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.
"""PyTorch BERT model."""

import logging
import os
import typing

import torch
from torch import 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,
    cached_path,
    hf_bucket_url,
    is_remote_url,
)


logger = logging.getLogger(__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


class ModuleUtilsMixin:
    """
    A few utilities for torch.nn.Modules, to be used as a mixin.
    """

    def num_parameters(self, only_trainable: bool = False) -> int:
        """
        Get number of (optionally, trainable) parameters in the module.
        """
        params = filter(lambda x: x.requires_grad, self.parameters()) if only_trainable else self.parameters()
        return sum(p.numel() for p in params)

    @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 `mem_rss_diff` attribute for each module and can be reset to zero with `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):
        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):
        return next(self.parameters()).device


class PreTrainedModel(nn.Module, ModuleUtilsMixin):
    r""" Base class for all models.

        :class:`~transformers.PreTrainedModel` takes care of storing the configuration of the models and handles methods for loading/downloading/saving models
        as well as a few methods common to all models to (i) resize the input embeddings and (ii) prune heads in the self-attention heads.

        Class attributes (overridden by derived classes):
            - ``config_class``: a class derived from :class:`~transformers.PretrainedConfig` to use as configuration class for this model architecture.
            - ``pretrained_model_archive_map``: a python ``dict`` of with `short-cut-names` (string) as keys and `url` (string) of associated pretrained weights as values.
            - ``load_tf_weights``: a python ``method`` for loading a TensorFlow checkpoint in a PyTorch model, taking as arguments:

                - ``model``: an instance of the relevant subclass of :class:`~transformers.PreTrainedModel`,
                - ``config``: an instance of the relevant subclass of :class:`~transformers.PretrainedConfig`,
                - ``path``: a path (string) to the TensorFlow checkpoint.

            - ``base_model_prefix``: 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.
    """
    config_class = None
    pretrained_model_archive_map = {}
    base_model_prefix = ""

    @property
    def dummy_inputs(self):
        """ Dummy inputs to do a forward pass in the network.

        Returns:
            torch.Tensor with dummy inputs
        """
        return {"input_ids": torch.tensor(DUMMY_INPUTS)}

    def __init__(self, config, *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 in model
        self.config = config

    @property
    def base_model(self):
        return getattr(self, self.base_model_prefix, self)

    def get_input_embeddings(self):
        """
        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):
        """
        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):
        """
        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 `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:
            self._tie_or_clone_weights(output_embeddings, self.get_input_embeddings())

    def _tie_or_clone_weights(self, output_embeddings, input_embeddings):
        """ Tie or clone module weights depending of weither 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=None):
        """ Resize input token embeddings matrix of the model if new_num_tokens != config.vocab_size.
        Take care of tying weights embeddings afterwards if the model class has a `tie_weights()` method.

        Arguments:

            new_num_tokens: (`optional`) int:
                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 None: does nothing and just returns a pointer to the input tokens ``torch.nn.Embeddings`` Module of the model.

        Return: ``torch.nn.Embeddings``
            Pointer to the input tokens Embeddings Module of the model
        """
        base_model = getattr(self, self.base_model_prefix, self)  # get the base model if needed
        model_embeds = base_model._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
        base_model.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)
        return self.get_input_embeddings()

    def _get_resized_embeddings(self, old_embeddings, new_num_tokens=None):
        """ 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:
            new_num_tokens: (`optional`) int
                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 None: return the provided token Embedding Module.
        Return: ``torch.nn.Embeddings``
            Pointer to the resized Embedding Module or the old Embedding Module if new_num_tokens is 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

        # Build new embeddings
        new_embeddings = nn.Embedding(new_num_tokens, old_embedding_dim)
        new_embeddings.to(old_embeddings.weight.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 init_weights(self):
        """ Initialize 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):
        """ Prunes heads of the base model.

            Arguments:

                heads_to_prune: dict with keys being selected layer indices (`int`) and associated values being the list of heads to prune in said layer (list of `int`).
                E.g. {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):
        """ 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.
        """
        assert os.path.isdir(
            save_directory
        ), "Saving path should be a directory where the model and configuration can be saved"

        # Only save the model itself if we are using distributed training
        model_to_save = self.module if hasattr(self, "module") else self

        # Attach architecture to the config
        model_to_save.config.architectures = [model_to_save.__class__.__name__]

        # Save configuration file
        model_to_save.config.save_pretrained(save_directory)

        # If we save using the predefined names, we can load using `from_pretrained`
        output_model_file = os.path.join(save_directory, WEIGHTS_NAME)
        torch.save(model_to_save.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, *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 pre-trained 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: either:
              - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
              - a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``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 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.
              - None if you are both providing the configuration and state dictionary (resp. with keyword arguments ``config`` and ``state_dict``)

            model_args: (`optional`) Sequence of positional arguments:
                All remaning positional arguments will be passed to the underlying model's ``__init__`` method

            config: (`optional`) one of:
                - an instance of a class derived from :class:`~transformers.PretrainedConfig`, or
                - a string 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 ``shortcut-name`` string of a pretrained model), or
                    - the model was saved using :func:`~transformers.PreTrainedModel.save_pretrained` and is reloaded by suppling the save directory.
                    - the model is loaded by suppling a local directory as ``pretrained_model_name_or_path`` and a configuration JSON file named `config.json` is found in the directory.

