.
+    """
+
+    @add_start_docstrings(LOGITS_PROCESSOR_INPUTS_DOCSTRING)
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> torch.FloatTensor:
+        for processor in self:
+            function_args = inspect.signature(processor.__call__).parameters
+            if len(function_args) > 2:
+                assert all(
+                    arg in kwargs for arg in list(function_args.keys())[2:]
+                ), f"Make sure that all the required parameters: {list(function_args.keys())} for {processor.__class__} are passed to the logits processor."
+                scores = processor(input_ids, scores, **kwargs)
+            else:
+                scores = processor(input_ids, scores)
+        return scores
+
+
+class MinLengthLogitsProcessor(LogitsProcessor):
+    r"""
+    :class:`transformers.LogitsProcessor` enforcing a min-length by setting EOS probability to 0.
+
+    Args:
+        min_length (:obj:`int`):
+            The minimum length below which the score of :obj:`eos_token_id` is set to :obj:`-float("Inf")`.
+        eos_token_id (:obj:`int`):
+            The id of the `end-of-sequence` token.
+    """
+
+    def __init__(self, min_length: int, eos_token_id: int):
+        if not isinstance(min_length, int) or min_length < 0:
+            raise ValueError(f"`min_length` has to be a positive integer, but is {min_length}")
+
+        if not isinstance(eos_token_id, int) or eos_token_id < 0:
+            raise ValueError(f"`eos_token_id` has to be a positive integer, but is {eos_token_id}")
+
+        self.min_length = min_length
+        self.eos_token_id = eos_token_id
+
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        cur_len = input_ids.shape[-1]
+        if cur_len < self.min_length:
+            scores[:, self.eos_token_id] = -float("inf")
+        return scores
+
+
+class TemperatureLogitsWarper(LogitsWarper):
+    r"""
+    :class:`transformers.LogitsWarper` for temperature (exponential scaling output probability distribution).
+
+    Args:
+        temperature (:obj:`float`):
+            The value used to module the logits distribution.
+    """
+
+    def __init__(self, temperature: float):
+        if not isinstance(temperature, float) or not (temperature > 0):
+            raise ValueError(f"`temperature` has to be a strictly positive float, but is {temperature}")
+
+        self.temperature = temperature
+
+    def __call__(self, input_ids: torch.Tensor, scores: torch.Tensor) -> torch.Tensor:
+        scores = scores / self.temperature
+        return scores
+
+
+class RepetitionPenaltyLogitsProcessor(LogitsProcessor):
+    r"""
+    :class:`transformers.LogitsProcessor` enforcing an exponential penalty on repeated sequences.
+
+    Args:
+        repetition_penalty (:obj:`float`):
+            The parameter for repetition penalty. 1.0 means no penalty. See `this paper
+            <https://arxiv.org/pdf/1909.05858.pdf>`__ for more details.
+    """
+
+    def __init__(self, penalty: float):
+        if not isinstance(penalty, float) or not (penalty > 0):
+            raise ValueError(f"`penalty` has to be a strictly positive float, but is {penalty}")
+
+        self.penalty = penalty
+
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        score = torch.gather(scores, 1, input_ids)
+
+        # if score < 0 then repetition penalty has to be multiplied to reduce the previous token probability
+        score = torch.where(score < 0, score * self.penalty, score / self.penalty)
+
+        scores.scatter_(1, input_ids, score)
+        return scores
+
+
+class TopPLogitsWarper(LogitsWarper):
+    """
+    :class:`transformers.LogitsWarper` that performs top-p, i.e. restricting to top tokens summing to prob_cut_off <=
+    prob_cut_off.
+
+    Args:
+        top_p (:obj:`float`):
+            If set to < 1, only the most probable tokens with probabilities that add up to :obj:`top_p` or higher are
+            kept for generation.
+        filter_value (:obj:`float`, `optional`, defaults to :obj:`-float("Inf")`):
+            All filtered values will be set to this float value.
+        min_tokens_to_keep (:obj:`int`, `optional`, defaults to 1):
+            Minimum number of tokens that cannot be filtered.
+    """
+
+    def __init__(self, top_p: float, filter_value: float = -float("Inf"), min_tokens_to_keep: int = 1):
+        if not isinstance(top_p, float) or (top_p < 0 or top_p > 1.0):
+            raise ValueError(f"`top_p` has to be a float > 0 and < 1, but is {top_p}")
+
+        self.top_p = top_p
+        self.filter_value = filter_value
+        self.min_tokens_to_keep = min_tokens_to_keep
+
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        sorted_logits, sorted_indices = torch.sort(scores, descending=True)
+        cumulative_probs = sorted_logits.softmax(dim=-1).cumsum(dim=-1)
+
+        # Remove tokens with cumulative top_p above the threshold (token with 0 are kept)
+        sorted_indices_to_remove = cumulative_probs > self.top_p
+        if self.min_tokens_to_keep > 1:
+            # Keep at least min_tokens_to_keep (set to min_tokens_to_keep-1 because we add the first one below)
+            sorted_indices_to_remove[..., : self.min_tokens_to_keep - 1] = 0
+        # Shift the indices to the right to keep also the first token above the threshold
+        sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
+        sorted_indices_to_remove[..., 0] = 0
+
+        # scatter sorted tensors to original indexing
+        indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
+        scores = scores.masked_fill(indices_to_remove, self.filter_value)
+        return scores
+
+
+class TopKLogitsWarper(LogitsWarper):
+    r"""
+    :class:`transformers.LogitsWarper` that performs top-k, i.e. restricting to the k highest probability elements.
+
+    Args:
+        top_k (:obj:`int`):
+            The number of highest probability vocabulary tokens to keep for top-k-filtering.
+        filter_value (:obj:`float`, `optional`, defaults to :obj:`-float("Inf")`):
+            All filtered values will be set to this float value.
+        min_tokens_to_keep (:obj:`int`, `optional`, defaults to 1):
+            Minimum number of tokens that cannot be filtered.
+    """
+
+    def __init__(self, top_k: int, filter_value: float = -float("Inf"), min_tokens_to_keep: int = 1):
+        if not isinstance(top_k, int) or top_k <= 0:
+            raise ValueError(f"`top_k` has to be a strictly positive integer, but is {top_k}")
+
+        self.top_k = top_k
+        self.filter_value = filter_value
+        self.min_tokens_to_keep = min_tokens_to_keep
+
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        top_k = min(max(self.top_k, self.min_tokens_to_keep), scores.size(-1))  # Safety check
+        # Remove all tokens with a probability less than the last token of the top-k
+        indices_to_remove = scores < torch.topk(scores, top_k)[0][..., -1, None]
+        scores = scores.masked_fill(indices_to_remove, self.filter_value)
+        return scores
+
+
+def _get_ngrams(ngram_size: int, prev_input_ids: torch.Tensor, num_hypos: int):
+    generated_ngrams = [{} for _ in range(num_hypos)]
+    for idx in range(num_hypos):
+        gen_tokens = prev_input_ids[idx].tolist()
+        generated_ngram = generated_ngrams[idx]
+        for ngram in zip(*[gen_tokens[i:] for i in range(ngram_size)]):
+            prev_ngram_tuple = tuple(ngram[:-1])
+            generated_ngram[prev_ngram_tuple] = generated_ngram.get(prev_ngram_tuple, []) + [ngram[-1]]
+    return generated_ngrams
+
+
+def _get_generated_ngrams(banned_ngrams, prev_input_ids, ngram_size, cur_len):
+    # Before decoding the next token, prevent decoding of ngrams that have already appeared
+    start_idx = cur_len + 1 - ngram_size
+    ngram_idx = tuple(prev_input_ids[start_idx:cur_len].tolist())
+    return banned_ngrams.get(ngram_idx, [])
+
+
+def _calc_banned_ngram_tokens(
+    ngram_size: int, prev_input_ids: torch.Tensor, num_hypos: int, cur_len: int
+) -> List[Iterable[int]]:
+    """Copied from fairseq for no_repeat_ngram in beam_search"""
+    if cur_len + 1 < ngram_size:
+        # return no banned tokens if we haven't generated no_repeat_ngram_size tokens yet
+        return [[] for _ in range(num_hypos)]
+
+    generated_ngrams = _get_ngrams(ngram_size, prev_input_ids, num_hypos)
+
+    banned_tokens = [
+        _get_generated_ngrams(generated_ngrams[hypo_idx], prev_input_ids[hypo_idx], ngram_size, cur_len)
+        for hypo_idx in range(num_hypos)
+    ]
+    return banned_tokens
+
+
+class NoRepeatNGramLogitsProcessor(LogitsProcessor):
+    r"""
+    :class:`transformers.LogitsProcessor` that enforces no repetition of n-grams. See `Fairseq
+    <https://github.com/pytorch/fairseq/blob/a07cb6f40480928c9e0548b737aadd36ee66ac76/fairseq/sequence_generator.py#L345>`__.
+
+    Args:
+        ngram_size (:obj:`int`):
+            All ngrams of size :obj:`ngram_size` can only occur once.
+    """
+
+    def __init__(self, ngram_size: int):
+        if not isinstance(ngram_size, int) or ngram_size <= 0:
+            raise ValueError(f"`ngram_size` has to be a strictly positive integer, but is {ngram_size}")
+        self.ngram_size = ngram_size
+
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        num_batch_hypotheses = scores.shape[0]
+        cur_len = input_ids.shape[-1]
+        banned_batch_tokens = _calc_banned_ngram_tokens(self.ngram_size, input_ids, num_batch_hypotheses, cur_len)
+
+        for i, banned_tokens in enumerate(banned_batch_tokens):
+            scores[i, banned_tokens] = -float("inf")
+
+        return scores
+
+
+class EncoderNoRepeatNGramLogitsProcessor(LogitsProcessor):
+    r"""
+    :class:`transformers.LogitsProcessor` that enforces no repetition of encoder input ids n-grams for the decoder ids.
+    See `ParlAI <https://github.com/facebookresearch/ParlAI/blob/master/parlai/core/torch_generator_agent.py#L1350>`__.
+
+    Args:
+        encoder_ngram_size (:obj:`int`):
+            All ngrams of size :obj:`ngram_size` can only occur within the encoder input ids.
+        encoder_input_ids (:obj:`int`):
+            The encoder_input_ids that should not be repeated within the decoder ids.
+    """
+
+    def __init__(self, encoder_ngram_size: int, encoder_input_ids: torch.LongTensor):
+        if not isinstance(encoder_ngram_size, int) or encoder_ngram_size <= 0:
+            raise ValueError(
+                f"`encoder_ngram_size` has to be a strictly positive integer, but is {encoder_ngram_size}"
+            )
+        self.ngram_size = encoder_ngram_size
+        if len(encoder_input_ids.shape) == 1:
+            encoder_input_ids = encoder_input_ids.unsqueeze(0)
+        self.batch_size = encoder_input_ids.shape[0]
+        self.generated_ngrams = _get_ngrams(encoder_ngram_size, encoder_input_ids, self.batch_size)
+
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        # B x num_beams
+        num_hypos = scores.shape[0]
+        num_beams = num_hypos // self.batch_size
+        cur_len = input_ids.shape[-1]
+        banned_batch_tokens = [
+            _get_generated_ngrams(
+                self.generated_ngrams[hypo_idx // num_beams], input_ids[hypo_idx], self.ngram_size, cur_len
+            )
+            for hypo_idx in range(num_hypos)
+        ]
+
+        for i, banned_tokens in enumerate(banned_batch_tokens):
+            scores[i, banned_tokens] = -float("inf")
+
+        return scores
+
+
+class NoBadWordsLogitsProcessor(LogitsProcessor):
+    """
+    :class:`transformers.LogitsProcessor` that enforces that specified sequences will never be sampled.
+
+    Args:
+        bad_words_ids (:obj:`List[List[int]]`):
+            List of list of token ids that are not allowed to be generated. In order to get the tokens of the words
+            that should not appear in the generated text, use :obj:`tokenizer(bad_word,
+            add_prefix_space=True).input_ids`.
+        eos_token_id (:obj:`int`):
+            The id of the `end-of-sequence` token.
+    """
+
+    def __init__(self, bad_words_ids: Iterable[Iterable[int]], eos_token_id: int):
+
+        if not isinstance(bad_words_ids, List) or len(bad_words_ids) == 0:
+            raise ValueError(f"`bad_words_ids` has to be a non-emtpy list, but is {bad_words_ids}.")
+        if any(not isinstance(bad_word_ids, list) for bad_word_ids in bad_words_ids):
+            raise ValueError(f"`bad_words_ids` has to be a list of lists, but is {bad_words_ids}.")
+        if any(
+            any((not isinstance(token_id, (int, np.integer)) or token_id < 0) for token_id in bad_word_ids)
+            for bad_word_ids in bad_words_ids
+        ):
+            raise ValueError(
+                f"Each list in `bad_words_ids` has to be a list of positive integers, but is {bad_words_ids}."
+            )
+
+        self.bad_words_ids = list(filter(lambda bad_token_seq: bad_token_seq != [eos_token_id], bad_words_ids))
+
+        for banned_token_seq in self.bad_words_ids:
+            assert len(banned_token_seq) > 0, "Banned words token sequences {} cannot have an empty list".format(
+                bad_words_ids
+            )
+
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        banned_tokens = self._calc_banned_bad_words_ids(input_ids)
+        scores = self._set_scores_to_inf_for_banned_tokens(scores, banned_tokens)
+
+        return scores
+
+    def _tokens_match(self, prev_tokens: torch.LongTensor, tokens: List[int]) -> bool:
+        if len(tokens) == 0:
+            # if bad word tokens is just one token always ban it
+            return True
+        elif len(tokens) > len(prev_tokens):
+            # if bad word tokens are longer then prev input_ids they can't be equal
+            return False
+        elif prev_tokens[-len(tokens) :].tolist() == tokens:
+            # if tokens match
+            return True
+        else:
+            return False
+
+    def _calc_banned_bad_words_ids(self, prev_input_ids: Iterable[int]) -> Iterable[int]:
+        banned_tokens = []
+        for prev_input_ids_slice in prev_input_ids:
+            banned_tokens_slice = []
+            for banned_token_seq in self.bad_words_ids:
+                if self._tokens_match(prev_input_ids_slice, banned_token_seq[:-1]) is False:
+                    # if tokens do not match continue
+                    continue
+
+                banned_tokens_slice.append(banned_token_seq[-1])
+
+            banned_tokens.append(banned_tokens_slice)
+
+        return banned_tokens
+
+    def _set_scores_to_inf_for_banned_tokens(self, scores: torch.Tensor, banned_tokens: List[List[int]]) -> None:
+        """
+        Modifies the scores in place by setting the banned token positions to `-inf`. Banned token is expected to be a
+        list of list of banned tokens to ban in the format [[batch index, vocabulary position],...
+
+        Args:
+            scores: logits distribution of shape (batch size, vocabulary size)
+            banned_tokens: list of list of tokens to ban of length (batch_size)
+        """
+        banned_mask_list = []
+        for idx, batch_banned_tokens in enumerate(banned_tokens):
+            for token in batch_banned_tokens:
+                banned_mask_list.append([idx, token])
+        if not banned_mask_list:
+            return scores
+
+        banned_mask = torch.LongTensor(banned_mask_list)
+        indices = torch.ones(len(banned_mask))
+        # A sparse tensor is generated from a list of coordinates: [[0, 1], [0, 2], [2, 0]]. A conversion to dense tensor generates:
+        # [ 0  1  1 ]
+        # [ 0  0  0 ]
+        # [ 1  0  0 ]
+
+        banned_mask = (
+            torch.sparse.LongTensor(banned_mask.t(), indices, scores.size()).to(scores.device).to_dense().bool()
+        )
+        scores = scores.masked_fill(banned_mask, -float("inf"))
+        return scores
+
+
+class PrefixConstrainedLogitsProcessor(LogitsProcessor):
+    r"""
+    :class:`transformers.LogitsProcessor` that enforces contrained generation and is useful for prefix-conditioned
+    constrained generation. See `Autoregressive Entity Retrieval <https://arxiv.org/abs/2010.00904>`__ for more
+    information.
+
+    Args:
+        prefix_allowed_tokens_fn: (:obj:`Callable[[int, torch.Tensor], List[int]]`):
+            This function constraints the beam search to allowed tokens only at each step. This function takes 2
+            arguments :obj:`inputs_ids` and the batch ID :obj:`batch_id`. It has to return a list with the allowed
+            tokens for the next generation step conditioned on the previously generated tokens :obj:`inputs_ids` and
+            the batch ID :obj:`batch_id`.
+    """
+
+    def __init__(self, prefix_allowed_tokens_fn: Callable[[int, torch.Tensor], List[int]], num_beams: int):
+        self._prefix_allowed_tokens_fn = prefix_allowed_tokens_fn
+        self._num_beams = num_beams
+
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        mask = torch.full_like(scores, -math.inf)
+        for batch_id, beam_sent in enumerate(input_ids.view(-1, self._num_beams, input_ids.shape[-1])):
+            for beam_id, sent in enumerate(beam_sent):
+                mask[batch_id * self._num_beams + beam_id, self._prefix_allowed_tokens_fn(batch_id, sent)] = 0
+
+        return scores + mask
+
+
+class HammingDiversityLogitsProcessor(LogitsProcessor):
+    r"""
+    :class:`transformers.LogitsProcessor` that enforces diverse beam search. Note that this logits processor is only
+    effective for :meth:`transformers.PretrainedModel.group_beam_search`. See `Diverse Beam Search: Decoding Diverse
+    Solutions from Neural Sequence Models <https://arxiv.org/pdf/1610.02424.pdf>`__ for more details.
+
+    Args:
+        diversity_penalty (:obj:`float`):
+            This value is subtracted from a beam's score if it generates a token same as any beam from other group at a
+            particular time. Note that :obj:`diversity_penalty` is only effective if ``group beam search`` is enabled.
+        num_beams (:obj:`int`):
+            Number of beams used for group beam search. See `this paper <https://arxiv.org/pdf/1610.02424.pdf>`__ for
+            more details.
+        num_beam_groups (:obj:`int`):
+            Number of groups to divide :obj:`num_beams` into in order to ensure diversity among different groups of
+            beams. See `this paper <https://arxiv.org/pdf/1610.02424.pdf>`__ for more details.
+    """
+
+    def __init__(self, diversity_penalty: float, num_beams: int, num_beam_groups: int):
+        if not isinstance(diversity_penalty, float) or (not diversity_penalty > 0.0):
+            raise ValueError("`diversity_penalty` should be a float strictly larger than 0.")
+        self._diversity_penalty = diversity_penalty
+        if not isinstance(num_beams, int) or num_beams < 2:
+            raise ValueError("`num_beams` should be an integer strictly larger than 1.")
+        self._num_beams = num_beams
+        if not isinstance(num_beam_groups, int) or num_beam_groups < 2:
+            raise ValueError("`num_beam_groups` should be an integer strictly larger than 1.")
+        if num_beam_groups > num_beams:
+            raise ValueError("`beam_groups` has to be smaller or equal to `num_beams`.")
+        self._num_sub_beams = num_beams // num_beam_groups
+
+    def __call__(
+        self,
+        input_ids: torch.LongTensor,
+        scores: torch.FloatTensor,
+        current_tokens: torch.LongTensor,
+        beam_group_idx: int,
+    ) -> torch.FloatTensor:
+        # hamming diversity: penalise using same token in current group which was used in previous groups at
+        # the same time step
+        batch_size = current_tokens.shape[0] // self._num_beams
+        group_start_idx = beam_group_idx * self._num_sub_beams
+        group_end_idx = min(group_start_idx + self._num_sub_beams, self._num_beams)
+        group_size = group_end_idx - group_start_idx
+        vocab_size = scores.shape[-1]
+
+        if group_start_idx == 0:
+            return scores
+
+        for batch_idx in range(batch_size):
+            # predicted tokens of last time step of previous groups
+            previous_group_tokens = current_tokens[
+                batch_idx * self._num_beams : batch_idx * self._num_beams + group_start_idx
+            ]
+            token_frequency = torch.bincount(previous_group_tokens, minlength=vocab_size).to(scores.device)
+            scores[batch_idx * group_size : (batch_idx + 1) * group_size] -= self._diversity_penalty * token_frequency
+
+        return scores
+
+
+class ForcedBOSTokenLogitsProcessor(LogitsProcessor):
+    r"""
+    :class:`~transformers.LogitsProcessor` that enforces the specified token as the first generated token.
+
+    Args:
+        bos_token_id (:obj:`int`):
+            The id of the token to force as the first generated token.