            state_dict: (`optional`) dict:
                an optional state dictionnary for the model 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: (`optional`) string:
                Path to a directory in which a downloaded pre-trained model
                configuration should be cached if the standard cache should not be used.

            force_download: (`optional`) boolean, default False:
                Force to (re-)download the model weights and configuration files and override the cached versions if they exists.

            resume_download: (`optional`) boolean, default False:
                Do not delete incompletely recieved file. Attempt to resume the download if such a file exists.

            proxies: (`optional`) dict, default None:
                A dictionary of proxy servers to use by protocol or endpoint, e.g.: {'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}.
                The proxies are used on each request.

            output_loading_info: (`optional`) boolean:
                Set to ``True`` to also return a dictionnary containing missing keys, unexpected keys and error messages.

            kwargs: (`optional`) Remaining dictionary of keyword arguments:
                Can be used to update the configuration object (after it being loaded) and initiate the model. (e.g. ``output_attention=True``). Behave 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.

        Examples::

            # For example purposes. Not runnable.
            model = BertModel.from_pretrained('bert-base-uncased')    # Download model and configuration from S3 and cache.
            model = BertModel.from_pretrained('./test/saved_model/')  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
            model = BertModel.from_pretrained('bert-base-uncased', output_attention=True)  # Update configuration during loading
            assert model.config.output_attention == True
            # Loading from a TF checkpoint file instead of a PyTorch model (slower)
            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)

        # 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,
                **kwargs,
            )
        else:
            model_kwargs = kwargs

        # Load model
        if pretrained_model_name_or_path is not None:
            if pretrained_model_name_or_path in cls.pretrained_model_archive_map:
                archive_file = cls.pretrained_model_archive_map[pretrained_model_name_or_path]
            elif 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
                    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
                    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, postfix=(TF2_WEIGHTS_NAME if from_tf else WEIGHTS_NAME),
                )

            # redirect to the cache, if necessary
            try:
                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,
                )
            except EnvironmentError:
                if pretrained_model_name_or_path in cls.pretrained_model_archive_map:
                    msg = "Couldn't reach server at '{}' to download pretrained weights.".format(archive_file)
                else:
                    msg = (
                        "Model name '{}' was not found in model name list ({}). "
                        "We assumed '{}' was a path or url to model weight files named one of {} but "
                        "couldn't find any such file at this path or url.".format(
                            pretrained_model_name_or_path,
                            ", ".join(cls.pretrained_model_archive_map.keys()),
                            archive_file,
                            [WEIGHTS_NAME, TF2_WEIGHTS_NAME, TF_WEIGHTS_NAME],
                        )
                    )
                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

        # 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(
                    "Unable to load weights from pytorch checkpoint 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 transformers 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
            if not hasattr(model, cls.base_model_prefix) and any(
                s.startswith(cls.base_model_prefix) for s in state_dict.keys()
            ):
                start_prefix = cls.base_model_prefix + "."
            if hasattr(model, cls.base_model_prefix) and not any(
                s.startswith(cls.base_model_prefix) for s in state_dict.keys()
            ):
                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)

            if len(missing_keys) > 0:
                logger.info(
                    "Weights of {} not initialized from pretrained model: {}".format(
                        model.__class__.__name__, missing_keys
                    )
                )
            if len(unexpected_keys) > 0:
                logger.info(
                    "Weights from pretrained model not used in {}: {}".format(
                        model.__class__.__name__, unexpected_keys
                    )
                )
            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)
                    )
                )
        model.tie_weights()  # make sure token embedding weights are still tied if needed

        # Set model in evaluation mode to desactivate 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

    def prepare_inputs_for_generation(self, input_ids, **kwargs):
        return {"input_ids": input_ids}

    def prepare_scores_for_generation(self, scores, **kwargs):
        return scores

    def _do_output_past(self, outputs):
        """During generation, decide whether to pass the `past` variable to the next forward pass."""
        has_output_past = getattr(self.config, "output_past", False)
        mem_len = getattr(self.config, "mem_len", 0)
        if len(outputs) <= 1:
            return False
        if mem_len > 0 or has_output_past:
            return True
        return False

    def enforce_repetition_penalty_(self, lprobs, batch_size, num_beams, prev_output_tokens, repetition_penalty):
        """repetition penalty (from CTRL paper https://arxiv.org/abs/1909.05858). """
        for i in range(batch_size * num_beams):
            for previous_token in set(prev_output_tokens[i].tolist()):
                # if score < 0 then repetition penalty has to multiplied to reduce the previous token probability
                if lprobs[i, previous_token] < 0:
                    lprobs[i, previous_token] *= repetition_penalty
                else:
                    lprobs[i, previous_token] /= repetition_penalty

    @torch.no_grad()
    def generate(
        self,
        input_ids=None,
        max_length=None,
        min_length=None,
        do_sample=None,
        early_stopping=None,
        num_beams=None,
        temperature=None,
        top_k=None,
        top_p=None,
        repetition_penalty=None,
        bos_token_id=None,
        pad_token_id=None,
        eos_token_id=None,
        length_penalty=None,
        no_repeat_ngram_size=None,
        num_return_sequences=None,
        attention_mask=None,
        decoder_start_token_id=None,
    ):
        r""" Generates sequences for models with a LM head. The method currently supports greedy decoding, beam-search decoding, sampling with temperature, sampling with top-k or nucleus sampling.

        Adapted in part from `Facebook's XLM beam search code`_.

        .. _`Facebook's XLM beam search code`:
           https://github.com/facebookresearch/XLM/blob/9e6f6814d17be4fe5b15f2e6c43eb2b2d76daeb4/src/model/transformer.py#L529


        Parameters:

            input_ids: (`optional`) `torch.LongTensor` of shape `(batch_size, sequence_length)`
                The sequence used as a prompt for the generation. If `None` the method initializes
                it as an empty `torch.LongTensor` of shape `(1,)`.

            max_length: (`optional`) int
                The max length of the sequence to be generated.  Between `min_length` and infinity. Default to 20.

            min_length: (`optional`) int
                The min length of the sequence to be generated.  Between 0 and infinity. Default to 0.

            do_sample: (`optional`) bool
                If set to `False` greedy decoding is used. Otherwise sampling is used. Defaults to `False` as defined in `configuration_utils.PretrainedConfig`.

            early_stopping: (`optional`) bool
                if set to `True` beam search is stopped when at least `num_beams` sentences finished per batch. Defaults to `False` as defined in `configuration_utils.PretrainedConfig`.

            num_beams: (`optional`) int
                Number of beams for beam search. Must be between 1 and infinity. 1 means no beam search. Default to 1.

            temperature: (`optional`) float
                The value used to module the next token probabilities. Must be strictly positive. Default to 1.0.

            top_k: (`optional`) int
                The number of highest probability vocabulary tokens to keep for top-k-filtering. Between 1 and infinity. Default to 50.

            top_p: (`optional`) float
                The cumulative probability of parameter highest probability vocabulary tokens to keep for nucleus sampling. Must be between 0 and 1. Default to 1.

            repetition_penalty: (`optional`) float
                The parameter for repetition penalty. Between 1.0 and infinity. 1.0 means no penalty. Default to 1.0.

            pad_token_id: (`optional`) int
                Padding token. Default to specicic model pad_token_id or None if it does not exist.

            bos_token_id: (`optional`) int
                BOS token. Defaults to bos_token_id as defined in the models config.

            pad_token_id: (`optional`) int
                Pad token. Defaults to pad_token_id as defined in the models config.

            eos_token_ids: (`optional`) int or list of int
                End of sequence token or list of tokens to stop the generation. Default to eos_token_ids as defined in the models config.

            length_penalty: (`optional`) float
                Exponential penalty to the length. Default to 1.

            no_repeat_ngram_size: (`optional`) int
                If set to int > 0, all ngrams of size `no_repeat_ngram_size` can only occur once.

            num_return_sequences: (`optional`) int
                The number of independently computed returned sequences for each element in the batch. Default to 1.

            attention_mask (`optional`) obj: `torch.LongTensor` of same shape as `input_ids`
                Mask to avoid performing attention on padding token indices.
                Mask values selected in ``[0, 1]``:
                ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
                Defaults to `None`.

            `What are attention masks? <../glossary.html#attention-mask>`__

            decoder_start_token_id=None: (`optional`) int
                If an encoder-decoder model starts decoding with a different token than BOS.
                Defaults to `None` and is changed to `BOS` later.

        Return:

            output: `torch.LongTensor` of shape `(batch_size * num_return_sequences, sequence_length)`
                sequence_length is either equal to max_length or shorter if all batches finished early due to the `eos_token_id`

        Examples::

            tokenizer = AutoTokenizer.from_pretrained('distilgpt2')   # Initialize tokenizer
            model = AutoModelWithLMHead.from_pretrained('distilgpt2')    # Download model and configuration from S3 and cache.
            outputs = model.generate(max_length=40)  # do greedy decoding
            print('Generated: {}'.format(tokenizer.decode(outputs[0], skip_special_tokens=True)))

            tokenizer = AutoTokenizer.from_pretrained('openai-gpt')   # Initialize tokenizer
            model = AutoModelWithLMHead.from_pretrained('openai-gpt')    # Download model and configuration from S3 and cache.
            input_context = 'The dog'
            input_ids = tokenizer.encode(input_context, return_tensors='pt')  # encode input context
            outputs = model.generate(input_ids=input_ids, num_beams=5, num_return_sequences=3, temperature=1.5)  # generate 3 independent sequences using beam search decoding (5 beams) with sampling from initial context 'The dog'
            for i in range(3): #  3 output sequences were generated
                print('Generated {}: {}'.format(i, tokenizer.decode(outputs[i], skip_special_tokens=True)))