+    """
+
+    def __init__(self, bos_token_id: int):
+        self.bos_token_id = bos_token_id
+
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        cur_len = input_ids.shape[-1]
+        if cur_len == 1:
+            num_tokens = scores.shape[1]
+            scores[:, [i for i in range(num_tokens) if i != self.bos_token_id]] = -float("inf")
+            scores[:, self.bos_token_id] = 0
+        return scores
+
+
+class ForcedEOSTokenLogitsProcessor(LogitsProcessor):
+    r"""
+    :class:`~transformers.LogitsProcessor` that enforces the specified token as the last generated token when
+    :obj:`max_length` is reached.
+
+    Args:
+        max_length (:obj:`int`):
+            The maximum length of the sequence to be generated.
+        eos_token_id (:obj:`int`):
+            The id of the token to force as the last generated token when :obj:`max_length` is reached.
+    """
+
+    def __init__(self, max_length: int, eos_token_id: int):
+        self.max_length = max_length
+        self.eos_token_id = eos_token_id
+
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        cur_len = input_ids.shape[-1]
+        if cur_len == self.max_length - 1:
+            num_tokens = scores.shape[1]
+            scores[:, [i for i in range(num_tokens) if i != self.eos_token_id]] = -float("inf")
+            scores[:, self.eos_token_id] = 0
+        return scores

generation_stopping_criteria.py

+import time
+import warnings
+from abc import ABC
+from typing import Optional
+
+import torch
+
+from file_utils import add_start_docstrings
+
+
+LOGITS_PROCESSOR_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using :class:`~transformers.BertTokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.vocab_size)`):
+            Prediction scores of a language modeling head. These can be scores for each vocabulary token before SoftMax
+            or scores for each vocabulary token after SoftMax.
+        kwargs:
+            Additional stopping critera specific kwargs.
+
+    Return:
+        :obj:`bool`. :obj:`False` indicates we should continue, :obj:`True` indicates we should stop.
+
+"""
+
+
+class StoppingCriteria(ABC):
+    """Abstract base class for all stopping criteria that can be applied during generation."""
+
+    @add_start_docstrings(LOGITS_PROCESSOR_INPUTS_DOCSTRING)
+    def __call__(self, input_ids: torch.LongTensor, score: torch.FloatTensor, **kwargs) -> bool:
+        raise NotImplementedError("StoppingCriteria needs to be subclassed")
+
+
+class MaxLengthCriteria(StoppingCriteria):
+    """
+    This class can be used to stop generation whenever the full generated number of tokens exceeds :obj:`max_length`.
+    Keep in mind for decoder-only type of transformers, this will include the initial prompted tokens.
+
+    Args:
+        max_length (:obj:`int`):
+            The maximum length that the output sequence can have in number of tokens.
+    """
+
+    def __init__(self, max_length: int):
+        self.max_length = max_length
+
+    @add_start_docstrings(LOGITS_PROCESSOR_INPUTS_DOCSTRING)
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
+        return input_ids.shape[-1] > self.max_length
+
+
+class MaxTimeCriteria(StoppingCriteria):
+    """
+    This class can be used to stop generation whenever the full generation exceeds some amount of time. By default, the
+    time will start being counted when you initialize this function. You can override this by passing an
+    :obj:`initial_time`.
+
+    Args:
+        max_time (:obj:`float`):
+            The maximum allowed time in seconds for the generation.
+        initial_time (:obj:`float`, `optional`, defaults to :obj:`time.time()`):
+            The start of the generation allowed time.
+    """
+
+    def __init__(self, max_time: float, initial_timestamp: Optional[float] = None):
+        self.max_time = max_time
+        self.initial_timestamp = time.time() if initial_timestamp is None else initial_timestamp
+
+    @add_start_docstrings(LOGITS_PROCESSOR_INPUTS_DOCSTRING)
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
+        return time.time() - self.initial_timestamp > self.max_time
+
+
+class StoppingCriteriaList(list):
+    @add_start_docstrings(LOGITS_PROCESSOR_INPUTS_DOCSTRING)
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
+        return any(criteria(input_ids, scores) for criteria in self)
+
+
+def validate_stopping_criteria(stopping_criteria: StoppingCriteriaList, max_length: int):
+    found = False
+    for stopping_criterium in stopping_criteria:
+        if isinstance(stopping_criterium, MaxLengthCriteria):
+            found = True
+            if stopping_criterium.max_length != max_length:
+                warnings.warn(
+                    "You set different `max_length` for stopping criteria and `max_length` parameter", UserWarning
+                )
+    if not found:
+        stopping_criteria.append(MaxLengthCriteria(max_length=max_length))

modeling_bert.py

+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch BERT model. """
+
+
+import math
+import os
+import warnings
+from dataclasses import dataclass
+from typing import Optional, Tuple
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import CrossEntropyLoss, MSELoss
+
+from activations import ACT2FN
+from file_utils import (
+    ModelOutput,
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    replace_return_docstrings,
+)
+from modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    BaseModelOutputWithPoolingAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
+    MaskedLMOutput,
+    MultipleChoiceModelOutput,
+    NextSentencePredictorOutput,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutput,
+    TokenClassifierOutput,
+)
+from modeling_utils import (
+    PreTrainedModel,
+    apply_chunking_to_forward,
+    find_pruneable_heads_and_indices,
+    prune_linear_layer,
+)
+from utils import logging
+from configuration_bert import BertConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "bert-base-uncased"
+_CONFIG_FOR_DOC = "BertConfig"
+_TOKENIZER_FOR_DOC = "BertTokenizer"
+
+BERT_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "bert-base-uncased",
+    "bert-large-uncased",
+    "bert-base-cased",
+    "bert-large-cased",
+    "bert-base-multilingual-uncased",
+    "bert-base-multilingual-cased",
+    "bert-base-chinese",
+    "bert-base-german-cased",
+    "bert-large-uncased-whole-word-masking",
+    "bert-large-cased-whole-word-masking",
+    "bert-large-uncased-whole-word-masking-finetuned-squad",
+    "bert-large-cased-whole-word-masking-finetuned-squad",
+    "bert-base-cased-finetuned-mrpc",
+    "bert-base-german-dbmdz-cased",
+    "bert-base-german-dbmdz-uncased",
+    "cl-tohoku/bert-base-japanese",
+    "cl-tohoku/bert-base-japanese-whole-word-masking",
+    "cl-tohoku/bert-base-japanese-char",
+    "cl-tohoku/bert-base-japanese-char-whole-word-masking",
+    "TurkuNLP/bert-base-finnish-cased-v1",
+    "TurkuNLP/bert-base-finnish-uncased-v1",
+    "wietsedv/bert-base-dutch-cased",
+    # See all BERT models at https://huggingface.co/models?filter=bert
+]
+
+
+def load_tf_weights_in_bert(model, config, tf_checkpoint_path):
+    """Load tf checkpoints in a pytorch model."""
+    try:
+        import re
+
+        import numpy as np
+        import tensorflow as tf
+    except ImportError:
+        logger.error(
+            "Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
+            "https://www.tensorflow.org/install/ for installation instructions."
+        )
+        raise
+    tf_path = os.path.abspath(tf_checkpoint_path)
+    logger.info("Converting TensorFlow checkpoint from {}".format(tf_path))
+    # Load weights from TF model
+    init_vars = tf.train.list_variables(tf_path)
+    names = []
+    arrays = []
+    for name, shape in init_vars:
+        logger.info("Loading TF weight {} with shape {}".format(name, shape))
+        array = tf.train.load_variable(tf_path, name)
+        names.append(name)
+        arrays.append(array)
+
+    for name, array in zip(names, arrays):
+        name = name.split("/")
+        # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
+        # which are not required for using pretrained model
+        if any(
+            n in ["adam_v", "adam_m", "AdamWeightDecayOptimizer", "AdamWeightDecayOptimizer_1", "global_step"]
+            for n in name
+        ):
+            logger.info("Skipping {}".format("/".join(name)))
+            continue
+        pointer = model
+        for m_name in name:
+            if re.fullmatch(r"[A-Za-z]+_\d+", m_name):
+                scope_names = re.split(r"_(\d+)", m_name)
+            else:
+                scope_names = [m_name]
+            if scope_names[0] == "kernel" or scope_names[0] == "gamma":
+                pointer = getattr(pointer, "weight")
+            elif scope_names[0] == "output_bias" or scope_names[0] == "beta":
+                pointer = getattr(pointer, "bias")
+            elif scope_names[0] == "output_weights":
+                pointer = getattr(pointer, "weight")
+            elif scope_names[0] == "squad":
+                pointer = getattr(pointer, "classifier")
+            else:
+                try:
+                    pointer = getattr(pointer, scope_names[0])
+                except AttributeError:
+                    logger.info("Skipping {}".format("/".join(name)))
+                    continue
+            if len(scope_names) >= 2:
+                num = int(scope_names[1])
+                pointer = pointer[num]
+        if m_name[-11:] == "_embeddings":
+            pointer = getattr(pointer, "weight")
+        elif m_name == "kernel":
+            array = np.transpose(array)
+        try:
+            assert (
+                pointer.shape == array.shape
+            ), f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched"
+        except AssertionError as e:
+            e.args += (pointer.shape, array.shape)
+            raise
+        logger.info("Initialize PyTorch weight {}".format(name))
+        pointer.data = torch.from_numpy(array)
+    return model
+
+
+class BertEmbeddings(nn.Module):
+    """Construct the embeddings from word, position and token_type embeddings."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id)
+        self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.hidden_size)
+        self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.hidden_size)
+
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
+        self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
+
+    def forward(
+        self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None, past_key_values_length=0
+    ):
+        if input_ids is not None:
+            input_shape = input_ids.size()
+        else:
+            input_shape = inputs_embeds.size()[:-1]
+
+        seq_length = input_shape[1]
+
+        if position_ids is None:
+            position_ids = self.position_ids[:, past_key_values_length : seq_length + past_key_values_length]
+
+        if token_type_ids is None:
+            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+        token_type_embeddings = self.token_type_embeddings(token_type_ids)
+
+        embeddings = inputs_embeds + token_type_embeddings
+        if self.position_embedding_type == "absolute":
+            position_embeddings = self.position_embeddings(position_ids)
+            embeddings += position_embeddings
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+
+class BertSelfAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
+            raise ValueError(
+                "The hidden size (%d) is not a multiple of the number of attention "
+                "heads (%d)" % (config.hidden_size, config.num_attention_heads)
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size)
+        self.key = nn.Linear(config.hidden_size, self.all_head_size)
+        self.value = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            self.max_position_embeddings = config.max_position_embeddings
+            self.distance_embedding = nn.Embedding(2 * config.max_position_embeddings - 1, self.attention_head_size)
+
+        self.is_decoder = config.is_decoder
+
+    def transpose_for_scores(self, x):
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        x = x.view(*new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+    ):
+        mixed_query_layer = self.query(hidden_states)
+
+        # If this is instantiated as a cross-attention module, the keys
+        # and values come from an encoder; the attention mask needs to be
+        # such that the encoder's padding tokens are not attended to.
+        is_cross_attention = encoder_hidden_states is not None
+
+        if is_cross_attention and past_key_value is not None:
+            # reuse k,v, cross_attentions
+            key_layer = past_key_value[0]
+            value_layer = past_key_value[1]
+            attention_mask = encoder_attention_mask
+        elif is_cross_attention:
+            key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
+            value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
+            attention_mask = encoder_attention_mask
+        elif past_key_value is not None:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
+            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
+        else:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_layer, value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            seq_length = hidden_states.size()[1]
+            position_ids_l = torch.arange(seq_length, dtype=torch.long, device=hidden_states.device).view(-1, 1)
+            position_ids_r = torch.arange(seq_length, dtype=torch.long, device=hidden_states.device).view(1, -1)
+            distance = position_ids_l - position_ids_r
+            positional_embedding = self.distance_embedding(distance + self.max_position_embeddings - 1)
+            positional_embedding = positional_embedding.to(dtype=query_layer.dtype)  # fp16 compatibility
+
+            if self.position_embedding_type == "relative_key":
+                relative_position_scores = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores
+            elif self.position_embedding_type == "relative_key_query":
+                relative_position_scores_query = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                relative_position_scores_key = torch.einsum("bhrd,lrd->bhlr", key_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores_query + relative_position_scores_key
+
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in BertModel forward() function)
+            attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(*new_context_layer_shape)
+
+        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
+
+        if self.is_decoder:
+            outputs = outputs + (past_key_value,)
+        return outputs
+
+
+class BertSelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class BertAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.self = BertSelfAttention(config)
+        self.output = BertSelfOutput(config)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads, self.self.num_attention_heads, self.self.attention_head_size, self.pruned_heads
+        )
+
+        # Prune linear layers
+        self.self.query = prune_linear_layer(self.self.query, index)
+        self.self.key = prune_linear_layer(self.self.key, index)
+        self.self.value = prune_linear_layer(self.self.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
+        self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+    ):
+        self_outputs = self.self(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            past_key_value,
+            output_attentions,
+        )
+        attention_output = self.output(self_outputs[0], hidden_states)
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+class BertIntermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+class BertOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class BertLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+        self.attention = BertAttention(config)
+        self.is_decoder = config.is_decoder
+        self.add_cross_attention = config.add_cross_attention
+        if self.add_cross_attention:
+            assert self.is_decoder, f"{self} should be used as a decoder model if cross attention is added"
+            self.crossattention = BertAttention(config)
+        self.intermediate = BertIntermediate(config)
+        self.output = BertOutput(config)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+    ):
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
+        self_attention_outputs = self.attention(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            output_attentions=output_attentions,
+            past_key_value=self_attn_past_key_value,
+        )
+        attention_output = self_attention_outputs[0]
+
+        # if decoder, the last output is tuple of self-attn cache
+        if self.is_decoder:
+            outputs = self_attention_outputs[1:-1]
+            present_key_value = self_attention_outputs[-1]
+        else:
+            outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
+
+        cross_attn_present_key_value = None
+        if self.is_decoder and encoder_hidden_states is not None:
+            assert hasattr(
+                self, "crossattention"
+            ), f"If `encoder_hidden_states` are passed, {self} has to be instantiated with cross-attention layers by setting `config.add_cross_attention=True`"
+
+            # cross_attn cached key/values tuple is at positions 3,4 of past_key_value tuple
+            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
+            cross_attention_outputs = self.crossattention(
+                attention_output,
+                attention_mask,
+                head_mask,
+                encoder_hidden_states,
+                encoder_attention_mask,
+                cross_attn_past_key_value,
+                output_attentions,
+            )
+            attention_output = cross_attention_outputs[0]
+            outputs = outputs + cross_attention_outputs[1:-1]  # add cross attentions if we output attention weights
+
+            # add cross-attn cache to positions 3,4 of present_key_value tuple
+            cross_attn_present_key_value = cross_attention_outputs[-1]
+            present_key_value = present_key_value + cross_attn_present_key_value
+
+        layer_output = apply_chunking_to_forward(
+            self.feed_forward_chunk, self.chunk_size_feed_forward, self.seq_len_dim, attention_output
+        )
+        outputs = (layer_output,) + outputs
+
+        # if decoder, return the attn key/values as the last output
+        if self.is_decoder:
+            outputs = outputs + (present_key_value,)
+
+        return outputs
+
+    def feed_forward_chunk(self, attention_output):
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.output(intermediate_output, attention_output)
+        return layer_output
+
+
+class BertEncoder(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList([BertLayer(config) for _ in range(config.num_hidden_layers)])
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_values=None,
+        use_cache=None,
+        output_attentions=False,
+        output_hidden_states=False,
+        return_dict=True,
+    ):
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+        all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
+
+        next_decoder_cache = () if use_cache else None
+        for i, layer_module in enumerate(self.layer):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+            past_key_value = past_key_values[i] if past_key_values is not None else None
+
+            if getattr(self.config, "gradient_checkpointing", False) and self.training:
+
+                if use_cache:
+                    logger.warn(
+                        "`use_cache=True` is incompatible with `config.gradient_checkpointing=True`. Setting "
+                        "`use_cache=False`..."
+                    )
+                    use_cache = False
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, past_key_value, output_attentions)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module),
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                )
+            else:
+                layer_outputs = layer_module(
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    past_key_value,
+                    output_attentions,
+                )
+
+            hidden_states = layer_outputs[0]
+            if use_cache:
+                next_decoder_cache += (layer_outputs[-1],)
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (layer_outputs[1],)
+                if self.config.add_cross_attention:
+                    all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [
+                    hidden_states,
+                    next_decoder_cache,
+                    all_hidden_states,
+                    all_self_attentions,
+                    all_cross_attentions,
+                ]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_decoder_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+class BertPooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.activation = nn.Tanh()
+
+    def forward(self, hidden_states):
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        first_token_tensor = hidden_states[:, 0]
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+
+class BertPredictionHeadTransform(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        if isinstance(config.hidden_act, str):
+            self.transform_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.transform_act_fn = config.hidden_act
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+    def forward(self, hidden_states):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.transform_act_fn(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states)
+        return hidden_states
+
+
+class BertLMPredictionHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.transform = BertPredictionHeadTransform(config)
+
+        # The output weights are the same as the input embeddings, but there is
+        # an output-only bias for each token.
+        self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+
+        # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
+        self.decoder.bias = self.bias
+
+    def forward(self, hidden_states):
+        hidden_states = self.transform(hidden_states)
+        hidden_states = self.decoder(hidden_states)
+        return hidden_states
+
+
+class BertOnlyMLMHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.predictions = BertLMPredictionHead(config)
+
+    def forward(self, sequence_output):
+        prediction_scores = self.predictions(sequence_output)
+        return prediction_scores
+
+
+class BertOnlyNSPHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.seq_relationship = nn.Linear(config.hidden_size, 2)
+
+    def forward(self, pooled_output):
+        seq_relationship_score = self.seq_relationship(pooled_output)
+        return seq_relationship_score
+
+
+class BertPreTrainingHeads(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.predictions = BertLMPredictionHead(config)
+        self.seq_relationship = nn.Linear(config.hidden_size, 2)
+
+    def forward(self, sequence_output, pooled_output):
+        prediction_scores = self.predictions(sequence_output)
+        seq_relationship_score = self.seq_relationship(pooled_output)
+        return prediction_scores, seq_relationship_score
+
+
+class BertPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = BertConfig
+    load_tf_weights = load_tf_weights_in_bert
+    base_model_prefix = "bert"
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    def _init_weights(self, module):
+        """ Initialize the weights """
+        if isinstance(module, nn.Linear):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+
+@dataclass
+class BertForPreTrainingOutput(ModelOutput):
+    """
+    Output type of :class:`~transformers.BertForPreTraining`.
+
+    Args:
+        loss (`optional`, returned when ``labels`` is provided, ``torch.FloatTensor`` of shape :obj:`(1,)`):
+            Total loss as the sum of the masked language modeling loss and the next sequence prediction
+            (classification) loss.
+        prediction_logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        seq_relationship_logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, 2)`):
+            Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation
+            before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    prediction_logits: torch.FloatTensor = None
+    seq_relationship_logits: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+BERT_START_DOCSTRING = r"""
+
+    This model inherits from :class:`~transformers.PreTrainedModel`. Check the superclass documentation for the generic
+    methods the library implements for all its model (such as downloading or saving, resizing the input embeddings,
+    pruning heads etc.)
+
+    This model is also a PyTorch `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`__
+    subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to
+    general usage and behavior.
+
+    Parameters:
+        config (:class:`~transformers.BertConfig`): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model
+            weights.
+"""
+
+BERT_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using :class:`~transformers.BertTokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        attention_mask (:obj:`torch.FloatTensor` of shape :obj:`({0})`, `optional`):
+            Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            `What are attention masks? <../glossary.html#attention-mask>`__
+        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`({0})`, `optional`):
+            Segment token indices to indicate first and second portions of the inputs. Indices are selected in ``[0,
+            1]``:
+
+            - 0 corresponds to a `sentence A` token,
+            - 1 corresponds to a `sentence B` token.
+
+            `What are token type IDs? <../glossary.html#token-type-ids>`_
+        position_ids (:obj:`torch.LongTensor` of shape :obj:`({0})`, `optional`):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range ``[0,
+            config.max_position_embeddings - 1]``.
+
+            `What are position IDs? <../glossary.html#position-ids>`_
+        head_mask (:obj:`torch.FloatTensor` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in ``[0, 1]``:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`({0}, hidden_size)`, `optional`):
+            Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert :obj:`input_ids` indices into associated
+            vectors than the model's internal embedding lookup matrix.