            tokenizer = AutoTokenizer.from_pretrained('distilgpt2')   # Initialize tokenizer
            model = AutoModelWithLMHead.from_pretrained('distilgpt2')    # Download model and configuration from S3 and cache.
            input_context = 'The dog'
            input_ids = tokenizer.encode(input_context, return_tensors='pt')  # encode input context
            outputs = model.generate(input_ids=input_ids, max_length=40, temperature=0.7, num_return_sequences=3)  # 3 generate sequences using by sampling
            for i in range(3): #  3 output sequences were generated
                print('Generated {}: {}'.format(i, tokenizer.decode(outputs[i], skip_special_tokens=True)))

            tokenizer = AutoTokenizer.from_pretrained('ctrl')   # Initialize tokenizer
            model = AutoModelWithLMHead.from_pretrained('ctrl')    # Download model and configuration from S3 and cache.
            input_context = 'Legal My neighbor is'  # "Legal" is one of the control codes for ctrl
            input_ids = tokenizer.encode(input_context, return_tensors='pt')  # encode input context
            outputs = model.generate(input_ids=input_ids, max_length=50, temperature=0.7, repetition_penalty=1.2)  # generate sequences
            print('Generated: {}'.format(tokenizer.decode(outputs[0], skip_special_tokens=True)))

        """

        # We cannot generate if the model does not have a LM head
        if self.get_output_embeddings() is None:
            raise AttributeError(
                "You tried to generate sequences with a model that does not have a LM Head."
                "Please use another model class (e.g. `OpenAIGPTLMHeadModel`, `XLNetLMHeadModel`, `GPT2LMHeadModel`, `CTRLLMHeadModel`, `T5WithLMHeadModel`, `TransfoXLLMHeadModel`, `XLMWithLMHeadModel`, `BartForConditionalGeneration` )"
            )

        max_length = max_length if max_length is not None else self.config.max_length
        min_length = min_length if min_length is not None else self.config.min_length
        do_sample = do_sample if do_sample is not None else self.config.do_sample
        early_stopping = early_stopping if early_stopping is not None else self.config.early_stopping
        num_beams = num_beams if num_beams is not None else self.config.num_beams
        temperature = temperature if temperature is not None else self.config.temperature
        top_k = top_k if top_k is not None else self.config.top_k
        top_p = top_p if top_p is not None else self.config.top_p
        repetition_penalty = repetition_penalty if repetition_penalty is not None else self.config.repetition_penalty
        bos_token_id = bos_token_id if bos_token_id is not None else self.config.bos_token_id
        pad_token_id = pad_token_id if pad_token_id is not None else self.config.pad_token_id
        eos_token_id = eos_token_id if eos_token_id is not None else self.config.eos_token_id
        length_penalty = length_penalty if length_penalty is not None else self.config.length_penalty
        no_repeat_ngram_size = (
            no_repeat_ngram_size if no_repeat_ngram_size is not None else self.config.no_repeat_ngram_size
        )
        num_return_sequences = (
            num_return_sequences if num_return_sequences is not None else self.config.num_return_sequences
        )
        decoder_start_token_id = (
            decoder_start_token_id if decoder_start_token_id is not None else self.config.decoder_start_token_id
        )

        if input_ids is not None:
            batch_size = input_ids.shape[0]  # overriden by the input batch_size
        else:
            batch_size = 1

        assert isinstance(max_length, int) and max_length > 0, "`max_length` should be a strictly positive integer."
        assert isinstance(min_length, int) and min_length >= 0, "`min_length` should be a positive integer."
        assert isinstance(do_sample, bool), "`do_sample` should be a boolean."
        assert isinstance(early_stopping, bool), "`early_stopping` should be a boolean."
        assert isinstance(num_beams, int) and num_beams > 0, "`num_beams` should be a strictly positive integer."
        assert temperature > 0, "`temperature` should be strictly positive."
        assert isinstance(top_k, int) and top_k >= 0, "`top_k` should be a positive integer."
        assert 0 <= top_p <= 1, "`top_p` should be between 0 and 1."
        assert repetition_penalty >= 1.0, "`repetition_penalty` should be >= 1."
        assert input_ids is not None or (
            isinstance(bos_token_id, int) and bos_token_id >= 0
        ), "If input_ids is not defined, `bos_token_id` should be a positive integer."
        assert pad_token_id is None or (
            isinstance(pad_token_id, int) and (pad_token_id >= 0)
        ), "`pad_token_id` should be a positive integer."
        assert (eos_token_id is None) or (
            isinstance(eos_token_id, int) and (eos_token_id >= 0)
        ), "`eos_token_id` should be a positive integer."
        assert length_penalty > 0, "`length_penalty` should be strictly positive."
        assert (
            isinstance(no_repeat_ngram_size, int) and no_repeat_ngram_size >= 0
        ), "`no_repeat_ngram_size` should be a positive integer."
        assert (
            isinstance(num_return_sequences, int) and num_return_sequences > 0
        ), "`num_return_sequences` should be a strictly positive integer."

        if input_ids is None:
            assert isinstance(bos_token_id, int) and bos_token_id >= 0, (
                "you should either supply a context to complete as `input_ids` input "
                "or a `bos_token_id` (integer >= 0) as a first token to start the generation."
            )
            input_ids = torch.full(
                (batch_size, 1), bos_token_id, dtype=torch.long, device=next(self.parameters()).device,
            )
        else:
            assert input_ids.dim() == 2, "Input prompt should be of shape (batch_size, sequence length)."