+        output_attentions (:obj:`bool`, `optional`):
+            Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
+            tensors for more detail.
+        output_hidden_states (:obj:`bool`, `optional`):
+            Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
+            more detail.
+        return_dict (:obj:`bool`, `optional`):
+            Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare Bert Model transformer outputting raw hidden-states without any specific head on top.",
+    BERT_START_DOCSTRING,
+)
+class BertModel(BertPreTrainedModel):
+    """
+
+    The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of
+    cross-attention is added between the self-attention layers, following the architecture described in `Attention is
+    all you need <https://arxiv.org/abs/1706.03762>`__ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,
+    Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.
+
+    To behave as an decoder the model needs to be initialized with the :obj:`is_decoder` argument of the configuration
+    set to :obj:`True`. To be used in a Seq2Seq model, the model needs to initialized with both :obj:`is_decoder`
+    argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`encoder_hidden_states` is then expected as an
+    input to the forward pass.
+    """
+
+    def __init__(self, config, add_pooling_layer=True):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = BertEmbeddings(config)
+        self.encoder = BertEncoder(config)
+
+        self.pooler = BertPooler(config) if add_pooling_layer else None
+
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutputWithPoolingAndCrossAttentions,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_values=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        encoder_hidden_states  (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
+            the model is configured as a decoder.
+        encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+
+            If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if self.config.is_decoder:
+            use_cache = use_cache if use_cache is not None else self.config.use_cache
+        else:
+            use_cache = False
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            batch_size, seq_length = input_shape
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+            batch_size, seq_length = input_shape
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        # past_key_values_length
+        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
+
+        if attention_mask is None:
+            attention_mask = torch.ones(((batch_size, seq_length + past_key_values_length)), device=device)
+        if token_type_ids is None:
+            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape, device)
+
+        # If a 2D or 3D attention mask is provided for the cross-attention
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        if self.config.is_decoder and encoder_hidden_states is not None:
+            encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+            if encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
+            encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+        else:
+            encoder_extended_attention_mask = None
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            token_type_ids=token_type_ids,
+            inputs_embeds=inputs_embeds,
+            past_key_values_length=past_key_values_length,
+        )
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_extended_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = encoder_outputs[0]
+        pooled_output = self.pooler(sequence_output) if self.pooler is not None else None
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPoolingAndCrossAttentions(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            past_key_values=encoder_outputs.past_key_values,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+            cross_attentions=encoder_outputs.cross_attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Bert Model with two heads on top as done during the pretraining: a `masked language modeling` head and a `next
+    sentence prediction (classification)` head.
+    """,
+    BERT_START_DOCSTRING,
+)
+class BertForPreTraining(BertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.bert = BertModel(config)
+        self.cls = BertPreTrainingHeads(config)
+
+        self.init_weights()
+
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.cls.predictions.decoder = new_embeddings
+
+    @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @replace_return_docstrings(output_type=BertForPreTrainingOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        next_sentence_label=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape ``(batch_size, sequence_length)``, `optional`):
+            Labels for computing the masked language modeling loss. Indices should be in ``[-100, 0, ...,
+            config.vocab_size]`` (see ``input_ids`` docstring) Tokens with indices set to ``-100`` are ignored
+            (masked), the loss is only computed for the tokens with labels in ``[0, ..., config.vocab_size]``
+        next_sentence_label (``torch.LongTensor`` of shape ``(batch_size,)``, `optional`):
+            Labels for computing the next sequence prediction (classification) loss. Input should be a sequence pair
+            (see :obj:`input_ids` docstring) Indices should be in ``[0, 1]``:
+
+            - 0 indicates sequence B is a continuation of sequence A,
+            - 1 indicates sequence B is a random sequence.
+        kwargs (:obj:`Dict[str, any]`, optional, defaults to `{}`):
+            Used to hide legacy arguments that have been deprecated.
+
+        Returns:
+
+        Example::
+
+            >>> from transformers import BertTokenizer, BertForPreTraining
+            >>> import torch
+
+            >>> tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+            >>> model = BertForPreTraining.from_pretrained('bert-base-uncased')
+
+            >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
+            >>> outputs = model(**inputs)
+
+            >>> prediction_logits = outputs.prediction_logits
+            >>> seq_relationship_logits = outputs.seq_relationship_logits
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output, pooled_output = outputs[:2]
+        prediction_scores, seq_relationship_score = self.cls(sequence_output, pooled_output)
+
+        total_loss = None
+        if labels is not None and next_sentence_label is not None:
+            loss_fct = CrossEntropyLoss()
+            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
+            next_sentence_loss = loss_fct(seq_relationship_score.view(-1, 2), next_sentence_label.view(-1))
+            total_loss = masked_lm_loss + next_sentence_loss
+
+        if not return_dict:
+            output = (prediction_scores, seq_relationship_score) + outputs[2:]
+            return ((total_loss,) + output) if total_loss is not None else output
+
+        return BertForPreTrainingOutput(
+            loss=total_loss,
+            prediction_logits=prediction_scores,
+            seq_relationship_logits=seq_relationship_score,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """Bert Model with a `language modeling` head on top for CLM fine-tuning. """, BERT_START_DOCSTRING
+)
+class BertLMHeadModel(BertPreTrainedModel):
+
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+    _keys_to_ignore_on_load_missing = [r"position_ids", r"predictions.decoder.bias"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        if not config.is_decoder:
+            logger.warning("If you want to use `BertLMHeadModel` as a standalone, add `is_decoder=True.`")
+
+        self.bert = BertModel(config, add_pooling_layer=False)
+        self.cls = BertOnlyMLMHead(config)
+
+        self.init_weights()
+
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.cls.predictions.decoder = new_embeddings
+
+    @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @replace_return_docstrings(output_type=CausalLMOutputWithCrossAttentions, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        labels=None,
+        past_key_values=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        encoder_hidden_states  (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
+            the model is configured as a decoder.
+        encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Labels for computing the left-to-right language modeling loss (next word prediction). Indices should be in
+            ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring) Tokens with indices set to ``-100`` are
+            ignored (masked), the loss is only computed for the tokens with labels n ``[0, ..., config.vocab_size]``
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+
+            If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
+
+        Returns:
+
+        Example::
+
+            >>> from transformers import BertTokenizer, BertLMHeadModel, BertConfig
+            >>> import torch
+
+            >>> tokenizer = BertTokenizer.from_pretrained('bert-base-cased')
+            >>> config = BertConfig.from_pretrained("bert-base-cased")
+            >>> config.is_decoder = True
+            >>> model = BertLMHeadModel.from_pretrained('bert-base-cased', config=config)
+
+            >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
+            >>> outputs = model(**inputs)
+
+            >>> prediction_logits = outputs.logits
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        if labels is not None:
+            use_cache = False
+
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+        prediction_scores = self.cls(sequence_output)
+
+        lm_loss = None
+        if labels is not None:
+            # we are doing next-token prediction; shift prediction scores and input ids by one
+            shifted_prediction_scores = prediction_scores[:, :-1, :].contiguous()
+            labels = labels[:, 1:].contiguous()
+            loss_fct = CrossEntropyLoss()
+            lm_loss = loss_fct(shifted_prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
+
+        if not return_dict:
+            output = (prediction_scores,) + outputs[2:]
+            return ((lm_loss,) + output) if lm_loss is not None else output
+
+        return CausalLMOutputWithCrossAttentions(
+            loss=lm_loss,
+            logits=prediction_scores,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            cross_attentions=outputs.cross_attentions,
+        )
+
+    def prepare_inputs_for_generation(self, input_ids, past=None, attention_mask=None, **model_kwargs):
+        input_shape = input_ids.shape
+        # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly
+        if attention_mask is None:
+            attention_mask = input_ids.new_ones(input_shape)
+
+        # cut decoder_input_ids if past is used
+        if past is not None:
+            input_ids = input_ids[:, -1:]
+
+        return {"input_ids": input_ids, "attention_mask": attention_mask, "past_key_values": past}
+
+    def _reorder_cache(self, past, beam_idx):
+        reordered_past = ()
+        for layer_past in past:
+            reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
+        return reordered_past
+
+
+@add_start_docstrings("""Bert Model with a `language modeling` head on top. """, BERT_START_DOCSTRING)
+class BertForMaskedLM(BertPreTrainedModel):
+
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+    _keys_to_ignore_on_load_missing = [r"position_ids", r"predictions.decoder.bias"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        if config.is_decoder:
+            logger.warning(
+                "If you want to use `BertForMaskedLM` make sure `config.is_decoder=False` for "
+                "bi-directional self-attention."
+            )
+
+        self.bert = BertModel(config, add_pooling_layer=False)
+        self.cls = BertOnlyMLMHead(config)
+
+        self.init_weights()
+
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.cls.predictions.decoder = new_embeddings
+
+    @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=MaskedLMOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        labels=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Labels for computing the masked language modeling loss. Indices should be in ``[-100, 0, ...,
+            config.vocab_size]`` (see ``input_ids`` docstring) Tokens with indices set to ``-100`` are ignored
+            (masked), the loss is only computed for the tokens with labels in ``[0, ..., config.vocab_size]``
+        """
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+        prediction_scores = self.cls(sequence_output)
+
+        masked_lm_loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()  # -100 index = padding token
+            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
+
+        if not return_dict:
+            output = (prediction_scores,) + outputs[2:]
+            return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
+
+        return MaskedLMOutput(
+            loss=masked_lm_loss,
+            logits=prediction_scores,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def prepare_inputs_for_generation(self, input_ids, attention_mask=None, **model_kwargs):
+        input_shape = input_ids.shape
+        effective_batch_size = input_shape[0]
+
+        #  add a dummy token
+        assert self.config.pad_token_id is not None, "The PAD token should be defined for generation"
+        attention_mask = torch.cat([attention_mask, attention_mask.new_zeros((attention_mask.shape[0], 1))], dim=-1)
+        dummy_token = torch.full(
+            (effective_batch_size, 1), self.config.pad_token_id, dtype=torch.long, device=input_ids.device
+        )
+        input_ids = torch.cat([input_ids, dummy_token], dim=1)
+
+        return {"input_ids": input_ids, "attention_mask": attention_mask}
+
+
+@add_start_docstrings(
+    """Bert Model with a `next sentence prediction (classification)` head on top. """,
+    BERT_START_DOCSTRING,
+)
+class BertForNextSentencePrediction(BertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.bert = BertModel(config)
+        self.cls = BertOnlyNSPHead(config)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @replace_return_docstrings(output_type=NextSentencePredictorOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        **kwargs
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
+            Labels for computing the next sequence prediction (classification) loss. Input should be a sequence pair
+            (see ``input_ids`` docstring). Indices should be in ``[0, 1]``:
+
+            - 0 indicates sequence B is a continuation of sequence A,
+            - 1 indicates sequence B is a random sequence.
+
+        Returns:
+
+        Example::
+
+            >>> from transformers import BertTokenizer, BertForNextSentencePrediction
+            >>> import torch
+
+            >>> tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+            >>> model = BertForNextSentencePrediction.from_pretrained('bert-base-uncased')
+
+            >>> prompt = "In Italy, pizza served in formal settings, such as at a restaurant, is presented unsliced."
+            >>> next_sentence = "The sky is blue due to the shorter wavelength of blue light."
+            >>> encoding = tokenizer(prompt, next_sentence, return_tensors='pt')
+
+            >>> outputs = model(**encoding, labels=torch.LongTensor([1]))
+            >>> logits = outputs.logits
+            >>> assert logits[0, 0] < logits[0, 1] # next sentence was random
+        """
+
+        if "next_sentence_label" in kwargs:
+            warnings.warn(
+                "The `next_sentence_label` argument is deprecated and will be removed in a future version, use `labels` instead.",
+                FutureWarning,
+            )
+            labels = kwargs.pop("next_sentence_label")
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = outputs[1]
+
+        seq_relationship_scores = self.cls(pooled_output)
+
+        next_sentence_loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            next_sentence_loss = loss_fct(seq_relationship_scores.view(-1, 2), labels.view(-1))
+
+        if not return_dict:
+            output = (seq_relationship_scores,) + outputs[2:]
+            return ((next_sentence_loss,) + output) if next_sentence_loss is not None else output
+
+        return NextSentencePredictorOutput(
+            loss=next_sentence_loss,
+            logits=seq_relationship_scores,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Bert Model transformer with a sequence classification/regression head on top (a linear layer on top of the pooled
+    output) e.g. for GLUE tasks.
+    """,
+    BERT_START_DOCSTRING,
+)
+class BertForSequenceClassification(BertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.bert = BertModel(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=SequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
+            Labels for computing the sequence classification/regression loss. Indices should be in :obj:`[0, ...,
+            config.num_labels - 1]`. If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
+            If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = outputs[1]
+
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+
+        loss = None
+        if labels is not None:
+            if self.num_labels == 1:
+                #  We are doing regression
+                loss_fct = MSELoss()
+                loss = loss_fct(logits.view(-1), labels.view(-1))
+            else:
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Bert Model with a multiple choice classification head on top (a linear layer on top of the pooled output and a
+    softmax) e.g. for RocStories/SWAG tasks.
+    """,
+    BERT_START_DOCSTRING,
+)
+class BertForMultipleChoice(BertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.bert = BertModel(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, 1)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=MultipleChoiceModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
+            Labels for computing the multiple choice classification loss. Indices should be in ``[0, ...,
+            num_choices-1]`` where :obj:`num_choices` is the size of the second dimension of the input tensors. (See
+            :obj:`input_ids` above)
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        num_choices = input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]
+
+        input_ids = input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
+        attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
+        token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
+        position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
+        inputs_embeds = (
+            inputs_embeds.view(-1, inputs_embeds.size(-2), inputs_embeds.size(-1))
+            if inputs_embeds is not None
+            else None
+        )
+
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        pooled_output = outputs[1]
+
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+        reshaped_logits = logits.view(-1, num_choices)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(reshaped_logits, labels)
+
+        if not return_dict:
+            output = (reshaped_logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return MultipleChoiceModelOutput(
+            loss=loss,
+            logits=reshaped_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Bert Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
+    Named-Entity-Recognition (NER) tasks.
+    """,
+    BERT_START_DOCSTRING,
+)
+class BertForTokenClassification(BertPreTrainedModel):
+
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.bert = BertModel(config, add_pooling_layer=False)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Labels for computing the token classification loss. Indices should be in ``[0, ..., config.num_labels -
+            1]``.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+
+        sequence_output = self.dropout(sequence_output)
+        logits = self.classifier(sequence_output)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            # Only keep active parts of the loss
+            if attention_mask is not None:
+                active_loss = attention_mask.view(-1) == 1
+                active_logits = logits.view(-1, self.num_labels)
+                active_labels = torch.where(
+                    active_loss, labels.view(-1), torch.tensor(loss_fct.ignore_index).type_as(labels)
+                )
+                loss = loss_fct(active_logits, active_labels)
+            else:
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Bert Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear
+    layers on top of the hidden-states output to compute `span start logits` and `span end logits`).
+    """,
+    BERT_START_DOCSTRING,
+)
+class BertForQuestionAnswering(BertPreTrainedModel):
+
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.bert = BertModel(config, add_pooling_layer=False)
+        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=QuestionAnsweringModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        start_positions=None,
+        end_positions=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        start_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
+            Labels for position (index) of the start of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (:obj:`sequence_length`). Position outside of the
+            sequence are not taken into account for computing the loss.
+        end_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
+            Labels for position (index) of the end of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (:obj:`sequence_length`). Position outside of the
+            sequence are not taken into account for computing the loss.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+
+        logits = self.qa_outputs(sequence_output)
+        start_logits, end_logits = logits.split(1, dim=-1)
+        start_logits = start_logits.squeeze(-1)
+        end_logits = end_logits.squeeze(-1)
+
+        total_loss = None
+        if start_positions is not None and end_positions is not None:
+            # If we are on multi-GPU, split add a dimension
+            if len(start_positions.size()) > 1:
+                start_positions = start_positions.squeeze(-1)
+            if len(end_positions.size()) > 1:
+                end_positions = end_positions.squeeze(-1)
+            # sometimes the start/end positions are outside our model inputs, we ignore these terms
+            ignored_index = start_logits.size(1)
+            start_positions.clamp_(0, ignored_index)
+            end_positions.clamp_(0, ignored_index)
+
+            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
+            start_loss = loss_fct(start_logits, start_positions)
+            end_loss = loss_fct(end_logits, end_positions)
+            total_loss = (start_loss + end_loss) / 2
+
+        if not return_dict:
+            output = (start_logits, end_logits) + outputs[2:]
+            return ((total_loss,) + output) if total_loss is not None else output
+
+        return QuestionAnsweringModelOutput(
+            loss=total_loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

modeling_outputs.py

+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import Optional, Tuple
+
+import torch
+
+from file_utils import ModelOutput
+
+
+@dataclass
+class BaseModelOutput(ModelOutput):
+    """
+    Base class for model's outputs, with potential hidden states and attentions.
+
+    Args:
+        last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    last_hidden_state: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class BaseModelOutputWithPooling(ModelOutput):
+    """
+    Base class for model's outputs that also contains a pooling of the last hidden states.
+
+    Args:
+        last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        pooler_output (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, hidden_size)`):
+            Last layer hidden-state of the first token of the sequence (classification token) further processed by a
+            Linear layer and a Tanh activation function. The Linear layer weights are trained from the next sentence
+            prediction (classification) objective during pretraining.
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    last_hidden_state: torch.FloatTensor = None
+    pooler_output: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class BaseModelOutputWithPast(ModelOutput):
+    """
+    Base class for model's outputs that may also contain a past key/values (to speed up sequential decoding).
+
+    Args:
+        last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+
+            If :obj:`past_key_values` is used only the last hidden-state of the sequences of shape :obj:`(batch_size,
+            1, hidden_size)` is output.
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            Tuple of :obj:`tuple(torch.FloatTensor)` of length :obj:`config.n_layers`, with each tuple having 2 tensors
+            of shape :obj:`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and optionally if
+            ``config.is_encoder_decoder=True`` 2 additional tensors of shape :obj:`(batch_size, num_heads,
+            encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and optionally if
+            ``config.is_encoder_decoder=True`` in the cross-attention blocks) that can be used (see
+            :obj:`past_key_values` input) to speed up sequential decoding.
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    last_hidden_state: torch.FloatTensor = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class BaseModelOutputWithCrossAttentions(ModelOutput):
+    """
+    Base class for model's outputs, with potential hidden states and attentions.
+
+    Args:
+        last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+        cross_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` and ``config.add_cross_attention=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
+            weighted average in the cross-attention heads.
+    """
+
+    last_hidden_state: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+    cross_attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class BaseModelOutputWithPoolingAndCrossAttentions(ModelOutput):
+    """
+    Base class for model's outputs that also contains a pooling of the last hidden states.
+
+    Args:
+        last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        pooler_output (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, hidden_size)`):
+            Last layer hidden-state of the first token of the sequence (classification token) further processed by a
+            Linear layer and a Tanh activation function. The Linear layer weights are trained from the next sentence
+            prediction (classification) objective during pretraining.
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+        cross_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` and ``config.add_cross_attention=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
+            weighted average in the cross-attention heads.
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            Tuple of :obj:`tuple(torch.FloatTensor)` of length :obj:`config.n_layers`, with each tuple having 2 tensors
+            of shape :obj:`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and optionally if
+            ``config.is_encoder_decoder=True`` 2 additional tensors of shape :obj:`(batch_size, num_heads,
+            encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and optionally if
+            ``config.is_encoder_decoder=True`` in the cross-attention blocks) that can be used (see
+            :obj:`past_key_values` input) to speed up sequential decoding.
+    """
+
+    last_hidden_state: torch.FloatTensor = None
+    pooler_output: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+    cross_attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class BaseModelOutputWithPastAndCrossAttentions(ModelOutput):
+    """
+    Base class for model's outputs that may also contain a past key/values (to speed up sequential decoding).
+
+    Args:
+        last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+
+            If :obj:`past_key_values` is used only the last hidden-state of the sequences of shape :obj:`(batch_size,
+            1, hidden_size)` is output.
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            Tuple of :obj:`tuple(torch.FloatTensor)` of length :obj:`config.n_layers`, with each tuple having 2 tensors
+            of shape :obj:`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and optionally if
+            ``config.is_encoder_decoder=True`` 2 additional tensors of shape :obj:`(batch_size, num_heads,
+            encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and optionally if
+            ``config.is_encoder_decoder=True`` in the cross-attention blocks) that can be used (see
+            :obj:`past_key_values` input) to speed up sequential decoding.