        # not allow to duplicate outputs when greedy decoding
        if do_sample is False:
            if num_beams == 1:
                # no_beam_search greedy generation conditions
                assert (
                    num_return_sequences == 1
                ), "Greedy decoding will always produce the same output for num_beams == 1 and num_return_sequences > 1. Please set num_return_sequences = 1"

            else:
                # beam_search greedy generation conditions
                assert (
                    num_beams >= num_return_sequences
                ), "Greedy beam search decoding cannot return more sequences than it has beams. Please set num_beams >= num_return_sequences"

        # create attention mask if necessary
        # TODO (PVP): this should later be handled by the forward fn() in each model in the future see PR 3140
        if (attention_mask is None) and (pad_token_id is not None) and (pad_token_id in input_ids):
            attention_mask = input_ids.ne(pad_token_id).long()
        elif attention_mask is None:
            attention_mask = input_ids.new_ones(input_ids.shape)

        # set pad_token_id to eos_token_id if not set. Important that this is done after
        # attention_mask is created
        if pad_token_id is None and eos_token_id is not None:
            logger.warning(
                "Setting `pad_token_id` to {} (first `eos_token_id`) to generate sequence".format(eos_token_id)
            )
            pad_token_id = eos_token_id

        # current position and vocab size
        vocab_size = self.config.vocab_size

        # set effective batch size and effective batch multiplier according to do_sample
        if do_sample:
            effective_batch_size = batch_size * num_return_sequences
            effective_batch_mult = num_return_sequences
        else:
            effective_batch_size = batch_size
            effective_batch_mult = 1

        if self.config.is_encoder_decoder:
            if decoder_start_token_id is None:
                decoder_start_token_id = bos_token_id

            assert (
                decoder_start_token_id is not None
            ), "decoder_start_token_id or bos_token_id has to be defined for encoder-decoder generation"
            assert hasattr(self, "get_encoder"), "{} should have a 'get_encoder' function defined".format(self)
            assert callable(self.get_encoder), "{} should be a method".format(self.get_encoder)

            # get encoder and store encoder outputs
            encoder = self.get_encoder()

            encoder_outputs = encoder(input_ids, attention_mask=attention_mask)

        # Expand input ids if num_beams > 1 or num_return_sequences > 1
        if num_return_sequences > 1 or num_beams > 1:
            input_ids_len = input_ids.shape[-1]
            input_ids = input_ids.unsqueeze(1).expand(batch_size, effective_batch_mult * num_beams, input_ids_len)
            attention_mask = attention_mask.unsqueeze(1).expand(
                batch_size, effective_batch_mult * num_beams, input_ids_len
            )

            input_ids = input_ids.contiguous().view(
                effective_batch_size * num_beams, input_ids_len
            )  # shape: (batch_size * num_return_sequences * num_beams, cur_len)
            attention_mask = attention_mask.contiguous().view(
                effective_batch_size * num_beams, input_ids_len
            )  # shape: (batch_size * num_return_sequences * num_beams, cur_len)

        if self.config.is_encoder_decoder:
            # create empty decoder_input_ids
            input_ids = torch.full(
                (effective_batch_size * num_beams, 1),
                decoder_start_token_id,
                dtype=torch.long,
                device=next(self.parameters()).device,
            )
            cur_len = 1
            batch_idx = self.encoder_outputs_batch_dim_idx
            assert (
                batch_size == encoder_outputs[0].shape[batch_idx]
            ), f"expected encoder_outputs[0] to have 1st dimension bs={batch_size}, got {encoder_outputs[0].shape[1]} "
            expanded_idx = (
                torch.arange(batch_size)
                .view(-1, 1)
                .repeat(1, num_beams * effective_batch_mult)
                .view(-1)
                .to(input_ids.device)
            )
            encoder_outputs = (encoder_outputs[0].index_select(batch_idx, expanded_idx), *encoder_outputs[1:])
        else:
            encoder_outputs = None
            cur_len = input_ids.shape[-1]