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+        cross_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` and ``config.add_cross_attention=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
+            weighted average in the cross-attention heads.
+    """
+
+    last_hidden_state: torch.FloatTensor = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+    cross_attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class Seq2SeqModelOutput(ModelOutput):
+    """
+    Base class for model encoder's outputs that also contains : pre-computed hidden states that can speed up sequential
+    decoding.
+
+    Args:
+        last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the decoder of the model.
+
+            If :obj:`past_key_values` is used only the last hidden-state of the sequences of shape :obj:`(batch_size,
+            1, hidden_size)` is output.
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            Tuple of :obj:`tuple(torch.FloatTensor)` of length :obj:`config.n_layers`, with each tuple having 2 tensors
+            of shape :obj:`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
+            shape :obj:`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used (see :obj:`past_key_values` input) to speed up sequential decoding.
+        decoder_hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
+        decoder_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+        cross_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
+            weighted average in the cross-attention heads.
+        encoder_last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder of the model.
+        encoder_hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
+        encoder_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+    """
+
+    last_hidden_state: torch.FloatTensor = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
+    decoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    decoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
+    cross_attentions: Optional[Tuple[torch.FloatTensor]] = None
+    encoder_last_hidden_state: Optional[torch.FloatTensor] = None
+    encoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    encoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class CausalLMOutput(ModelOutput):
+    """
+    Base class for causal language model (or autoregressive) outputs.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Language modeling loss (for next-token prediction).
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class CausalLMOutputWithPast(ModelOutput):
+    """
+    Base class for causal language model (or autoregressive) outputs.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Language modeling loss (for next-token prediction).
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        past_key_values (:obj:`tuple(tupel(torch.FloatTensor))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            Tuple of :obj:`tuple(torch.FloatTensor)` of length :obj:`config.n_layers`, with each tuple having 2 tensors
+            of shape :obj:`(batch_size, num_heads, sequence_length, embed_size_per_head)`)
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
+            :obj:`past_key_values` input) to speed up sequential decoding.
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: torch.FloatTensor = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class CausalLMOutputWithCrossAttentions(ModelOutput):
+    """
+    Base class for causal language model (or autoregressive) outputs.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Language modeling loss (for next-token prediction).
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+        cross_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Cross attentions weights after the attention softmax, used to compute the weighted average in the
+            cross-attention heads.
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            Tuple of :obj:`torch.FloatTensor` tuples of length :obj:`config.n_layers`, with each tuple containing the
+            cached key, value states of the self-attention and the cross-attention layers if model is used in
+            encoder-decoder setting. Only relevant if ``config.is_decoder = True``.
+
+            Contains pre-computed hidden-states (key and values in the attention blocks) that can be used (see
+            :obj:`past_key_values` input) to speed up sequential decoding.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: torch.FloatTensor = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+    cross_attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class SequenceClassifierOutputWithPast(ModelOutput):
+    """
+    Base class for outputs of sentence classification models.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Classification (or regression if config.num_labels==1) loss.
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.num_labels)`):
+            Classification (or regression if config.num_labels==1) scores (before SoftMax).
+        past_key_values (:obj:`tuple(tupel(torch.FloatTensor))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            Tuple of :obj:`tuple(torch.FloatTensor)` of length :obj:`config.n_layers`, with each tuple having 2 tensors
+            of shape :obj:`(batch_size, num_heads, sequence_length, embed_size_per_head)`)
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
+            :obj:`past_key_values` input) to speed up sequential decoding.
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: torch.FloatTensor = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class MaskedLMOutput(ModelOutput):
+    """
+    Base class for masked language models outputs.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Masked language modeling (MLM) loss.
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class Seq2SeqLMOutput(ModelOutput):
+    """
+    Base class for sequence-to-sequence language models outputs.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Language modeling loss.
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            Tuple of :obj:`tuple(torch.FloatTensor)` of length :obj:`config.n_layers`, with each tuple having 2 tensors
+            of shape :obj:`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
+            shape :obj:`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used (see :obj:`past_key_values` input) to speed up sequential decoding.
+        decoder_hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
+        decoder_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+        cross_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
+            weighted average in the cross-attention heads.
+        encoder_last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder of the model.
+        encoder_hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
+        encoder_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: torch.FloatTensor = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
+    decoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    decoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
+    cross_attentions: Optional[Tuple[torch.FloatTensor]] = None
+    encoder_last_hidden_state: Optional[torch.FloatTensor] = None
+    encoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    encoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class NextSentencePredictorOutput(ModelOutput):
+    """
+    Base class for outputs of models predicting if two sentences are consecutive or not.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`next_sentence_label` is provided):
+            Next sequence prediction (classification) loss.
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, 2)`):
+            Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation
+            before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class SequenceClassifierOutput(ModelOutput):
+    """
+    Base class for outputs of sentence classification models.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Classification (or regression if config.num_labels==1) loss.
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.num_labels)`):
+            Classification (or regression if config.num_labels==1) scores (before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class Seq2SeqSequenceClassifierOutput(ModelOutput):
+    """
+    Base class for outputs of sequence-to-sequence sentence classification models.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`label` is provided):
+            Classification (or regression if config.num_labels==1) loss.
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.num_labels)`):
+            Classification (or regression if config.num_labels==1) scores (before SoftMax).
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            Tuple of :obj:`tuple(torch.FloatTensor)` of length :obj:`config.n_layers`, with each tuple having 2 tensors
+            of shape :obj:`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
+            shape :obj:`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used (see :obj:`past_key_values` input) to speed up sequential decoding.
+        decoder_hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
+        decoder_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+        cross_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
+            weighted average in the cross-attention heads.
+        encoder_last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder of the model.
+        encoder_hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
+        encoder_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: torch.FloatTensor = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
+    decoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    decoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
+    cross_attentions: Optional[Tuple[torch.FloatTensor]] = None
+    encoder_last_hidden_state: Optional[torch.FloatTensor] = None
+    encoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    encoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class MultipleChoiceModelOutput(ModelOutput):
+    """
+    Base class for outputs of multiple choice models.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape `(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Classification loss.
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, num_choices)`):
+            `num_choices` is the second dimension of the input tensors. (see `input_ids` above).
+
+            Classification scores (before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class TokenClassifierOutput(ModelOutput):
+    """
+    Base class for outputs of token classification models.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when ``labels`` is provided) :
+            Classification loss.
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.num_labels)`):
+            Classification scores (before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class QuestionAnsweringModelOutput(ModelOutput):
+    """
+    Base class for outputs of question answering models.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.
+        start_logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`):
+            Span-start scores (before SoftMax).
+        end_logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`):
+            Span-end scores (before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    start_logits: torch.FloatTensor = None
+    end_logits: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class Seq2SeqQuestionAnsweringModelOutput(ModelOutput):
+    """
+    Base class for outputs of sequence-to-sequence question answering models.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.
+        start_logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`):
+            Span-start scores (before SoftMax).
+        end_logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`):
+            Span-end scores (before SoftMax).
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            Tuple of :obj:`tuple(torch.FloatTensor)` of length :obj:`config.n_layers`, with each tuple having 2 tensors
+            of shape :obj:`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
+            shape :obj:`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used (see :obj:`past_key_values` input) to speed up sequential decoding.
+        decoder_hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.
+        decoder_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+        cross_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
+            weighted average in the cross-attention heads.
+        encoder_last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder of the model.
+        encoder_hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.
+        encoder_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
+            sequence_length, sequence_length)`.
+
+            Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the
+            self-attention heads.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    start_logits: torch.FloatTensor = None
+    end_logits: torch.FloatTensor = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
+    decoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    decoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
+    cross_attentions: Optional[Tuple[torch.FloatTensor]] = None
+    encoder_last_hidden_state: Optional[torch.FloatTensor] = None
+    encoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    encoder_attentions: Optional[Tuple[torch.FloatTensor]] = None

modeling_utils.py

+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors, Facebook AI Research authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import os
+import re
+import warnings
+from dataclasses import dataclass
+from typing import Any, Callable, Dict, List, Optional, Set, Tuple, Union
+
+import torch
+from torch import Tensor, device, dtype, nn
+from torch.nn import CrossEntropyLoss
+from torch.nn import functional as F
+
+from activations import get_activation
+from configuration_utils import PretrainedConfig
+from file_utils import (
+    DUMMY_INPUTS,
+    TF2_WEIGHTS_NAME,
+    TF_WEIGHTS_NAME,
+    WEIGHTS_NAME,
+    ModelOutput,
+    cached_path,
+    hf_bucket_url,
+    is_offline_mode,
+    is_remote_url,
+    replace_return_docstrings,
+)
+from generation_utils import GenerationMixin
+from utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+try:
+    from torch.nn import Identity
+except ImportError:
+    # Older PyTorch compatibility
+    class Identity(nn.Module):
+        r"""A placeholder identity operator that is argument-insensitive."""
+
+        def __init__(self, *args, **kwargs):
+            super().__init__()
+
+        def forward(self, input):
+            return input
+
+
+def find_pruneable_heads_and_indices(
+    heads: List[int], n_heads: int, head_size: int, already_pruned_heads: Set[int]
+) -> Tuple[Set[int], torch.LongTensor]:
+    """
+    Finds the heads and their indices taking :obj:`already_pruned_heads` into account.
+
+    Args:
+        heads (:obj:`List[int]`): List of the indices of heads to prune.
+        n_heads (:obj:`int`): The number of heads in the model.
+        head_size (:obj:`int`): The size of each head.
+        already_pruned_heads (:obj:`Set[int]`): A set of already pruned heads.
+
+    Returns:
+        :obj:`Tuple[Set[int], torch.LongTensor]`: A tuple with the remaining heads and their corresponding indices.
+    """
+    mask = torch.ones(n_heads, head_size)
+    heads = set(heads) - already_pruned_heads  # Convert to set and remove already pruned heads
+    for head in heads:
+        # Compute how many pruned heads are before the head and move the index accordingly
+        head = head - sum(1 if h < head else 0 for h in already_pruned_heads)
+        mask[head] = 0
+    mask = mask.view(-1).contiguous().eq(1)
+    index: torch.LongTensor = torch.arange(len(mask))[mask].long()
+    return heads, index
+
+
+def get_parameter_device(parameter: Union[nn.Module, GenerationMixin, "ModuleUtilsMixin"]):
+    try:
+        return next(parameter.parameters()).device
+    except StopIteration:
+        # For nn.DataParallel compatibility in PyTorch 1.5
+
+        def find_tensor_attributes(module: nn.Module) -> List[Tuple[str, Tensor]]:
+            tuples = [(k, v) for k, v in module.__dict__.items() if torch.is_tensor(v)]
+            return tuples
+
+        gen = parameter._named_members(get_members_fn=find_tensor_attributes)
+        first_tuple = next(gen)
+        return first_tuple[1].device
+
+
+def get_parameter_dtype(parameter: Union[nn.Module, GenerationMixin, "ModuleUtilsMixin"]):
+    try:
+        return next(parameter.parameters()).dtype
+    except StopIteration:
+        # For nn.DataParallel compatibility in PyTorch 1.5
+
+        def find_tensor_attributes(module: nn.Module) -> List[Tuple[str, Tensor]]:
+            tuples = [(k, v) for k, v in module.__dict__.items() if torch.is_tensor(v)]
+            return tuples
+
+        gen = parameter._named_members(get_members_fn=find_tensor_attributes)
+        first_tuple = next(gen)
+        return first_tuple[1].dtype
+
+
+class ModuleUtilsMixin:
+    """
+    A few utilities for :obj:`torch.nn.Modules`, to be used as a mixin.
+    """
+
+    @staticmethod
+    def _hook_rss_memory_pre_forward(module, *args, **kwargs):
+        try:
+            import psutil
+        except (ImportError):
+            raise ImportError("You need to install psutil (pip install psutil) to use memory tracing.")
+
+        process = psutil.Process(os.getpid())
+        mem = process.memory_info()
+        module.mem_rss_pre_forward = mem.rss
+        return None
+
+    @staticmethod
+    def _hook_rss_memory_post_forward(module, *args, **kwargs):
+        try:
+            import psutil
+        except (ImportError):
+            raise ImportError("You need to install psutil (pip install psutil) to use memory tracing.")
+
+        process = psutil.Process(os.getpid())
+        mem = process.memory_info()
+        module.mem_rss_post_forward = mem.rss
+        mem_rss_diff = module.mem_rss_post_forward - module.mem_rss_pre_forward
+        module.mem_rss_diff = mem_rss_diff + (module.mem_rss_diff if hasattr(module, "mem_rss_diff") else 0)
+        return None
+
+    def add_memory_hooks(self):
+        """
+        Add a memory hook before and after each sub-module forward pass to record increase in memory consumption.
+
+        Increase in memory consumption is stored in a :obj:`mem_rss_diff` attribute for each module and can be reset to
+        zero with :obj:`model.reset_memory_hooks_state()`.
+        """
+        for module in self.modules():
+            module.register_forward_pre_hook(self._hook_rss_memory_pre_forward)
+            module.register_forward_hook(self._hook_rss_memory_post_forward)
+        self.reset_memory_hooks_state()
+
+    def reset_memory_hooks_state(self):
+        """
+        Reset the :obj:`mem_rss_diff` attribute of each module (see
+        :func:`~transformers.modeling_utils.ModuleUtilsMixin.add_memory_hooks`).
+        """
+        for module in self.modules():
+            module.mem_rss_diff = 0
+            module.mem_rss_post_forward = 0
+            module.mem_rss_pre_forward = 0
+
+    @property
+    def device(self) -> device:
+        """
+        :obj:`torch.device`: The device on which the module is (assuming that all the module parameters are on the same
+        device).
+        """
+        return get_parameter_device(self)
+
+    @property
+    def dtype(self) -> dtype:
+        """
+        :obj:`torch.dtype`: The dtype of the module (assuming that all the module parameters have the same dtype).
+        """
+        return get_parameter_dtype(self)
+
+    def invert_attention_mask(self, encoder_attention_mask: Tensor) -> Tensor:
+        """
+        Invert an attention mask (e.g., switches 0. and 1.).
+
+        Args:
+            encoder_attention_mask (:obj:`torch.Tensor`): An attention mask.
+
+        Returns:
+            :obj:`torch.Tensor`: The inverted attention mask.
+        """
+        if encoder_attention_mask.dim() == 3:
+            encoder_extended_attention_mask = encoder_attention_mask[:, None, :, :]
+        if encoder_attention_mask.dim() == 2:
+            encoder_extended_attention_mask = encoder_attention_mask[:, None, None, :]
+        # T5 has a mask that can compare sequence ids, we can simulate this here with this transposition
+        # Cf. https://github.com/tensorflow/mesh/blob/8d2465e9bc93129b913b5ccc6a59aa97abd96ec6/mesh_tensorflow
+        # /transformer/transformer_layers.py#L270
+        # encoder_extended_attention_mask = (encoder_extended_attention_mask ==
+        # encoder_extended_attention_mask.transpose(-1, -2))
+        encoder_extended_attention_mask = encoder_extended_attention_mask.to(dtype=self.dtype)  # fp16 compatibility
+
+        if self.dtype == torch.float16:
+            encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -1e4
+        elif self.dtype == torch.float32:
+            encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -1e9
+        else:
+            raise ValueError(
+                "{} not recognized. `dtype` should be set to either `torch.float32` or `torch.float16`".format(
+                    self.dtype
+                )
+            )
+
+        return encoder_extended_attention_mask
+
+    def get_extended_attention_mask(self, attention_mask: Tensor, input_shape: Tuple[int], device: device) -> Tensor:
+        """
+        Makes broadcastable attention and causal masks so that future and masked tokens are ignored.
+
+        Arguments:
+            attention_mask (:obj:`torch.Tensor`):
+                Mask with ones indicating tokens to attend to, zeros for tokens to ignore.
+            input_shape (:obj:`Tuple[int]`):
+                The shape of the input to the model.
+            device: (:obj:`torch.device`):
+                The device of the input to the model.
+
+        Returns:
+            :obj:`torch.Tensor` The extended attention mask, with a the same dtype as :obj:`attention_mask.dtype`.
+        """
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        if attention_mask.dim() == 3:
+            extended_attention_mask = attention_mask[:, None, :, :]
+        elif attention_mask.dim() == 2:
+            # Provided a padding mask of dimensions [batch_size, seq_length]
+            # - if the model is a decoder, apply a causal mask in addition to the padding mask
+            # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
+            if self.config.is_decoder:
+                batch_size, seq_length = input_shape
+                seq_ids = torch.arange(seq_length, device=device)
+                causal_mask = seq_ids[None, None, :].repeat(batch_size, seq_length, 1) <= seq_ids[None, :, None]
+                # in case past_key_values are used we need to add a prefix ones mask to the causal mask
+                # causal and attention masks must have same type with pytorch version < 1.3
+                causal_mask = causal_mask.to(attention_mask.dtype)
+
+                if causal_mask.shape[1] < attention_mask.shape[1]:
+                    prefix_seq_len = attention_mask.shape[1] - causal_mask.shape[1]
+                    causal_mask = torch.cat(
+                        [
+                            torch.ones(
+                                (batch_size, seq_length, prefix_seq_len), device=device, dtype=causal_mask.dtype
+                            ),
+                            causal_mask,
+                        ],
+                        axis=-1,
+                    )
+
+                extended_attention_mask = causal_mask[:, None, :, :] * attention_mask[:, None, None, :]
+            else:
+                extended_attention_mask = attention_mask[:, None, None, :]
+        else:
+            raise ValueError(
+                "Wrong shape for input_ids (shape {}) or attention_mask (shape {})".format(
+                    input_shape, attention_mask.shape
+                )
+            )
+
+        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+        # masked positions, this operation will create a tensor which is 0.0 for
+        # positions we want to attend and -10000.0 for masked positions.
+        # Since we are adding it to the raw scores before the softmax, this is
+        # effectively the same as removing these entirely.
+        extended_attention_mask = extended_attention_mask.to(dtype=self.dtype)  # fp16 compatibility
+        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
+        return extended_attention_mask
+
+    def get_head_mask(
+        self, head_mask: Optional[Tensor], num_hidden_layers: int, is_attention_chunked: bool = False
+    ) -> Tensor:
+        """
+        Prepare the head mask if needed.
+
+        Args:
+            head_mask (:obj:`torch.Tensor` with shape :obj:`[num_heads]` or :obj:`[num_hidden_layers x num_heads]`, `optional`):
+                The mask indicating if we should keep the heads or not (1.0 for keep, 0.0 for discard).
+            num_hidden_layers (:obj:`int`):
+                The number of hidden layers in the model.
+            is_attention_chunked: (:obj:`bool`, `optional, defaults to :obj:`False`):
+                Whether or not the attentions scores are computed by chunks or not.
+
+        Returns:
+            :obj:`torch.Tensor` with shape :obj:`[num_hidden_layers x batch x num_heads x seq_length x seq_length]` or
+            list with :obj:`[None]` for each layer.
+        """
+        if head_mask is not None:
+            head_mask = self._convert_head_mask_to_5d(head_mask, num_hidden_layers)
+            if is_attention_chunked is True:
+                head_mask = head_mask.unsqueeze(-1)
+        else:
+            head_mask = [None] * num_hidden_layers
+
+        return head_mask
+
+    def _convert_head_mask_to_5d(self, head_mask, num_hidden_layers):
+        """-> [num_hidden_layers x batch x num_heads x seq_length x seq_length]"""
+        if head_mask.dim() == 1:
+            head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
+            head_mask = head_mask.expand(num_hidden_layers, -1, -1, -1, -1)
+        elif head_mask.dim() == 2:
+            head_mask = head_mask.unsqueeze(1).unsqueeze(-1).unsqueeze(-1)  # We can specify head_mask for each layer
+        assert head_mask.dim() == 5, f"head_mask.dim != 5, instead {head_mask.dim()}"
+        head_mask = head_mask.to(dtype=self.dtype)  # switch to float if need + fp16 compatibility
+        return head_mask
+
+    def num_parameters(self, only_trainable: bool = False, exclude_embeddings: bool = False) -> int:
+        """
+        Get number of (optionally, trainable or non-embeddings) parameters in the module.
+
+        Args:
+            only_trainable (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Whether or not to return only the number of trainable parameters
+
+            exclude_embeddings (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Whether or not to return only the number of non-embeddings parameters
+
+        Returns:
+            :obj:`int`: The number of parameters.