        if num_beams > 1:
            output = self._generate_beam_search(
                input_ids,
                cur_len=cur_len,
                max_length=max_length,
                min_length=min_length,
                do_sample=do_sample,
                early_stopping=early_stopping,
                temperature=temperature,
                top_k=top_k,
                top_p=top_p,
                repetition_penalty=repetition_penalty,
                no_repeat_ngram_size=no_repeat_ngram_size,
                bos_token_id=bos_token_id,
                pad_token_id=pad_token_id,
                decoder_start_token_id=decoder_start_token_id,
                eos_token_id=eos_token_id,
                batch_size=effective_batch_size,
                num_return_sequences=num_return_sequences,
                length_penalty=length_penalty,
                num_beams=num_beams,
                vocab_size=vocab_size,
                encoder_outputs=encoder_outputs,
                attention_mask=attention_mask,
            )
        else:
            output = self._generate_no_beam_search(
                input_ids,
                cur_len=cur_len,
                max_length=max_length,
                min_length=min_length,
                do_sample=do_sample,
                temperature=temperature,
                top_k=top_k,
                top_p=top_p,
                repetition_penalty=repetition_penalty,
                no_repeat_ngram_size=no_repeat_ngram_size,
                bos_token_id=bos_token_id,
                pad_token_id=pad_token_id,
                decoder_start_token_id=decoder_start_token_id,
                eos_token_id=eos_token_id,
                batch_size=effective_batch_size,
                encoder_outputs=encoder_outputs,
                attention_mask=attention_mask,
            )

        return output

    def _generate_no_beam_search(
        self,
        input_ids,
        cur_len,
        max_length,
        min_length,
        do_sample,
        temperature,
        top_k,
        top_p,
        repetition_penalty,
        no_repeat_ngram_size,
        bos_token_id,
        pad_token_id,
        eos_token_id,
        decoder_start_token_id,
        batch_size,
        encoder_outputs,
        attention_mask,
    ):
        """ Generate sequences for each example without beam search (num_beams == 1).
            All returned sequence are generated independantly.
        """
        # length of generated sentences / unfinished sentences
        unfinished_sents = input_ids.new(batch_size).fill_(1)
        sent_lengths = input_ids.new(batch_size).fill_(max_length)

        past = encoder_outputs  # defined for encoder-decoder models, None for decoder-only models

        while cur_len < max_length:
            model_inputs = self.prepare_inputs_for_generation(input_ids, past=past, attention_mask=attention_mask)

            outputs = self(**model_inputs)
            next_token_logits = outputs[0][:, -1, :]

            # if model has past, then set the past variable to speed up decoding
            if self._do_output_past(outputs):
                past = outputs[1]

            # repetition penalty from CTRL paper (https://arxiv.org/abs/1909.05858)
            if repetition_penalty != 1.0:
                self.enforce_repetition_penalty_(next_token_logits, batch_size, 1, input_ids, repetition_penalty)

            if no_repeat_ngram_size > 0:
                # calculate a list of banned tokens to prevent repetitively generating the same ngrams
                # from fairseq: https://github.com/pytorch/fairseq/blob/a07cb6f40480928c9e0548b737aadd36ee66ac76/fairseq/sequence_generator.py#L345
                banned_tokens = calc_banned_tokens(input_ids, batch_size, no_repeat_ngram_size, cur_len)
                for batch_idx in range(batch_size):
                    next_token_logits[batch_idx, banned_tokens[batch_idx]] = -float("inf")

            # set eos token prob to zero if min_length is not reached
            if eos_token_id is not None and cur_len < min_length:
                next_token_logits[:, eos_token_id] = -float("inf")

            if do_sample:
                # Temperature (higher temperature => more likely to sample low probability tokens)
                if temperature != 1.0:
                    next_token_logits = next_token_logits / temperature
                # Top-p/top-k filtering
                next_token_logits = top_k_top_p_filtering(next_token_logits, top_k=top_k, top_p=top_p)
                # Sample
                probs = F.softmax(next_token_logits, dim=-1)
                next_token = torch.multinomial(probs, num_samples=1).squeeze(1)
            else:
                # Greedy decoding
                next_token = torch.argmax(next_token_logits, dim=-1)

            # update generations and finished sentences
            if eos_token_id is not None:
                # pad finished sentences if eos_token_id exist
                tokens_to_add = next_token * unfinished_sents + (pad_token_id) * (1 - unfinished_sents)
            else:
                tokens_to_add = next_token

            input_ids = torch.cat([input_ids, tokens_to_add.unsqueeze(-1)], dim=-1)

            if eos_token_id is not None:
                eos_in_sents = tokens_to_add == eos_token_id
                # if sentence is unfinished and the token to add is eos, sent_lengths is filled with current length
                is_sents_unfinished_and_token_to_add_is_eos = unfinished_sents.mul(eos_in_sents.long()).bool()
                sent_lengths.masked_fill_(is_sents_unfinished_and_token_to_add_is_eos, cur_len + 1)
                # unfinished_sents is set to zero if eos in sentence
                unfinished_sents.mul_((~eos_in_sents).long())

            # stop when there is a </s> in each sentence, or if we exceed the maximul length
            if unfinished_sents.max() == 0:
                break