+        """
+
+        def parameter_filter(x):
+            return (x.requires_grad or not only_trainable) and not (
+                isinstance(x, torch.nn.Embedding) and exclude_embeddings
+            )
+
+        params = filter(parameter_filter, self.parameters()) if only_trainable else self.parameters()
+        return sum(p.numel() for p in params)
+
+    def estimate_tokens(self, input_dict: Dict[str, Union[torch.Tensor, Any]]) -> int:
+        """
+        Helper function to estimate the total number of tokens from the model inputs.
+
+        Args:
+            inputs (:obj:`dict`): The model inputs.
+
+        Returns:
+            :obj:`int`: The total number of tokens.
+        """
+        token_inputs = [tensor for key, tensor in input_dict.items() if "input" in key]
+        if token_inputs:
+            return sum([token_input.numel() for token_input in token_inputs])
+        else:
+            warnings.warn(
+                "Could not estimate the number of tokens of the input, floating-point operations will not be computed"
+            )
+            return 0
+
+    def floating_point_ops(
+        self, input_dict: Dict[str, Union[torch.Tensor, Any]], exclude_embeddings: bool = True
+    ) -> int:
+        """
+        Get number of (optionally, non-embeddings) floating-point operations for the forward and backward passes of a
+        batch with this transformer model. Default approximation neglects the quadratic dependency on the number of
+        tokens (valid if :obj:`12 * d_model << sequence_length`) as laid out in `this paper
+        <https://arxiv.org/pdf/2001.08361.pdf>`__ section 2.1. Should be overridden for transformers with parameter
+        re-use e.g. Albert or Universal Transformers, or if doing long-range modeling with very high sequence lengths.
+
+        Args:
+            batch_size (:obj:`int`):
+                The batch size for the forward pass.
+
+            sequence_length (:obj:`int`):
+                The number of tokens in each line of the batch.
+
+            exclude_embeddings (:obj:`bool`, `optional`, defaults to :obj:`True`):
+                Whether or not to count embedding and softmax operations.
+
+        Returns:
+            :obj:`int`: The number of floating-point operations.
+        """
+
+        return 6 * self.estimate_tokens(input_dict) * self.num_parameters(exclude_embeddings=exclude_embeddings)
+
+
+class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin):
+    r"""
+    Base class for all models.
+
+    :class:`~transformers.PreTrainedModel` takes care of storing the configuration of the models and handles methods
+    for loading, downloading and saving models as well as a few methods common to all models to:
+
+        * resize the input embeddings,
+        * prune heads in the self-attention heads.
+
+    Class attributes (overridden by derived classes):
+
+        - **config_class** (:class:`~transformers.PretrainedConfig`) -- A subclass of
+          :class:`~transformers.PretrainedConfig` to use as configuration class for this model architecture.
+        - **load_tf_weights** (:obj:`Callable`) -- A python `method` for loading a TensorFlow checkpoint in a PyTorch
+          model, taking as arguments:
+
+            - **model** (:class:`~transformers.PreTrainedModel`) -- An instance of the model on which to load the
+              TensorFlow checkpoint.
+            - **config** (:class:`~transformers.PreTrainedConfig`) -- An instance of the configuration associated to
+              the model.
+            - **path** (:obj:`str`) -- A path to the TensorFlow checkpoint.
+
+        - **base_model_prefix** (:obj:`str`) -- A string indicating the attribute associated to the base model in
+          derived classes of the same architecture adding modules on top of the base model.
+        - **is_parallelizable** (:obj:`bool`) -- A flag indicating whether this model supports model parallelization.
+    """
+    config_class = None
+    base_model_prefix = ""
+    # a list of re pattern of tensor names to ignore from the model when loading the model weights
+    # (and avoid unnecessary warnings).
+    _keys_to_ignore_on_load_missing = None
+    # a list of re pattern of tensor names to ignore from the weights when loading the model weights
+    # (and avoid unnecessary warnings).
+    _keys_to_ignore_on_load_unexpected = None
+    # a list of of tensor names to ignore when saving the model (useful for keys that aren't
+    # trained, but which are deterministic)
+    _keys_to_ignore_on_save = None
+
+    is_parallelizable = False
+
+    @property
+    def dummy_inputs(self) -> Dict[str, torch.Tensor]:
+        """
+        :obj:`Dict[str, torch.Tensor]`: Dummy inputs to do a forward pass in the network.
+        """
+        return {"input_ids": torch.tensor(DUMMY_INPUTS)}
+
+    def __init__(self, config: PretrainedConfig, *inputs, **kwargs):
+        super().__init__()
+        if not isinstance(config, PretrainedConfig):
+            raise ValueError(
+                "Parameter config in `{}(config)` should be an instance of class `PretrainedConfig`. "
+                "To create a model from a pretrained model use "
+                "`model = {}.from_pretrained(PRETRAINED_MODEL_NAME)`".format(
+                    self.__class__.__name__, self.__class__.__name__
+                )
+            )
+        # Save config and origin of the pretrained weights if given in model
+        self.config = config
+        self.name_or_path = config.name_or_path
+
+    @property
+    def base_model(self) -> nn.Module:
+        """
+        :obj:`torch.nn.Module`: The main body of the model.
+        """
+        return getattr(self, self.base_model_prefix, self)
+
+    def get_input_embeddings(self) -> nn.Module:
+        """
+        Returns the model's input embeddings.
+
+        Returns:
+            :obj:`nn.Module`: A torch module mapping vocabulary to hidden states.
+        """
+        base_model = getattr(self, self.base_model_prefix, self)
+        if base_model is not self:
+            return base_model.get_input_embeddings()
+        else:
+            raise NotImplementedError
+
+    def set_input_embeddings(self, value: nn.Module):
+        """
+        Set model's input embeddings.
+
+        Args:
+            value (:obj:`nn.Module`): A module mapping vocabulary to hidden states.
+        """
+        base_model = getattr(self, self.base_model_prefix, self)
+        if base_model is not self:
+            base_model.set_input_embeddings(value)
+        else:
+            raise NotImplementedError
+
+    def get_output_embeddings(self) -> nn.Module:
+        """
+        Returns the model's output embeddings.
+
+        Returns:
+            :obj:`nn.Module`: A torch module mapping hidden states to vocabulary.
+        """
+        return None  # Overwrite for models with output embeddings
+
+    def tie_weights(self):
+        """
+        Tie the weights between the input embeddings and the output embeddings.
+
+        If the :obj:`torchscript` flag is set in the configuration, can't handle parameter sharing so we are cloning
+        the weights instead.
+        """
+        output_embeddings = self.get_output_embeddings()
+        if output_embeddings is not None and self.config.tie_word_embeddings:
+            self._tie_or_clone_weights(output_embeddings, self.get_input_embeddings())
+
+        if self.config.is_encoder_decoder and self.config.tie_encoder_decoder:
+            if hasattr(self, self.base_model_prefix):
+                self = getattr(self, self.base_model_prefix)
+            self._tie_encoder_decoder_weights(self.encoder, self.decoder, self.base_model_prefix)
+
+    @staticmethod
+    def _tie_encoder_decoder_weights(encoder: nn.Module, decoder: nn.Module, base_model_prefix: str):
+        uninitialized_encoder_weights: List[str] = []
+        if decoder.__class__ != encoder.__class__:
+            logger.info(
+                f"{decoder.__class__} and {encoder.__class__} are not equal. In this case make sure that all encoder weights are correctly initialized."
+            )
+
+        def tie_encoder_to_decoder_recursively(
+            decoder_pointer: nn.Module,
+            encoder_pointer: nn.Module,
+            module_name: str,
+            uninitialized_encoder_weights: List[str],
+            depth=0,
+        ):
+            assert isinstance(decoder_pointer, nn.Module) and isinstance(
+                encoder_pointer, nn.Module
+            ), f"{decoder_pointer} and {encoder_pointer} have to be of type torch.nn.Module"
+            if hasattr(decoder_pointer, "weight"):
+                assert hasattr(encoder_pointer, "weight")
+                encoder_pointer.weight = decoder_pointer.weight
+                if hasattr(decoder_pointer, "bias"):
+                    assert hasattr(encoder_pointer, "bias")
+                    encoder_pointer.bias = decoder_pointer.bias
+                return
+
+            encoder_modules = encoder_pointer._modules
+            decoder_modules = decoder_pointer._modules
+            if len(decoder_modules) > 0:
+                assert (
+                    len(encoder_modules) > 0
+                ), f"Encoder module {encoder_pointer} does not match decoder module {decoder_pointer}"
+
+                all_encoder_weights = set([module_name + "/" + sub_name for sub_name in encoder_modules.keys()])
+                encoder_layer_pos = 0
+                for name, module in decoder_modules.items():
+                    if name.isdigit():
+                        encoder_name = str(int(name) + encoder_layer_pos)
+                        decoder_name = name
+                        if not isinstance(decoder_modules[decoder_name], type(encoder_modules[encoder_name])) and len(
+                            encoder_modules
+                        ) != len(decoder_modules):
+                            # this can happen if the name corresponds to the position in a list module list of layers
+                            # in this case the decoder has added a cross-attention that the encoder does not have
+                            # thus skip this step and subtract one layer pos from encoder
+                            encoder_layer_pos -= 1
+                            continue
+                    elif name not in encoder_modules:
+                        continue
+                    elif depth > 500:
+                        raise ValueError(
+                            "Max depth of recursive function `tie_encoder_to_decoder` reached. It seems that there is a circular dependency between two or more `nn.Modules` of your model."
+                        )
+                    else:
+                        decoder_name = encoder_name = name
+                    tie_encoder_to_decoder_recursively(
+                        decoder_modules[decoder_name],
+                        encoder_modules[encoder_name],
+                        module_name + "/" + name,
+                        uninitialized_encoder_weights,
+                        depth=depth + 1,
+                    )
+                    all_encoder_weights.remove(module_name + "/" + encoder_name)
+
+                uninitialized_encoder_weights += list(all_encoder_weights)
+
+        # tie weights recursively
+        tie_encoder_to_decoder_recursively(decoder, encoder, base_model_prefix, uninitialized_encoder_weights)
+        if len(uninitialized_encoder_weights) > 0:
+            logger.warning(
+                f"The following encoder weights were not tied to the decoder {uninitialized_encoder_weights}"
+            )
+
+    def _tie_or_clone_weights(self, output_embeddings, input_embeddings):
+        """Tie or clone module weights depending of whether we are using TorchScript or not"""
+        if self.config.torchscript:
+            output_embeddings.weight = nn.Parameter(input_embeddings.weight.clone())
+        else:
+            output_embeddings.weight = input_embeddings.weight
+
+        if getattr(output_embeddings, "bias", None) is not None:
+            output_embeddings.bias.data = torch.nn.functional.pad(
+                output_embeddings.bias.data,
+                (
+                    0,
+                    output_embeddings.weight.shape[0] - output_embeddings.bias.shape[0],
+                ),
+                "constant",
+                0,
+            )
+        if hasattr(output_embeddings, "out_features") and hasattr(input_embeddings, "num_embeddings"):
+            output_embeddings.out_features = input_embeddings.num_embeddings
+
+    def resize_token_embeddings(self, new_num_tokens: Optional[int] = None) -> torch.nn.Embedding:
+        """
+        Resizes input token embeddings matrix of the model if :obj:`new_num_tokens != config.vocab_size`.
+
+        Takes care of tying weights embeddings afterwards if the model class has a :obj:`tie_weights()` method.
+
+        Arguments:
+            new_num_tokens (:obj:`int`, `optional`):
+                The number of new tokens in the embedding matrix. Increasing the size will add newly initialized
+                vectors at the end. Reducing the size will remove vectors from the end. If not provided or :obj:`None`,
+                just returns a pointer to the input tokens :obj:`torch.nn.Embedding` module of the model without doing
+                anything.
+
+        Return:
+            :obj:`torch.nn.Embedding`: Pointer to the input tokens Embeddings Module of the model.
+        """
+        model_embeds = self._resize_token_embeddings(new_num_tokens)
+        if new_num_tokens is None:
+            return model_embeds
+
+        # Update base model and current model config
+        self.config.vocab_size = new_num_tokens
+        self.vocab_size = new_num_tokens
+
+        # Tie weights again if needed
+        self.tie_weights()
+
+        return model_embeds
+
+    def _resize_token_embeddings(self, new_num_tokens):
+        old_embeddings = self.get_input_embeddings()
+        new_embeddings = self._get_resized_embeddings(old_embeddings, new_num_tokens)
+        self.set_input_embeddings(new_embeddings)
+
+        # if word embeddings are not tied, make sure that lm head is resized as well
+        if self.get_output_embeddings() is not None and not self.config.tie_word_embeddings:
+            old_lm_head = self.get_output_embeddings()
+            new_lm_head = self._get_resized_lm_head(old_lm_head, new_num_tokens)
+            self.set_output_embeddings(new_lm_head)
+
+        return self.get_input_embeddings()
+
+    def _get_resized_embeddings(
+        self, old_embeddings: torch.nn.Embedding, new_num_tokens: Optional[int] = None
+    ) -> torch.nn.Embedding:
+        """
+        Build a resized Embedding Module from a provided token Embedding Module. Increasing the size will add newly
+        initialized vectors at the end. Reducing the size will remove vectors from the end
+
+        Args:
+            old_embeddings (:obj:`torch.nn.Embedding`):
+                Old embeddings to be resized.
+            new_num_tokens (:obj:`int`, `optional`):
+                New number of tokens in the embedding matrix.
+
+                Increasing the size will add newly initialized vectors at the end. Reducing the size will remove
+                vectors from the end. If not provided or :obj:`None`, just returns a pointer to the input tokens
+                :obj:`torch.nn.Embedding`` module of the model without doing anything.
+
+        Return:
+            :obj:`torch.nn.Embedding`: Pointer to the resized Embedding Module or the old Embedding Module if
+            :obj:`new_num_tokens` is :obj:`None`
+        """
+        if new_num_tokens is None:
+            return old_embeddings
+
+        old_num_tokens, old_embedding_dim = old_embeddings.weight.size()
+        if old_num_tokens == new_num_tokens:
+            return old_embeddings
+
+        if not isinstance(old_embeddings, nn.Embedding):
+            raise TypeError(
+                f"Old embeddings are of type {type(old_embeddings)}, which is not an instance of {nn.Embedding}."
+                f"You should either use a different resize function or make sure that `old_embeddings` are an instance of {nn.Embedding}."
+            )
+
+        # Build new embeddings
+        new_embeddings = nn.Embedding(new_num_tokens, old_embedding_dim).to(self.device)
+
+        # initialize all new embeddings (in particular added tokens)
+        self._init_weights(new_embeddings)
+
+        # Copy token embeddings from the previous weights
+        num_tokens_to_copy = min(old_num_tokens, new_num_tokens)
+        new_embeddings.weight.data[:num_tokens_to_copy, :] = old_embeddings.weight.data[:num_tokens_to_copy, :]
+
+        return new_embeddings
+
+    def _get_resized_lm_head(
+        self, old_lm_head: torch.nn.Linear, new_num_tokens: Optional[int] = None, transposed: Optional[bool] = False
+    ) -> torch.nn.Linear:
+        """
+        Build a resized Linear Module from a provided old Linear Module. Increasing the size will add newly initialized
+        vectors at the end. Reducing the size will remove vectors from the end
+
+        Args:
+            old_lm_head (:obj:`torch.nn.Linear`):
+                Old lm head liner layer to be resized.
+            new_num_tokens (:obj:`int`, `optional`):
+                New number of tokens in the linear matrix.
+
+                Increasing the size will add newly initialized vectors at the end. Reducing the size will remove
+                vectors from the end. If not provided or :obj:`None`, just returns a pointer to the input tokens
+                :obj:`torch.nn.Linear`` module of the model without doing anything.
+            transposed (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Whether ``old_lm_head`` is transposed or not. If True ``old_lm_head.size()`` is ``lm_head_dim,
+                vocab_size`` else ``vocab_size, lm_head_dim``.
+
+        Return:
+            :obj:`torch.nn.Linear`: Pointer to the resized Linear Module or the old Linear Module if
+            :obj:`new_num_tokens` is :obj:`None`
+        """
+        if new_num_tokens is None:
+            return old_lm_head
+
+        old_num_tokens, old_lm_head_dim = (
+            old_lm_head.weight.size() if not transposed else old_lm_head.weight.t().size()
+        )
+
+        if old_num_tokens == new_num_tokens:
+            return old_lm_head
+
+        if not isinstance(old_lm_head, nn.Linear):
+            raise TypeError(
+                f"Old language model head is of type {type(old_lm_head)}, which is not an instance of {nn.Linear}."
+                f"You should either use a different resize function or make sure that `old_embeddings` are an instance of {nn.Linear}."
+            )
+
+        # Build new lm head
+        new_lm_head_shape = (old_lm_head_dim, new_num_tokens) if not transposed else (new_num_tokens, old_lm_head_dim)
+        has_new_lm_head_bias = old_lm_head.bias is not None
+        new_lm_head = nn.Linear(*new_lm_head_shape, bias=has_new_lm_head_bias).to(self.device)
+
+        # initialize new lm head (in particular added tokens)
+        self._init_weights(new_lm_head)
+
+        num_tokens_to_copy = min(old_num_tokens, new_num_tokens)
+
+        # Copy old lm head weights to new lm head
+        if not transposed:
+            new_lm_head.weight.data[:num_tokens_to_copy, :] = old_lm_head.weight.data[:num_tokens_to_copy, :]
+        else:
+            new_lm_head.weight.data[:, :num_tokens_to_copy] = old_lm_head.weight.data[:, :num_tokens_to_copy]
+
+        # Copy bias weights to new lm head
+        if has_new_lm_head_bias:
+            new_lm_head.bias.data[:num_tokens_to_copy] = old_lm_head.bias.data[:num_tokens_to_copy]
+
+        return new_lm_head
+
+    def init_weights(self):
+        """
+        Initializes and prunes weights if needed.
+        """
+        # Initialize weights
+        self.apply(self._init_weights)
+
+        # Prune heads if needed
+        if self.config.pruned_heads:
+            self.prune_heads(self.config.pruned_heads)
+
+        # Tie weights if needed
+        self.tie_weights()
+
+    def prune_heads(self, heads_to_prune: Dict[int, List[int]]):
+        """
+        Prunes heads of the base model.
+
+        Arguments:
+            heads_to_prune (:obj:`Dict[int, List[int]]`):
+                Dictionary with keys being selected layer indices (:obj:`int`) and associated values being the list of
+                heads to prune in said layer (list of :obj:`int`). For instance {1: [0, 2], 2: [2, 3]} will prune heads
+                0 and 2 on layer 1 and heads 2 and 3 on layer 2.
+        """
+        # save new sets of pruned heads as union of previously stored pruned heads and newly pruned heads
+        for layer, heads in heads_to_prune.items():
+            union_heads = set(self.config.pruned_heads.get(layer, [])) | set(heads)
+            self.config.pruned_heads[layer] = list(union_heads)  # Unfortunately we have to store it as list for JSON
+
+        self.base_model._prune_heads(heads_to_prune)
+
+    def save_pretrained(
+        self,
+        save_directory: Union[str, os.PathLike],
+        save_config: bool = True,
+        state_dict: Optional[dict] = None,
+        save_function: Callable = torch.save,
+    ):
+        """
+        Save a model and its configuration file to a directory, so that it can be re-loaded using the
+        `:func:`~transformers.PreTrainedModel.from_pretrained`` class method.
+
+        Arguments:
+            save_directory (:obj:`str` or :obj:`os.PathLike`):
+                Directory to which to save. Will be created if it doesn't exist.
+            save_config (:obj:`bool`, `optional`, defaults to :obj:`True`):
+                Whether or not to save the config of the model. Useful when in distributed training like TPUs and need
+                to call this function on all processes. In this case, set :obj:`save_config=True` only on the main
+                process to avoid race conditions.
+            state_dict (nested dictionary of :obj:`torch.Tensor`):
+                The state dictionary of the model to save. Will default to :obj:`self.state_dict()`, but can be used to
+                only save parts of the model or if special precautions need to be taken when recovering the state
+                dictionary of a model (like when using model parallelism).
+            save_function (:obj:`Callable`):
+                The function to use to save the state dictionary. Useful on distributed training like TPUs when one
+                need to replace :obj:`torch.save` by another method.
+        """
+        if os.path.isfile(save_directory):
+            logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
+            return
+        os.makedirs(save_directory, exist_ok=True)
+
+        # Only save the model itself if we are using distributed training
+        model_to_save = unwrap_model(self)
+
+        # Attach architecture to the config
+        model_to_save.config.architectures = [model_to_save.__class__.__name__]
+
+        # Save the config
+        if save_config:
+            model_to_save.config.save_pretrained(save_directory)
+
+        # Save the model
+        if state_dict is None:
+            state_dict = model_to_save.state_dict()
+
+        # Handle the case where some state_dict keys shouldn't be saved
+        if self._keys_to_ignore_on_save is not None:
+            state_dict = {k: v for k, v in state_dict.items() if k not in self._keys_to_ignore_on_save}
+
+        # If we save using the predefined names, we can load using `from_pretrained`
+        output_model_file = os.path.join(save_directory, WEIGHTS_NAME)
+        save_function(state_dict, output_model_file)
+
+        logger.info("Model weights saved in {}".format(output_model_file))
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], *model_args, **kwargs):
+        r"""
+        Instantiate a pretrained pytorch model from a pre-trained model configuration.
+
+        The model is set in evaluation mode by default using ``model.eval()`` (Dropout modules are deactivated). To
+        train the model, you should first set it back in training mode with ``model.train()``.
+
+        The warning `Weights from XXX not initialized from pretrained model` means that the weights of XXX do not come
+        pretrained with the rest of the model. It is up to you to train those weights with a downstream fine-tuning
+        task.
+
+        The warning `Weights from XXX not used in YYY` means that the layer XXX is not used by YYY, therefore those
+        weights are discarded.
+
+        Parameters:
+            pretrained_model_name_or_path (:obj:`str` or :obj:`os.PathLike`, `optional`):
+                Can be either:
+
+                    - A string, the `model id` of a pretrained model hosted inside a model repo on huggingface.co.
+                      Valid model ids can be located at the root-level, like ``bert-base-uncased``, or namespaced under
+                      a user or organization name, like ``dbmdz/bert-base-german-cased``.
+                    - A path to a `directory` containing model weights saved using
+                      :func:`~transformers.PreTrainedModel.save_pretrained`, e.g., ``./my_model_directory/``.
+                    - A path or url to a `tensorflow index checkpoint file` (e.g, ``./tf_model/model.ckpt.index``). In
+                      this case, ``from_tf`` should be set to :obj:`True` and a configuration object should be provided
+                      as ``config`` argument. This loading path is slower than converting the TensorFlow checkpoint in
+                      a PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.
+                    - :obj:`None` if you are both providing the configuration and state dictionary (resp. with keyword
+                      arguments ``config`` and ``state_dict``).
+            model_args (sequence of positional arguments, `optional`):
+                All remaning positional arguments will be passed to the underlying model's ``__init__`` method.
+            config (:obj:`Union[PretrainedConfig, str, os.PathLike]`, `optional`):
+                Can be either:
+
+                    - an instance of a class derived from :class:`~transformers.PretrainedConfig`,
+                    - a string or path valid as input to :func:`~transformers.PretrainedConfig.from_pretrained`.
+
+                Configuration for the model to use instead of an automatically loaded configuation. Configuration can
+                be automatically loaded when:
+
+                    - The model is a model provided by the library (loaded with the `model id` string of a pretrained
+                      model).
+                    - The model was saved using :func:`~transformers.PreTrainedModel.save_pretrained` and is reloaded
+                      by supplying the save directory.
+                    - The model is loaded by supplying a local directory as ``pretrained_model_name_or_path`` and a
+                      configuration JSON file named `config.json` is found in the directory.
+            state_dict (:obj:`Dict[str, torch.Tensor]`, `optional`):
+                A state dictionary to use instead of a state dictionary loaded from saved weights file.
+
+                This option can be used if you want to create a model from a pretrained configuration but load your own
+                weights. In this case though, you should check if using
+                :func:`~transformers.PreTrainedModel.save_pretrained` and
+                :func:`~transformers.PreTrainedModel.from_pretrained` is not a simpler option.
+            cache_dir (:obj:`Union[str, os.PathLike]`, `optional`):
+                Path to a directory in which a downloaded pretrained model configuration should be cached if the
+                standard cache should not be used.
+            from_tf (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Load the model weights from a TensorFlow checkpoint save file (see docstring of
+                ``pretrained_model_name_or_path`` argument).
+            force_download (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            resume_download (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Whether or not to delete incompletely received files. Will attempt to resume the download if such a
+                file exists.
+            proxies (:obj:`Dict[str, str], `optional`):
+                A dictionary of proxy servers to use by protocol or endpoint, e.g., :obj:`{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            output_loading_info(:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Whether ot not to also return a dictionary containing missing keys, unexpected keys and error messages.
+            local_files_only(:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Whether or not to only look at local files (i.e., do not try to download the model).
+            use_auth_token (:obj:`str` or `bool`, `optional`):
+                The token to use as HTTP bearer authorization for remote files. If :obj:`True`, will use the token
+                generated when running :obj:`transformers-cli login` (stored in :obj:`~/.huggingface`).
+            revision(:obj:`str`, `optional`, defaults to :obj:`"main"`):
+                The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
+                git-based system for storing models and other artifacts on huggingface.co, so ``revision`` can be any
+                identifier allowed by git.
+            mirror(:obj:`str`, `optional`, defaults to :obj:`None`):
+                Mirror source to accelerate downloads in China. If you are from China and have an accessibility
+                problem, you can set this option to resolve it. Note that we do not guarantee the timeliness or safety.
+                Please refer to the mirror site for more information.
+            kwargs (remaining dictionary of keyword arguments, `optional`):
+                Can be used to update the configuration object (after it being loaded) and initiate the model (e.g.,
+                :obj:`output_attentions=True`). Behaves differently depending on whether a ``config`` is provided or
+                automatically loaded:
+
+                    - If a configuration is provided with ``config``, ``**kwargs`` will be directly passed to the
+                      underlying model's ``__init__`` method (we assume all relevant updates to the configuration have
+                      already been done)
+                    - If a configuration is not provided, ``kwargs`` will be first passed to the configuration class
+                      initialization function (:func:`~transformers.PretrainedConfig.from_pretrained`). Each key of
+                      ``kwargs`` that corresponds to a configuration attribute will be used to override said attribute
+                      with the supplied ``kwargs`` value. Remaining keys that do not correspond to any configuration
+                      attribute will be passed to the underlying model's ``__init__`` function.
+
+        .. note::
+
+            Passing :obj:`use_auth_token=True` is required when you want to use a private model.
+
+        Examples::
+
+            >>> from transformers import BertConfig, BertModel
+            >>> # Download model and configuration from huggingface.co and cache.
+            >>> model = BertModel.from_pretrained('bert-base-uncased')
+            >>> # Model was saved using `save_pretrained('./test/saved_model/')` (for example purposes, not runnable).
+            >>> model = BertModel.from_pretrained('./test/saved_model/')
+            >>> # Update configuration during loading.
+            >>> model = BertModel.from_pretrained('bert-base-uncased', output_attentions=True)
+            >>> assert model.config.output_attentions == True
+            >>> # Loading from a TF checkpoint file instead of a PyTorch model (slower, for example purposes, not runnable).
+            >>> config = BertConfig.from_json_file('./tf_model/my_tf_model_config.json')
+            >>> model = BertModel.from_pretrained('./tf_model/my_tf_checkpoint.ckpt.index', from_tf=True, config=config)
+        """
+        config = kwargs.pop("config", None)
+        state_dict = kwargs.pop("state_dict", None)
+        cache_dir = kwargs.pop("cache_dir", None)
+        from_tf = kwargs.pop("from_tf", False)
+        force_download = kwargs.pop("force_download", False)
+        resume_download = kwargs.pop("resume_download", False)
+        proxies = kwargs.pop("proxies", None)
+        output_loading_info = kwargs.pop("output_loading_info", False)
+        local_files_only = kwargs.pop("local_files_only", False)
+        use_auth_token = kwargs.pop("use_auth_token", None)
+        revision = kwargs.pop("revision", None)
+        mirror = kwargs.pop("mirror", None)
+
+        if is_offline_mode() and not local_files_only:
+            logger.info("Offline mode: forcing local_files_only=True")
+            local_files_only = True
+
+        # Load config if we don't provide a configuration
+        if not isinstance(config, PretrainedConfig):
+            config_path = config if config is not None else pretrained_model_name_or_path
+            config, model_kwargs = cls.config_class.from_pretrained(
+                config_path,
+                *model_args,
+                cache_dir=cache_dir,
+                return_unused_kwargs=True,
+                force_download=force_download,
+                resume_download=resume_download,
+                proxies=proxies,
+                local_files_only=local_files_only,
+                use_auth_token=use_auth_token,
+                revision=revision,
+                **kwargs,
+            )
+        else:
+            model_kwargs = kwargs
+
+        # Load model
+        if pretrained_model_name_or_path is not None:
+            pretrained_model_name_or_path = str(pretrained_model_name_or_path)
+            if os.path.isdir(pretrained_model_name_or_path):
+                if from_tf and os.path.isfile(os.path.join(pretrained_model_name_or_path, TF_WEIGHTS_NAME + ".index")):
+                    # Load from a TF 1.0 checkpoint in priority if from_tf
+                    archive_file = os.path.join(pretrained_model_name_or_path, TF_WEIGHTS_NAME + ".index")
+                elif from_tf and os.path.isfile(os.path.join(pretrained_model_name_or_path, TF2_WEIGHTS_NAME)):
+                    # Load from a TF 2.0 checkpoint in priority if from_tf
+                    archive_file = os.path.join(pretrained_model_name_or_path, TF2_WEIGHTS_NAME)
+                elif os.path.isfile(os.path.join(pretrained_model_name_or_path, WEIGHTS_NAME)):
+                    # Load from a PyTorch checkpoint
+                    archive_file = os.path.join(pretrained_model_name_or_path, WEIGHTS_NAME)
+                else:
+                    raise EnvironmentError(
+                        "Error no file named {} found in directory {} or `from_tf` set to False".format(
+                            [WEIGHTS_NAME, TF2_WEIGHTS_NAME, TF_WEIGHTS_NAME + ".index"],
+                            pretrained_model_name_or_path,
+                        )
+                    )
+            elif os.path.isfile(pretrained_model_name_or_path) or is_remote_url(pretrained_model_name_or_path):
+                archive_file = pretrained_model_name_or_path
+            elif os.path.isfile(pretrained_model_name_or_path + ".index"):
+                assert (
+                    from_tf
+                ), "We found a TensorFlow checkpoint at {}, please set from_tf to True to load from this checkpoint".format(
+                    pretrained_model_name_or_path + ".index"
+                )
+                archive_file = pretrained_model_name_or_path + ".index"
+            else:
+                archive_file = hf_bucket_url(
+                    pretrained_model_name_or_path,
+                    filename=(TF2_WEIGHTS_NAME if from_tf else WEIGHTS_NAME),
+                    revision=revision,
+                    mirror=mirror,
+                )
+
+            try:
+                # Load from URL or cache if already cached
+                resolved_archive_file = cached_path(
+                    archive_file,
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    proxies=proxies,
+                    resume_download=resume_download,
+                    local_files_only=local_files_only,
+                    use_auth_token=use_auth_token,
+                )
+            except EnvironmentError as err:
+                logger.error(err)
+                msg = (
+                    f"Can't load weights for '{pretrained_model_name_or_path}'. Make sure that:\n\n"
+                    f"- '{pretrained_model_name_or_path}' is a correct model identifier listed on 'https://huggingface.co/models'\n\n"
+                    f"- or '{pretrained_model_name_or_path}' is the correct path to a directory containing a file named one of {WEIGHTS_NAME}, {TF2_WEIGHTS_NAME}, {TF_WEIGHTS_NAME}.\n\n"
+                )
+                raise EnvironmentError(msg)
+
+            if resolved_archive_file == archive_file:
+                logger.info("loading weights file {}".format(archive_file))
+            else:
+                logger.info("loading weights file {} from cache at {}".format(archive_file, resolved_archive_file))
+        else:
+            resolved_archive_file = None
+
+        config.name_or_path = pretrained_model_name_or_path
+
+        # Instantiate model.
+        model = cls(config, *model_args, **model_kwargs)
+
+        if state_dict is None and not from_tf:
+            try:
+                state_dict = torch.load(resolved_archive_file, map_location="cpu")
+            except Exception:
+                raise OSError(
+                    f"Unable to load weights from pytorch checkpoint file for '{pretrained_model_name_or_path}' "
+                    f"at '{resolved_archive_file}'"
+                    "If you tried to load a PyTorch model from a TF 2.0 checkpoint, please set from_tf=True. "
+                )
+
+        missing_keys = []
+        unexpected_keys = []
+        error_msgs = []
+
+        if from_tf:
+            if resolved_archive_file.endswith(".index"):
+                # Load from a TensorFlow 1.X checkpoint - provided by original authors
+                model = cls.load_tf_weights(model, config, resolved_archive_file[:-6])  # Remove the '.index'
+            else:
+                # Load from our TensorFlow 2.0 checkpoints
+                try:
+                    from .modeling_tf_pytorch_utils import load_tf2_checkpoint_in_pytorch_model
+
+                    model = load_tf2_checkpoint_in_pytorch_model(model, resolved_archive_file, allow_missing_keys=True)
+                except ImportError:
+                    logger.error(
+                        "Loading a TensorFlow model in PyTorch, requires both PyTorch and TensorFlow to be installed. Please see "
+                        "https://pytorch.org/ and https://www.tensorflow.org/install/ for installation instructions."
+                    )
+                    raise
+        else:
+            # Convert old format to new format if needed from a PyTorch state_dict
+            old_keys = []
+            new_keys = []
+            for key in state_dict.keys():
+                new_key = None
+                if "gamma" in key:
+                    new_key = key.replace("gamma", "weight")
+                if "beta" in key:
+                    new_key = key.replace("beta", "bias")
+                if new_key:
+                    old_keys.append(key)
+                    new_keys.append(new_key)
+            for old_key, new_key in zip(old_keys, new_keys):
+                state_dict[new_key] = state_dict.pop(old_key)
+
+            # copy state_dict so _load_from_state_dict can modify it
+            metadata = getattr(state_dict, "_metadata", None)
+            state_dict = state_dict.copy()
+            if metadata is not None:
+                state_dict._metadata = metadata
+
+            # PyTorch's `_load_from_state_dict` does not copy parameters in a module's descendants
+            # so we need to apply the function recursively.
+            def load(module: nn.Module, prefix=""):
+                local_metadata = {} if metadata is None else metadata.get(prefix[:-1], {})
+                module._load_from_state_dict(
+                    state_dict,
+                    prefix,
+                    local_metadata,
+                    True,
+                    missing_keys,
+                    unexpected_keys,
+                    error_msgs,
+                )
+                for name, child in module._modules.items():
+                    if child is not None:
+                        load(child, prefix + name + ".")
+
+            # Make sure we are able to load base models as well as derived models (with heads)
+            start_prefix = ""
+            model_to_load = model
+            has_prefix_module = any(s.startswith(cls.base_model_prefix) for s in state_dict.keys())
+            if not hasattr(model, cls.base_model_prefix) and has_prefix_module:
+                start_prefix = cls.base_model_prefix + "."
+            if hasattr(model, cls.base_model_prefix) and not has_prefix_module:
+                model_to_load = getattr(model, cls.base_model_prefix)
+
+            load(model_to_load, prefix=start_prefix)
+
+            if model.__class__.__name__ != model_to_load.__class__.__name__:
+                base_model_state_dict = model_to_load.state_dict().keys()
+                head_model_state_dict_without_base_prefix = [
+                    key.split(cls.base_model_prefix + ".")[-1] for key in model.state_dict().keys()
+                ]
+                missing_keys.extend(head_model_state_dict_without_base_prefix - base_model_state_dict)
+
+            # Some models may have keys that are not in the state by design, removing them before needlessly warning
+            # the user.
+            if cls._keys_to_ignore_on_load_missing is not None:
+                for pat in cls._keys_to_ignore_on_load_missing:
+                    missing_keys = [k for k in missing_keys if re.search(pat, k) is None]
+
+            if cls._keys_to_ignore_on_load_unexpected is not None:
+                for pat in cls._keys_to_ignore_on_load_unexpected:
+                    unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None]
+
+            if len(unexpected_keys) > 0:
+                logger.warning(
+                    f"Some weights of the model checkpoint at {pretrained_model_name_or_path} were not used when "
+                    f"initializing {model.__class__.__name__}: {unexpected_keys}\n"
+                    f"- This IS expected if you are initializing {model.__class__.__name__} from the checkpoint of a model trained on another task "
+                    f"or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).\n"
+                    f"- This IS NOT expected if you are initializing {model.__class__.__name__} from the checkpoint of a model that you expect "
+                    f"to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model)."
+                )
+            else:
+                logger.info(f"All model checkpoint weights were used when initializing {model.__class__.__name__}.\n")
+            if len(missing_keys) > 0:
+                logger.warning(
+                    f"Some weights of {model.__class__.__name__} were not initialized from the model checkpoint at {pretrained_model_name_or_path} "
+                    f"and are newly initialized: {missing_keys}\n"
+                    f"You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference."
+                )
+            else:
+                logger.info(
+                    f"All the weights of {model.__class__.__name__} were initialized from the model checkpoint at {pretrained_model_name_or_path}.\n"
+                    f"If your task is similar to the task the model of the checkpoint was trained on, "
+                    f"you can already use {model.__class__.__name__} for predictions without further training."
+                )
+            if len(error_msgs) > 0:
+                raise RuntimeError(
+                    "Error(s) in loading state_dict for {}:\n\t{}".format(
+                        model.__class__.__name__, "\n\t".join(error_msgs)
+                    )
+                )
+        # make sure token embedding weights are still tied if needed
+        model.tie_weights()
+
+        # Set model in evaluation mode to deactivate DropOut modules by default
+        model.eval()
+
+        if output_loading_info:
+            loading_info = {
+                "missing_keys": missing_keys,
+                "unexpected_keys": unexpected_keys,
+                "error_msgs": error_msgs,
+            }
+            return model, loading_info
+
+        return model
+
+
+class Conv1D(nn.Module):
+    """
+    1D-convolutional layer as defined by Radford et al. for OpenAI GPT (and also used in GPT-2).
+
+    Basically works like a linear layer but the weights are transposed.
+
+    Args:
+        nf (:obj:`int`): The number of output features.
+        nx (:obj:`int`): The number of input features.
+    """
+
+    def __init__(self, nf, nx):
+        super().__init__()
+        self.nf = nf
+        w = torch.empty(nx, nf)
+        nn.init.normal_(w, std=0.02)
+        self.weight = nn.Parameter(w)
+        self.bias = nn.Parameter(torch.zeros(nf))
+
+    def forward(self, x):
+        size_out = x.size()[:-1] + (self.nf,)
+        x = torch.addmm(self.bias, x.view(-1, x.size(-1)), self.weight)
+        x = x.view(*size_out)
+        return x
+
+
+class PoolerStartLogits(nn.Module):
+    """
+    Compute SQuAD start logits from sequence hidden states.
+
+    Args:
+        config (:class:`~transformers.PretrainedConfig`):
+            The config used by the model, will be used to grab the :obj:`hidden_size` of the model.
+    """
+
+    def __init__(self, config: PretrainedConfig):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, 1)
+
+    def forward(
+        self, hidden_states: torch.FloatTensor, p_mask: Optional[torch.FloatTensor] = None
+    ) -> torch.FloatTensor:
+        """
+        Args:
+            hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, seq_len, hidden_size)`):
+                The final hidden states of the model.
+            p_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, seq_len)`, `optional`):
+                Mask for tokens at invalid position, such as query and special symbols (PAD, SEP, CLS). 1.0 means token
+                should be masked.
+
+        Returns:
+            :obj:`torch.FloatTensor`: The start logits for SQuAD.
+        """
+        x = self.dense(hidden_states).squeeze(-1)
+
+        if p_mask is not None:
+            if get_parameter_dtype(self) == torch.float16:
+                x = x * (1 - p_mask) - 65500 * p_mask
+            else:
+                x = x * (1 - p_mask) - 1e30 * p_mask
+
+        return x
+
+
+class PoolerEndLogits(nn.Module):
+    """
+    Compute SQuAD end logits from sequence hidden states.
+
+    Args:
+        config (:class:`~transformers.PretrainedConfig`):
+            The config used by the model, will be used to grab the :obj:`hidden_size` of the model and the
+            :obj:`layer_norm_eps` to use.