            # extend attention_mask for new generated input if only decoder
            if self.config.is_encoder_decoder is False:
                attention_mask = torch.cat(
                    [attention_mask, attention_mask.new_ones((attention_mask.shape[0], 1))], dim=-1
                )

            cur_len = cur_len + 1

        # if there are different sentences lengths in the batch, some batches have to be padded
        if sent_lengths.min().item() != sent_lengths.max().item():
            assert pad_token_id is not None, "`Pad_token_id` has to be defined if batches have different lengths"
            # finished sents are filled with pad_token
            decoded = input_ids.new(batch_size, sent_lengths.max().item()).fill_(pad_token_id)
        else:
            decoded = input_ids

        for hypo_idx, hypo in enumerate(input_ids):
            decoded[hypo_idx, : sent_lengths[hypo_idx]] = hypo[: sent_lengths[hypo_idx]]

        return decoded

    def _generate_beam_search(
        self,
        input_ids,
        cur_len,
        max_length,
        min_length,
        do_sample,
        early_stopping,
        temperature,
        top_k,
        top_p,
        repetition_penalty,
        no_repeat_ngram_size,
        bos_token_id,
        pad_token_id,
        eos_token_id,
        decoder_start_token_id,
        batch_size,
        num_return_sequences,
        length_penalty,
        num_beams,
        vocab_size,
        encoder_outputs,
        attention_mask,
    ):
        """ Generate sequences for each example with beam search.
        """

        # generated hypotheses
        generated_hyps = [
            BeamHypotheses(num_beams, max_length, length_penalty, early_stopping=early_stopping)
            for _ in range(batch_size)
        ]

        # scores for each sentence in the beam
        beam_scores = torch.zeros((batch_size, num_beams), dtype=torch.float, device=input_ids.device)

        # for greedy decoding it is made sure that only tokens of the first beam are considered to avoid sampling the exact same tokens three times
        if do_sample is False:
            beam_scores[:, 1:] = -1e9
        beam_scores = beam_scores.view(-1)  # shape (batch_size * num_beams,)

        # cache compute states
        past = encoder_outputs  # defined for encoder-decoder models, None for decoder-only models

        # done sentences
        done = [False for _ in range(batch_size)]

        while cur_len < max_length:
            model_inputs = self.prepare_inputs_for_generation(input_ids, past=past, attention_mask=attention_mask)
            outputs = self(**model_inputs)  # (batch_size * num_beams, cur_len, vocab_size)
            next_token_logits = outputs[0][:, -1, :]  # (batch_size * num_beams, vocab_size)

            # if model has past, then set the past variable to speed up decoding
            if self._do_output_past(outputs):
                past = outputs[1]

            # repetition penalty (from CTRL paper https://arxiv.org/abs/1909.05858)
            if repetition_penalty != 1.0:
                self.enforce_repetition_penalty_(
                    next_token_logits, batch_size, num_beams, input_ids, repetition_penalty,
                )

            if temperature != 1.0:
                next_token_logits = next_token_logits / temperature

            scores = F.log_softmax(next_token_logits, dim=-1)  # (batch_size * num_beams, vocab_size)
            if self.config.is_encoder_decoder and do_sample is False:
                # TODO (PVP) still a bit hacky here - there might be a better solutino
                scores = self.prepare_scores_for_generation(scores, cur_len=cur_len, max_length=max_length)

            # set eos token prob to zero if min_length is not reached
            if eos_token_id is not None and cur_len < min_length:
                scores[:, eos_token_id] = -float("inf")

            if no_repeat_ngram_size > 0:
                # calculate a list of banned tokens to prevent repetitively generating the same ngrams
                num_batch_hypotheses = batch_size * num_beams
                # from fairseq: https://github.com/pytorch/fairseq/blob/a07cb6f40480928c9e0548b737aadd36ee66ac76/fairseq/sequence_generator.py#L345
                banned_batch_tokens = calc_banned_tokens(
                    input_ids, num_batch_hypotheses, no_repeat_ngram_size, cur_len
                )
                for i, banned_tokens in enumerate(banned_batch_tokens):
                    scores[i, banned_tokens] = -float("inf")

            assert scores.shape == (batch_size * num_beams, vocab_size), "Shapes of scores: {} != {}".format(
                scores.shape, (batch_size * num_beams, vocab_size)
            )

            if do_sample:
                _scores = scores + beam_scores[:, None].expand_as(scores)  # (batch_size * num_beams, vocab_size)
                # Top-p/top-k filtering
                _scores = top_k_top_p_filtering(
                    _scores, top_k=top_k, top_p=top_p, min_tokens_to_keep=2
                )  # (batch_size * num_beams, vocab_size)
                # re-organize to group the beam together to sample from all beam_idxs
                _scores = _scores.contiguous().view(
                    batch_size, num_beams * vocab_size
                )  # (batch_size, num_beams * vocab_size)

                # Sample 2 next tokens for each beam (so we have some spare tokens and match output of greedy beam search)
                probs = F.softmax(_scores, dim=-1)
                next_tokens = torch.multinomial(probs, num_samples=2 * num_beams)  # (batch_size, num_beams * 2)
                # Compute next scores
                next_scores = torch.gather(_scores, -1, next_tokens)  # (batch_size, num_beams * 2)
                # sort the sampled vector to make sure that the first num_beams samples are the best
                next_scores, next_scores_indices = torch.sort(next_scores, descending=True, dim=1)
                next_tokens = torch.gather(next_tokens, -1, next_scores_indices)  # (batch_size, num_beams * 2)

            else:
                next_scores = scores + beam_scores[:, None].expand_as(scores)  # (batch_size * num_beams, vocab_size)