+    """
+
+    def __init__(self, config: PretrainedConfig):
+        super().__init__()
+        self.dense_0 = nn.Linear(config.hidden_size * 2, config.hidden_size)
+        self.activation = nn.Tanh()
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dense_1 = nn.Linear(config.hidden_size, 1)
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        start_states: Optional[torch.FloatTensor] = None,
+        start_positions: Optional[torch.LongTensor] = None,
+        p_mask: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        """
+        Args:
+            hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, seq_len, hidden_size)`):
+                The final hidden states of the model.
+            start_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, seq_len, hidden_size)`, `optional`):
+                The hidden states of the first tokens for the labeled span.
+            start_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
+                The position of the first token for the labeled span.
+            p_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, seq_len)`, `optional`):
+                Mask for tokens at invalid position, such as query and special symbols (PAD, SEP, CLS). 1.0 means token
+                should be masked.
+
+        .. note::
+
+            One of ``start_states`` or ``start_positions`` should be not obj:`None`. If both are set,
+            ``start_positions`` overrides ``start_states``.
+
+        Returns:
+            :obj:`torch.FloatTensor`: The end logits for SQuAD.
+        """
+        assert (
+            start_states is not None or start_positions is not None
+        ), "One of start_states, start_positions should be not None"
+        if start_positions is not None:
+            slen, hsz = hidden_states.shape[-2:]
+            start_positions = start_positions[:, None, None].expand(-1, -1, hsz)  # shape (bsz, 1, hsz)
+            start_states = hidden_states.gather(-2, start_positions)  # shape (bsz, 1, hsz)
+            start_states = start_states.expand(-1, slen, -1)  # shape (bsz, slen, hsz)
+
+        x = self.dense_0(torch.cat([hidden_states, start_states], dim=-1))
+        x = self.activation(x)
+        x = self.LayerNorm(x)
+        x = self.dense_1(x).squeeze(-1)
+
+        if p_mask is not None:
+            if get_parameter_dtype(self) == torch.float16:
+                x = x * (1 - p_mask) - 65500 * p_mask
+            else:
+                x = x * (1 - p_mask) - 1e30 * p_mask
+
+        return x
+
+
+class PoolerAnswerClass(nn.Module):
+    """
+    Compute SQuAD 2.0 answer class from classification and start tokens hidden states.
+
+    Args:
+        config (:class:`~transformers.PretrainedConfig`):
+            The config used by the model, will be used to grab the :obj:`hidden_size` of the model.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        self.dense_0 = nn.Linear(config.hidden_size * 2, config.hidden_size)
+        self.activation = nn.Tanh()
+        self.dense_1 = nn.Linear(config.hidden_size, 1, bias=False)
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        start_states: Optional[torch.FloatTensor] = None,
+        start_positions: Optional[torch.LongTensor] = None,
+        cls_index: Optional[torch.LongTensor] = None,
+    ) -> torch.FloatTensor:
+        """
+        Args:
+            hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, seq_len, hidden_size)`):
+                The final hidden states of the model.
+            start_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, seq_len, hidden_size)`, `optional`):
+                The hidden states of the first tokens for the labeled span.
+            start_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
+                The position of the first token for the labeled span.
+            cls_index (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
+                Position of the CLS token for each sentence in the batch. If :obj:`None`, takes the last token.
+
+        .. note::
+
+            One of ``start_states`` or ``start_positions`` should be not obj:`None`. If both are set,
+            ``start_positions`` overrides ``start_states``.
+
+        Returns:
+            :obj:`torch.FloatTensor`: The SQuAD 2.0 answer class.
+        """
+        # No dependency on end_feature so that we can obtain one single `cls_logits` for each sample.
+        hsz = hidden_states.shape[-1]
+        assert (
+            start_states is not None or start_positions is not None
+        ), "One of start_states, start_positions should be not None"
+        if start_positions is not None:
+            start_positions = start_positions[:, None, None].expand(-1, -1, hsz)  # shape (bsz, 1, hsz)
+            start_states = hidden_states.gather(-2, start_positions).squeeze(-2)  # shape (bsz, hsz)
+
+        if cls_index is not None:
+            cls_index = cls_index[:, None, None].expand(-1, -1, hsz)  # shape (bsz, 1, hsz)
+            cls_token_state = hidden_states.gather(-2, cls_index).squeeze(-2)  # shape (bsz, hsz)
+        else:
+            cls_token_state = hidden_states[:, -1, :]  # shape (bsz, hsz)
+
+        x = self.dense_0(torch.cat([start_states, cls_token_state], dim=-1))
+        x = self.activation(x)
+        x = self.dense_1(x).squeeze(-1)
+
+        return x
+
+
+@dataclass
+class SquadHeadOutput(ModelOutput):
+    """
+    Base class for outputs of question answering models using a :class:`~transformers.modeling_utils.SQuADHead`.
+
+    Args:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned if both :obj:`start_positions` and :obj:`end_positions` are provided):
+            Classification loss as the sum of start token, end token (and is_impossible if provided) classification
+            losses.
+        start_top_log_probs (``torch.FloatTensor`` of shape ``(batch_size, config.start_n_top)``, `optional`, returned if ``start_positions`` or ``end_positions`` is not provided):
+            Log probabilities for the top config.start_n_top start token possibilities (beam-search).
+        start_top_index (``torch.LongTensor`` of shape ``(batch_size, config.start_n_top)``, `optional`, returned if ``start_positions`` or ``end_positions`` is not provided):
+            Indices for the top config.start_n_top start token possibilities (beam-search).
+        end_top_log_probs (``torch.FloatTensor`` of shape ``(batch_size, config.start_n_top * config.end_n_top)``, `optional`, returned if ``start_positions`` or ``end_positions`` is not provided):
+            Log probabilities for the top ``config.start_n_top * config.end_n_top`` end token possibilities
+            (beam-search).
+        end_top_index (``torch.LongTensor`` of shape ``(batch_size, config.start_n_top * config.end_n_top)``, `optional`, returned if ``start_positions`` or ``end_positions`` is not provided):
+            Indices for the top ``config.start_n_top * config.end_n_top`` end token possibilities (beam-search).
+        cls_logits (``torch.FloatTensor`` of shape ``(batch_size,)``, `optional`, returned if ``start_positions`` or ``end_positions`` is not provided):
+            Log probabilities for the ``is_impossible`` label of the answers.
+
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    start_top_log_probs: Optional[torch.FloatTensor] = None
+    start_top_index: Optional[torch.LongTensor] = None
+    end_top_log_probs: Optional[torch.FloatTensor] = None
+    end_top_index: Optional[torch.LongTensor] = None
+    cls_logits: Optional[torch.FloatTensor] = None
+
+
+class SQuADHead(nn.Module):
+    r"""
+    A SQuAD head inspired by XLNet.
+
+    Args:
+        config (:class:`~transformers.PretrainedConfig`):
+            The config used by the model, will be used to grab the :obj:`hidden_size` of the model and the
+            :obj:`layer_norm_eps` to use.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        self.start_n_top = config.start_n_top
+        self.end_n_top = config.end_n_top
+
+        self.start_logits = PoolerStartLogits(config)
+        self.end_logits = PoolerEndLogits(config)
+        self.answer_class = PoolerAnswerClass(config)
+
+    @replace_return_docstrings(output_type=SquadHeadOutput, config_class=PretrainedConfig)
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        start_positions: Optional[torch.LongTensor] = None,
+        end_positions: Optional[torch.LongTensor] = None,
+        cls_index: Optional[torch.LongTensor] = None,
+        is_impossible: Optional[torch.LongTensor] = None,
+        p_mask: Optional[torch.FloatTensor] = None,
+        return_dict: bool = False,
+    ) -> Union[SquadHeadOutput, Tuple[torch.FloatTensor]]:
+        """
+        Args:
+            hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, seq_len, hidden_size)`):
+                Final hidden states of the model on the sequence tokens.
+            start_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
+                Positions of the first token for the labeled span.
+            end_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
+                Positions of the last token for the labeled span.
+            cls_index (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
+                Position of the CLS token for each sentence in the batch. If :obj:`None`, takes the last token.
+            is_impossible (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
+                Whether the question has a possible answer in the paragraph or not.
+            p_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, seq_len)`, `optional`):
+                Mask for tokens at invalid position, such as query and special symbols (PAD, SEP, CLS). 1.0 means token
+                should be masked.
+            return_dict (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.
+
+        Returns:
+        """
+        start_logits = self.start_logits(hidden_states, p_mask=p_mask)
+
+        if start_positions is not None and end_positions is not None:
+            # If we are on multi-GPU, let's remove the dimension added by batch splitting
+            for x in (start_positions, end_positions, cls_index, is_impossible):
+                if x is not None and x.dim() > 1:
+                    x.squeeze_(-1)
+
+            # during training, compute the end logits based on the ground truth of the start position
+            end_logits = self.end_logits(hidden_states, start_positions=start_positions, p_mask=p_mask)
+
+            loss_fct = CrossEntropyLoss()
+            start_loss = loss_fct(start_logits, start_positions)
+            end_loss = loss_fct(end_logits, end_positions)
+            total_loss = (start_loss + end_loss) / 2
+
+            if cls_index is not None and is_impossible is not None:
+                # Predict answerability from the representation of CLS and START
+                cls_logits = self.answer_class(hidden_states, start_positions=start_positions, cls_index=cls_index)
+                loss_fct_cls = nn.BCEWithLogitsLoss()
+                cls_loss = loss_fct_cls(cls_logits, is_impossible)
+
+                # note(zhiliny): by default multiply the loss by 0.5 so that the scale is comparable to start_loss and end_loss
+                total_loss += cls_loss * 0.5
+
+            return SquadHeadOutput(loss=total_loss) if return_dict else (total_loss,)
+
+        else:
+            # during inference, compute the end logits based on beam search
+            bsz, slen, hsz = hidden_states.size()
+            start_log_probs = F.softmax(start_logits, dim=-1)  # shape (bsz, slen)
+
+            start_top_log_probs, start_top_index = torch.topk(
+                start_log_probs, self.start_n_top, dim=-1
+            )  # shape (bsz, start_n_top)
+            start_top_index_exp = start_top_index.unsqueeze(-1).expand(-1, -1, hsz)  # shape (bsz, start_n_top, hsz)
+            start_states = torch.gather(hidden_states, -2, start_top_index_exp)  # shape (bsz, start_n_top, hsz)
+            start_states = start_states.unsqueeze(1).expand(-1, slen, -1, -1)  # shape (bsz, slen, start_n_top, hsz)
+
+            hidden_states_expanded = hidden_states.unsqueeze(2).expand_as(
+                start_states
+            )  # shape (bsz, slen, start_n_top, hsz)
+            p_mask = p_mask.unsqueeze(-1) if p_mask is not None else None
+            end_logits = self.end_logits(hidden_states_expanded, start_states=start_states, p_mask=p_mask)
+            end_log_probs = F.softmax(end_logits, dim=1)  # shape (bsz, slen, start_n_top)
+
+            end_top_log_probs, end_top_index = torch.topk(
+                end_log_probs, self.end_n_top, dim=1
+            )  # shape (bsz, end_n_top, start_n_top)
+            end_top_log_probs = end_top_log_probs.view(-1, self.start_n_top * self.end_n_top)
+            end_top_index = end_top_index.view(-1, self.start_n_top * self.end_n_top)
+
+            start_states = torch.einsum("blh,bl->bh", hidden_states, start_log_probs)
+            cls_logits = self.answer_class(hidden_states, start_states=start_states, cls_index=cls_index)
+
+            if not return_dict:
+                return (start_top_log_probs, start_top_index, end_top_log_probs, end_top_index, cls_logits)
+            else:
+                return SquadHeadOutput(
+                    start_top_log_probs=start_top_log_probs,
+                    start_top_index=start_top_index,
+                    end_top_log_probs=end_top_log_probs,
+                    end_top_index=end_top_index,
+                    cls_logits=cls_logits,
+                )
+
+
+class SequenceSummary(nn.Module):
+    r"""
+    Compute a single vector summary of a sequence hidden states.
+
+    Args:
+        config (:class:`~transformers.PretrainedConfig`):
+            The config used by the model. Relevant arguments in the config class of the model are (refer to the actual
+            config class of your model for the default values it uses):
+
+            - **summary_type** (:obj:`str`) -- The method to use to make this summary. Accepted values are:
+
+                - :obj:`"last"` -- Take the last token hidden state (like XLNet)
+                - :obj:`"first"` -- Take the first token hidden state (like Bert)
+                - :obj:`"mean"` -- Take the mean of all tokens hidden states
+                - :obj:`"cls_index"` -- Supply a Tensor of classification token position (GPT/GPT-2)
+                - :obj:`"attn"` -- Not implemented now, use multi-head attention
+
+            - **summary_use_proj** (:obj:`bool`) -- Add a projection after the vector extraction.
+            - **summary_proj_to_labels** (:obj:`bool`) -- If :obj:`True`, the projection outputs to
+              :obj:`config.num_labels` classes (otherwise to :obj:`config.hidden_size`).
+            - **summary_activation** (:obj:`Optional[str]`) -- Set to :obj:`"tanh"` to add a tanh activation to the
+              output, another string or :obj:`None` will add no activation.
+            - **summary_first_dropout** (:obj:`float`) -- Optional dropout probability before the projection and
+              activation.
+            - **summary_last_dropout** (:obj:`float`)-- Optional dropout probability after the projection and
+              activation.
+    """
+
+    def __init__(self, config: PretrainedConfig):
+        super().__init__()
+
+        self.summary_type = getattr(config, "summary_type", "last")
+        if self.summary_type == "attn":
+            # We should use a standard multi-head attention module with absolute positional embedding for that.
+            # Cf. https://github.com/zihangdai/xlnet/blob/master/modeling.py#L253-L276
+            # We can probably just use the multi-head attention module of PyTorch >=1.1.0
+            raise NotImplementedError
+
+        self.summary = Identity()
+        if hasattr(config, "summary_use_proj") and config.summary_use_proj:
+            if hasattr(config, "summary_proj_to_labels") and config.summary_proj_to_labels and config.num_labels > 0:
+                num_classes = config.num_labels
+            else:
+                num_classes = config.hidden_size
+            self.summary = nn.Linear(config.hidden_size, num_classes)
+
+        activation_string = getattr(config, "summary_activation", None)
+        self.activation: Callable = get_activation(activation_string) if activation_string else Identity()
+
+        self.first_dropout = Identity()
+        if hasattr(config, "summary_first_dropout") and config.summary_first_dropout > 0:
+            self.first_dropout = nn.Dropout(config.summary_first_dropout)
+
+        self.last_dropout = Identity()
+        if hasattr(config, "summary_last_dropout") and config.summary_last_dropout > 0:
+            self.last_dropout = nn.Dropout(config.summary_last_dropout)
+
+    def forward(
+        self, hidden_states: torch.FloatTensor, cls_index: Optional[torch.LongTensor] = None
+    ) -> torch.FloatTensor:
+        """
+        Compute a single vector summary of a sequence hidden states.
+
+        Args:
+            hidden_states (:obj:`torch.FloatTensor` of shape :obj:`[batch_size, seq_len, hidden_size]`):
+                The hidden states of the last layer.
+            cls_index (:obj:`torch.LongTensor` of shape :obj:`[batch_size]` or :obj:`[batch_size, ...]` where ... are optional leading dimensions of :obj:`hidden_states`, `optional`):
+                Used if :obj:`summary_type == "cls_index"` and takes the last token of the sequence as classification
+                token.
+
+        Returns:
+            :obj:`torch.FloatTensor`: The summary of the sequence hidden states.
+        """
+        if self.summary_type == "last":
+            output = hidden_states[:, -1]
+        elif self.summary_type == "first":
+            output = hidden_states[:, 0]
+        elif self.summary_type == "mean":
+            output = hidden_states.mean(dim=1)
+        elif self.summary_type == "cls_index":
+            if cls_index is None:
+                cls_index = torch.full_like(
+                    hidden_states[..., :1, :],
+                    hidden_states.shape[-2] - 1,
+                    dtype=torch.long,
+                )
+            else:
+                cls_index = cls_index.unsqueeze(-1).unsqueeze(-1)
+                cls_index = cls_index.expand((-1,) * (cls_index.dim() - 1) + (hidden_states.size(-1),))
+            # shape of cls_index: (bsz, XX, 1, hidden_size) where XX are optional leading dim of hidden_states
+            output = hidden_states.gather(-2, cls_index).squeeze(-2)  # shape (bsz, XX, hidden_size)
+        elif self.summary_type == "attn":
+            raise NotImplementedError
+
+        output = self.first_dropout(output)
+        output = self.summary(output)
+        output = self.activation(output)
+        output = self.last_dropout(output)
+
+        return output
+
+
+def unwrap_model(model: torch.nn.Module) -> torch.nn.Module:
+    """
+    Recursively unwraps a model from potential containers (as used in distributed training).
+
+    Args:
+        model (:obj:`torch.nn.Module`): The model to unwrap.
+    """
+    # since there could be multiple levels of wrapping, unwrap recursively
+    if hasattr(model, "module"):
+        return unwrap_model(model.module)
+    else:
+        return model
+
+
+def prune_linear_layer(layer: torch.nn.Linear, index: torch.LongTensor, dim: int = 0) -> torch.nn.Linear:
+    """
+    Prune a linear layer to keep only entries in index.
+
+    Used to remove heads.
+
+    Args:
+        layer (:obj:`torch.nn.Linear`): The layer to prune.
+        index (:obj:`torch.LongTensor`): The indices to keep in the layer.
+        dim (:obj:`int`, `optional`, defaults to 0): The dimension on which to keep the indices.
+
+    Returns:
+        :obj:`torch.nn.Linear`: The pruned layer as a new layer with :obj:`requires_grad=True`.
+    """
+    index = index.to(layer.weight.device)
+    W = layer.weight.index_select(dim, index).clone().detach()
+    if layer.bias is not None:
+        if dim == 1:
+            b = layer.bias.clone().detach()
+        else:
+            b = layer.bias[index].clone().detach()
+    new_size = list(layer.weight.size())
+    new_size[dim] = len(index)
+    new_layer = nn.Linear(new_size[1], new_size[0], bias=layer.bias is not None).to(layer.weight.device)
+    new_layer.weight.requires_grad = False
+    new_layer.weight.copy_(W.contiguous())
+    new_layer.weight.requires_grad = True
+    if layer.bias is not None:
+        new_layer.bias.requires_grad = False
+        new_layer.bias.copy_(b.contiguous())
+        new_layer.bias.requires_grad = True
+    return new_layer
+
+
+def prune_conv1d_layer(layer: Conv1D, index: torch.LongTensor, dim: int = 1) -> Conv1D:
+    """
+    Prune a Conv1D layer to keep only entries in index. A Conv1D work as a Linear layer (see e.g. BERT) but the weights
+    are transposed.
+
+    Used to remove heads.
+
+    Args:
+        layer (:class:`~transformers.modeling_utils.Conv1D`): The layer to prune.
+        index (:obj:`torch.LongTensor`): The indices to keep in the layer.
+        dim (:obj:`int`, `optional`, defaults to 1): The dimension on which to keep the indices.
+
+    Returns:
+        :class:`~transformers.modeling_utils.Conv1D`: The pruned layer as a new layer with :obj:`requires_grad=True`.
+    """
+    index = index.to(layer.weight.device)
+    W = layer.weight.index_select(dim, index).clone().detach()
+    if dim == 0:
+        b = layer.bias.clone().detach()
+    else:
+        b = layer.bias[index].clone().detach()
+    new_size = list(layer.weight.size())
+    new_size[dim] = len(index)
+    new_layer = Conv1D(new_size[1], new_size[0]).to(layer.weight.device)
+    new_layer.weight.requires_grad = False
+    new_layer.weight.copy_(W.contiguous())
+    new_layer.weight.requires_grad = True
+    new_layer.bias.requires_grad = False
+    new_layer.bias.copy_(b.contiguous())
+    new_layer.bias.requires_grad = True
+    return new_layer
+
+
+def prune_layer(
+    layer: Union[torch.nn.Linear, Conv1D], index: torch.LongTensor, dim: Optional[int] = None
+) -> Union[torch.nn.Linear, Conv1D]:
+    """
+    Prune a Conv1D or linear layer to keep only entries in index.
+
+    Used to remove heads.
+
+    Args:
+        layer (:obj:`Union[torch.nn.Linear, Conv1D]`): The layer to prune.
+        index (:obj:`torch.LongTensor`): The indices to keep in the layer.
+        dim (:obj:`int`, `optional`): The dimension on which to keep the indices.
+
+    Returns:
+        :obj:`torch.nn.Linear` or :class:`~transformers.modeling_utils.Conv1D`: The pruned layer as a new layer with
+        :obj:`requires_grad=True`.
+    """
+    if isinstance(layer, nn.Linear):
+        return prune_linear_layer(layer, index, dim=0 if dim is None else dim)
+    elif isinstance(layer, Conv1D):
+        return prune_conv1d_layer(layer, index, dim=1 if dim is None else dim)
+    else:
+        raise ValueError("Can't prune layer of class {}".format(layer.__class__))
+
+
+def apply_chunking_to_forward(
+    forward_fn: Callable[..., torch.Tensor], chunk_size: int, chunk_dim: int, *input_tensors
+) -> torch.Tensor:
+    """
+    This function chunks the :obj:`input_tensors` into smaller input tensor parts of size :obj:`chunk_size` over the
+    dimension :obj:`chunk_dim`. It then applies a layer :obj:`forward_fn` to each chunk independently to save memory.
+
+    If the :obj:`forward_fn` is independent across the :obj:`chunk_dim` this function will yield the same result as
+    directly applying :obj:`forward_fn` to :obj:`input_tensors`.
+
+    Args:
+        forward_fn (:obj:`Callable[..., torch.Tensor]`):
+            The forward function of the model.
+        chunk_size (:obj:`int`):
+            The chunk size of a chunked tensor: :obj:`num_chunks = len(input_tensors[0]) / chunk_size`.
+        chunk_dim (:obj:`int`):
+            The dimension over which the :obj:`input_tensors` should be chunked.
+        input_tensors (:obj:`Tuple[torch.Tensor]`):
+            The input tensors of ``forward_fn`` which will be chunked
+
+    Returns:
+        :obj:`torch.Tensor`: A tensor with the same shape as the :obj:`forward_fn` would have given if applied`.
+
+
+    Examples::
+
+        # rename the usual forward() fn to forward_chunk()
+        def forward_chunk(self, hidden_states):
+            hidden_states = self.decoder(hidden_states)
+            return hidden_states
+
+        # implement a chunked forward function
+        def forward(self, hidden_states):
+            return apply_chunking_to_forward(self.forward_chunk, self.chunk_size_lm_head, self.seq_len_dim, hidden_states)
+    """
+
+    assert len(input_tensors) > 0, "{} has to be a tuple/list of tensors".format(input_tensors)
+    tensor_shape = input_tensors[0].shape[chunk_dim]
+    assert all(
+        input_tensor.shape[chunk_dim] == tensor_shape for input_tensor in input_tensors
+    ), "All input tenors have to be of the same shape"
+
+    # inspect.signature exist since python 3.5 and is a python method -> no problem with backward compatibility
+    num_args_in_forward_chunk_fn = len(inspect.signature(forward_fn).parameters)
+    assert num_args_in_forward_chunk_fn == len(
+        input_tensors
+    ), "forward_chunk_fn expects {} arguments, but only {} input tensors are given".format(
+        num_args_in_forward_chunk_fn, len(input_tensors)
+    )
+
+    if chunk_size > 0:
+        assert (
+            input_tensors[0].shape[chunk_dim] % chunk_size == 0
+        ), "The dimension to be chunked {} has to be a multiple of the chunk size {}".format(
+            input_tensors[0].shape[chunk_dim], chunk_size
+        )
+
+        num_chunks = input_tensors[0].shape[chunk_dim] // chunk_size
+
+        # chunk input tensor into tuples
+        input_tensors_chunks = tuple(input_tensor.chunk(num_chunks, dim=chunk_dim) for input_tensor in input_tensors)
+        # apply forward fn to every tuple
+        output_chunks = tuple(forward_fn(*input_tensors_chunk) for input_tensors_chunk in zip(*input_tensors_chunks))
+        # concatenate output at same dimension
+        return torch.cat(output_chunks, dim=chunk_dim)
+
+    return forward_fn(*input_tensors)

run_interaction.py

+from argparse import ArgumentParser
+from dataset import Data_Encoder
+import torch
+from transformers import BertModel, BertTokenizer, BertLayer
+from transformers import AutoTokenizer, AutoModel
+# from modeling_bert import BertModel
+
+
+
+parser = ArgumentParser(description='BertAff Training.')
+parser.add_argument('-b', '--batch-size', default=16, type=int,
+                    metavar='N',
+                    help='mini-batch size (default: 16), this is the total '
+                         'batch size of all GPUs on the current node when '
+                         'using Data Parallel or Distributed Data Parallel')
+parser.add_argument('-j', '--workers', default=0, type=int, metavar='N',
+                    help='number of data loading workers (default: 0)')
+parser.add_argument('--epochs', default=50, type=int, metavar='N',
+                    help='number of total epochs to run')
+parser.add_argument('--task', choices=['train', 'test', 'channel', 'ER', 'GPCR', 'kinase'],
+                    default='', type=str, metavar='TASK',
+                    help='Task name. Could be train, test, channel, ER, GPCR, kinase.')
+parser.add_argument('--lr', '--learning-rate', default=1e-4, type=float,
+                    metavar='LR', help='initial learning rate', dest='lr')
+
+
+
+
+def get_task(task_name):
+    if task_name.lower() == 'train':
+        df_train = {"sps": './data/train_sps',
+                    "smile": './data/train_smile',
+                    "affinity": './data/train_ic50',
+                    }
+
+        tokenizer_config = {"vocab_file": './config/vocab.txt',
+                            "vocab_pair": './config/drug_codes_chembl.txt',
+                            "begin_id": '[CLS]',
+                            "separate_id": "[SEP]",
+                            "max_len": 256
+                            }
+        params = {'batch_size': 5,
+                  'shuffle': True,
+                  'num_workers': 0,
+                  'drop_last': True}
+        return df_train, tokenizer_config, params
+    elif task_name.lower() == 'test':
+        df_test = {"sps": './data/test_sps',
+                    "smile": './data/test_smile',
+                    "affinity": './data/test_ic50',
+                    }
+        tokenizer_config = {"vocab_file": './config/vocab.txt',
+                            "vocab_pair": './config/drug_codes_chembl.txt',
+                            "begin_id": '[CLS]',
+                            "separate_id": "[SEP]",
+                            "max_len": 256
+                            }
+        params = {'batch_size': 5,
+                  'shuffle': True,
+                  'num_workers': 0,
+                  'drop_last': True}
+
+        return df_test, tokenizer_config, params
+
+def main():
+    args = parser.parse_args()
+
+
+
+    dataset, config, params = get_task("train")
+    model = BertModel.from_pretrained("bert-base-uncased")
+
+if __name__ == '__main__':
+    # main()
+    tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
+    # model = BertModel('./config/config.json')
+    model = BertModel.from_pretrained("bert-base-uncased")
+    inputs = tokenizer("Hello world!", return_tensors="pt")
+    outputs = model(**inputs)

utils/hf_api.py

+# coding=utf-8
+# Copyright 2019-present, the HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import io
+import os
+from os.path import expanduser
+from typing import Dict, List, Optional, Tuple
+
+from tqdm import tqdm
+
+import requests
+
+
+ENDPOINT = "https://huggingface.co"
+
+
+class RepoObj:
+    """
+    HuggingFace git-based system, data structure that represents a file belonging to the current user.
+    """
+
+    def __init__(self, filename: str, lastModified: str, commit: str, size: int, **kwargs):
+        self.filename = filename
+        self.lastModified = lastModified
+        self.commit = commit
+        self.size = size
+
+
+class ModelSibling:
+    """
+    Data structure that represents a public file inside a model, accessible from huggingface.co
+    """
+
+    def __init__(self, rfilename: str, **kwargs):
+        self.rfilename = rfilename  # filename relative to the model root
+        for k, v in kwargs.items():
+            setattr(self, k, v)
+
+
+class ModelInfo:
+    """
+    Info about a public model accessible from huggingface.co
+    """
+
+    def __init__(
+        self,
+        modelId: Optional[str] = None,  # id of model
+        tags: List[str] = [],
+        pipeline_tag: Optional[str] = None,
+        siblings: Optional[List[Dict]] = None,  # list of files that constitute the model
+        **kwargs
+    ):
+        self.modelId = modelId
+        self.tags = tags
+        self.pipeline_tag = pipeline_tag
+        self.siblings = [ModelSibling(**x) for x in siblings] if siblings is not None else None
+        for k, v in kwargs.items():
+            setattr(self, k, v)
+
+
+class HfApi:
+    def __init__(self, endpoint=None):
+        self.endpoint = endpoint if endpoint is not None else ENDPOINT
+
+    def login(self, username: str, password: str) -> str:
+        """
+        Call HF API to sign in a user and get a token if credentials are valid.
+
+        Outputs: token if credentials are valid
+
+        Throws: requests.exceptions.HTTPError if credentials are invalid
+        """
+        path = "{}/api/login".format(self.endpoint)
+        r = requests.post(path, json={"username": username, "password": password})
+        r.raise_for_status()
+        d = r.json()
+        return d["token"]
+
+    def whoami(self, token: str) -> Tuple[str, List[str]]:
+        """
+        Call HF API to know "whoami"
+        """
+        path = "{}/api/whoami".format(self.endpoint)
+        r = requests.get(path, headers={"authorization": "Bearer {}".format(token)})
+        r.raise_for_status()
+        d = r.json()
+        return d["user"], d["orgs"]
+
+    def logout(self, token: str) -> None:
+        """
+        Call HF API to log out.
+        """
+        path = "{}/api/logout".format(self.endpoint)
+        r = requests.post(path, headers={"authorization": "Bearer {}".format(token)})
+        r.raise_for_status()
+
+    def model_list(self) -> List[ModelInfo]:
+        """
+        Get the public list of all the models on huggingface.co
+        """
+        path = "{}/api/models".format(self.endpoint)
+        r = requests.get(path)
+        r.raise_for_status()
+        d = r.json()
+        return [ModelInfo(**x) for x in d]
+
+    def list_repos_objs(self, token: str, organization: Optional[str] = None) -> List[RepoObj]:
+        """
+        HuggingFace git-based system, used for models.
+
+        Call HF API to list all stored files for user (or one of their organizations).
+        """
+        path = "{}/api/repos/ls".format(self.endpoint)
+        params = {"organization": organization} if organization is not None else None
+        r = requests.get(path, params=params, headers={"authorization": "Bearer {}".format(token)})
+        r.raise_for_status()
+        d = r.json()
+        return [RepoObj(**x) for x in d]
+
+    def create_repo(
+        self,
+        token: str,
+        name: str,
+        organization: Optional[str] = None,
+        private: Optional[bool] = None,
+        exist_ok=False,
+        lfsmultipartthresh: Optional[int] = None,
+    ) -> str:
+        """
+        HuggingFace git-based system, used for models.
+
+        Call HF API to create a whole repo.
+
+        Params:
+            private: Whether the model repo should be private (requires a paid huggingface.co account)
+
+            exist_ok: Do not raise an error if repo already exists
+
+            lfsmultipartthresh: Optional: internal param for testing purposes.
+        """
+        path = "{}/api/repos/create".format(self.endpoint)
+        json = {"name": name, "organization": organization, "private": private}
+        if lfsmultipartthresh is not None:
+            json["lfsmultipartthresh"] = lfsmultipartthresh
+        r = requests.post(
+            path,
+            headers={"authorization": "Bearer {}".format(token)},
+            json=json,
+        )
+        if exist_ok and r.status_code == 409:
+            return ""
+        r.raise_for_status()
+        d = r.json()
+        return d["url"]
+
+    def delete_repo(self, token: str, name: str, organization: Optional[str] = None):
+        """
+        HuggingFace git-based system, used for models.
+
+        Call HF API to delete a whole repo.
+
+        CAUTION(this is irreversible).
+        """
+        path = "{}/api/repos/delete".format(self.endpoint)
+        r = requests.delete(
+            path,
+            headers={"authorization": "Bearer {}".format(token)},
+            json={"name": name, "organization": organization},
+        )
+        r.raise_for_status()
+
+
+class TqdmProgressFileReader:
+    """
+    Wrap an io.BufferedReader `f` (such as the output of `open(…, "rb")`) and override `f.read()` so as to display a
+    tqdm progress bar.
+
+    see github.com/huggingface/transformers/pull/2078#discussion_r354739608 for implementation details.
+    """
+
+    def __init__(self, f: io.BufferedReader):
+        self.f = f
+        self.total_size = os.fstat(f.fileno()).st_size
+        self.pbar = tqdm(total=self.total_size, leave=False)
+        self.read = f.read
+        f.read = self._read
+
+    def _read(self, n=-1):
+        self.pbar.update(n)
+        return self.read(n)
+
+    def close(self):
+        self.pbar.close()
+
+
+class HfFolder:
+    path_token = expanduser("~/.huggingface/token")
+
+    @classmethod
+    def save_token(cls, token):
+        """
+        Save token, creating folder as needed.
+        """
+        os.makedirs(os.path.dirname(cls.path_token), exist_ok=True)
+        with open(cls.path_token, "w+") as f:
+            f.write(token)
+
+    @classmethod
+    def get_token(cls):
+        """
+        Get token or None if not existent.
+        """
+        try:
+            with open(cls.path_token, "r") as f:
+                return f.read()
+        except FileNotFoundError:
+            pass
+
+    @classmethod
+    def delete_token(cls):
+        """
+        Delete token. Do not fail if token does not exist.
+        """
+        try:
+            os.remove(cls.path_token)
+        except FileNotFoundError:
+            pass

utils/logging.py

+# coding=utf-8
+# Copyright 2020 Optuna, Hugging Face
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Logging utilities. """
+
+import logging
+import os
+import sys
+import threading
+from logging import CRITICAL  # NOQA
+from logging import DEBUG  # NOQA
+from logging import ERROR  # NOQA
+from logging import FATAL  # NOQA
+from logging import INFO  # NOQA
+from logging import NOTSET  # NOQA
+from logging import WARN  # NOQA
+from logging import WARNING  # NOQA
+from typing import Optional
+
+
+_lock = threading.Lock()
+_default_handler: Optional[logging.Handler] = None
+
+log_levels = {
+    "debug": logging.DEBUG,
+    "info": logging.INFO,
+    "warning": logging.WARNING,
+    "error": logging.ERROR,
+    "critical": logging.CRITICAL,
+}
+
+_default_log_level = logging.WARNING
+
+
+def _get_default_logging_level():
+    """
+    If TRANSFORMERS_VERBOSITY env var is set to one of the valid choices return that as the new default level. If it is
+    not - fall back to ``_default_log_level``
+    """
+    env_level_str = os.getenv("TRANSFORMERS_VERBOSITY", None)
+    if env_level_str:
+        if env_level_str in log_levels:
+            return log_levels[env_level_str]
+        else:
+            logging.getLogger().warning(
+                f"Unknown option TRANSFORMERS_VERBOSITY={env_level_str}, "
+                f"has to be one of: { ', '.join(log_levels.keys()) }"
+            )
+    return _default_log_level
+
+
+def _get_library_name() -> str:
+
+    return __name__.split(".")[0]
+
+
+def _get_library_root_logger() -> logging.Logger:
+
+    return logging.getLogger(_get_library_name())
+
+
+def _configure_library_root_logger() -> None:
+
+    global _default_handler
+
+    with _lock:
+        if _default_handler:
+            # This library has already configured the library root logger.
+            return
+        _default_handler = logging.StreamHandler()  # Set sys.stderr as stream.
+        _default_handler.flush = sys.stderr.flush
+
+        # Apply our default configuration to the library root logger.
+        library_root_logger = _get_library_root_logger()
+        library_root_logger.addHandler(_default_handler)
+        library_root_logger.setLevel(_get_default_logging_level())
+        library_root_logger.propagate = False
+
+
+def _reset_library_root_logger() -> None:
+
+    global _default_handler
+
+    with _lock:
+        if not _default_handler:
+            return
+
+        library_root_logger = _get_library_root_logger()
+        library_root_logger.removeHandler(_default_handler)
+        library_root_logger.setLevel(logging.NOTSET)
+        _default_handler = None
+
+
+def get_logger(name: Optional[str] = None) -> logging.Logger:
+    """
+    Return a logger with the specified name.
+
+    This function is not supposed to be directly accessed unless you are writing a custom transformers module.
+    """
+
+    if name is None:
+        name = _get_library_name()
+
+    _configure_library_root_logger()
+    return logging.getLogger(name)
+
+
+def get_verbosity() -> int:
+    """
+    Return the current level for the 🤗 Transformers's root logger as an int.
+
+    Returns:
+        :obj:`int`: The logging level.
+
+    .. note::
+
+        🤗 Transformers has following logging levels:
+
+        - 50: ``transformers.logging.CRITICAL`` or ``transformers.logging.FATAL``
+        - 40: ``transformers.logging.ERROR``
+        - 30: ``transformers.logging.WARNING`` or ``transformers.logging.WARN``
+        - 20: ``transformers.logging.INFO``
+        - 10: ``transformers.logging.DEBUG``
+    """
+
+    _configure_library_root_logger()
+    return _get_library_root_logger().getEffectiveLevel()
+
+
+def set_verbosity(verbosity: int) -> None:
+    """
+    Set the vebosity level for the 🤗 Transformers's root logger.
+
+    Args:
+        verbosity (:obj:`int`):
+            Logging level, e.g., one of:
+
+            - ``transformers.logging.CRITICAL`` or ``transformers.logging.FATAL``
+            - ``transformers.logging.ERROR``
+            - ``transformers.logging.WARNING`` or ``transformers.logging.WARN``
+            - ``transformers.logging.INFO``
+            - ``transformers.logging.DEBUG``
+    """
+
+    _configure_library_root_logger()
+    _get_library_root_logger().setLevel(verbosity)
+
+
+def set_verbosity_info():
+    """Set the verbosity to the :obj:`INFO` level."""
+    return set_verbosity(INFO)
+
+
+def set_verbosity_warning():
+    """Set the verbosity to the :obj:`WARNING` level."""
+    return set_verbosity(WARNING)
+
+
+def set_verbosity_debug():
+    """Set the verbosity to the :obj:`DEBUG` level."""
+    return set_verbosity(DEBUG)
+
+
+def set_verbosity_error():
+    """Set the verbosity to the :obj:`ERROR` level."""
+    return set_verbosity(ERROR)
+
+
+def disable_default_handler() -> None:
+    """Disable the default handler of the HuggingFace Transformers's root logger."""
+
+    _configure_library_root_logger()
+
+    assert _default_handler is not None
+    _get_library_root_logger().removeHandler(_default_handler)
+
+
+def enable_default_handler() -> None:
+    """Enable the default handler of the HuggingFace Transformers's root logger."""
+
+    _configure_library_root_logger()
+
+    assert _default_handler is not None
+    _get_library_root_logger().addHandler(_default_handler)
+
+
+def add_handler(handler: logging.Handler) -> None:
+    """adds a handler to the HuggingFace Transformers's root logger."""
+
+    _configure_library_root_logger()
+
+    assert handler is not None
+    _get_library_root_logger().addHandler(handler)
+
+
+def remove_handler(handler: logging.Handler) -> None:
+    """removes given handler from the HuggingFace Transformers's root logger."""
+
+    _configure_library_root_logger()
+
+    assert handler is not None and handler not in _get_library_root_logger().handlers
+    _get_library_root_logger().removeHandler(handler)
+
+
+def disable_propagation() -> None:
+    """
+    Disable propagation of the library log outputs. Note that log propagation is disabled by default.
+    """
+
+    _configure_library_root_logger()
+    _get_library_root_logger().propagate = False
+
+
+def enable_propagation() -> None:
+    """
+    Enable propagation of the library log outputs. Please disable the HuggingFace Transformers's default handler to
+    prevent double logging if the root logger has been configured.
+    """
+
+    _configure_library_root_logger()
+    _get_library_root_logger().propagate = True
+
+
+def enable_explicit_format() -> None:
+    """
+    Enable explicit formatting for every HuggingFace Transformers's logger. The explicit formatter is as follows:
+
+    ::
+
+        [LEVELNAME|FILENAME|LINE NUMBER] TIME >> MESSAGE
+
+    All handlers currently bound to the root logger are affected by this method.
+    """
+    handlers = _get_library_root_logger().handlers
+
+    for handler in handlers:
+        formatter = logging.Formatter("[%(levelname)s|%(filename)s:%(lineno)s] %(asctime)s >> %(message)s")
+        handler.setFormatter(formatter)
+
+
+def reset_format() -> None:
+    """
+    Resets the formatting for HuggingFace Transformers's loggers.
+
+    All handlers currently bound to the root logger are affected by this method.
+    """
+    handlers = _get_library_root_logger().handlers
+
+    for handler in handlers:
+        handler.setFormatter(None)