                # re-organize to group the beam together (we are keeping top hypothesis accross beams)
                next_scores = next_scores.view(
                    batch_size, num_beams * vocab_size
                )  # (batch_size, num_beams * vocab_size)

                next_scores, next_tokens = torch.topk(next_scores, 2 * num_beams, dim=1, largest=True, sorted=True)

            assert next_scores.size() == next_tokens.size() == (batch_size, 2 * num_beams)

            # next batch beam content
            next_batch_beam = []

            # for each sentence
            for batch_idx in range(batch_size):

                # if we are done with this sentence
                if done[batch_idx]:
                    assert (
                        len(generated_hyps[batch_idx]) >= num_beams
                    ), "Batch can only be done if at least {} beams have been generated".format(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"
                    next_batch_beam.extend([(0, pad_token_id, 0)] * num_beams)  # pad the batch
                    continue

                # next sentence beam content
                next_sent_beam = []

                # next tokens for this sentence
                for beam_token_rank, (beam_token_id, beam_token_score) in enumerate(
                    zip(next_tokens[batch_idx], next_scores[batch_idx])
                ):
                    # get beam and word IDs
                    beam_id = beam_token_id // vocab_size
                    token_id = beam_token_id % vocab_size

                    effective_beam_id = batch_idx * num_beams + beam_id

                    # add to generated hypotheses if end of sentence
                    if (eos_token_id is not None) and (token_id.item() is 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 >= num_beams
                        if is_beam_token_worse_than_top_num_beams:
                            continue
                        generated_hyps[batch_idx].add(
                            input_ids[effective_beam_id].clone(), beam_token_score.item(),
                        )
                    else:
                        # add next predicted word if it is not eos_token
                        next_sent_beam.append((beam_token_score, token_id, effective_beam_id))

                    # the beam for next step is full
                    if len(next_sent_beam) == num_beams:
                        break

                # Check if were done so that we can save a pad step if all(done)
                done[batch_idx] = done[batch_idx] or generated_hyps[batch_idx].is_done(
                    next_scores[batch_idx].max().item(), cur_len=cur_len
                )

                # update next beam content
                assert len(next_sent_beam) == num_beams, "Beam should always be full"
                next_batch_beam.extend(next_sent_beam)
                assert len(next_batch_beam) == num_beams * (batch_idx + 1)

            # stop when we are done with each sentence
            if all(done):
                break

            # sanity check / prepare next batch
            assert len(next_batch_beam) == batch_size * num_beams
            beam_scores = beam_scores.new([x[0] for x in next_batch_beam])
            beam_tokens = input_ids.new([x[1] for x in next_batch_beam])
            beam_idx = input_ids.new([x[2] for x in next_batch_beam])

            # re-order batch
            input_ids = input_ids[beam_idx, :]
            input_ids = torch.cat([input_ids, beam_tokens.unsqueeze(1)], dim=-1)

            # re-order internal states
            if past is not None:
                past = self._reorder_cache(past, beam_idx)

            # extend attention_mask for new generated input if only decoder
            if self.config.is_encoder_decoder is False:
                attention_mask = torch.cat(
                    [attention_mask, attention_mask.new_ones((attention_mask.shape[0], 1))], dim=-1
                )

            # update current length
            cur_len = cur_len + 1

        # finalize all open beam hypotheses and end to generated hypotheses
        for batch_idx in range(batch_size):
            if done[batch_idx]:
                continue

            # test that beam scores match previously calculated scores if not eos and batch_idx not done
            if eos_token_id is not None and all(
                (token_id % vocab_size).item() is not eos_token_id for token_id in next_tokens[batch_idx]
            ):
                assert torch.all(
                    next_scores[batch_idx, :num_beams] == beam_scores.view(batch_size, num_beams)[batch_idx]
                ), "If batch_idx is not done, final next scores: {} have to equal to accumulated beam_scores: {}".format(
                    next_scores[:, :num_beams][batch_idx], beam_scores.view(batch_size, num_beams)[batch_idx],
                )

            # need to add best num_beams hypotheses to generated hyps
            for beam_id in range(num_beams):
                effective_beam_id = batch_idx * num_beams + beam_id
                final_score = beam_scores[effective_beam_id].item()
                final_tokens = input_ids[effective_beam_id]
                generated_hyps[batch_idx].add(final_tokens, final_score)

        # depending on whether greedy generation is wanted or not define different output_batch_size and output_num_return_sequences_per_batch
        output_batch_size = batch_size if do_sample else batch_size * num_return_sequences
        output_num_return_sequences_per_batch = 1 if do_sample else num_return_sequences

        # select the best hypotheses
        sent_lengths = input_ids.new(output_batch_size)
        best = []

        # retrieve best hypotheses
        for i, hypotheses in enumerate(generated_hyps):
            sorted_hyps = sorted(hypotheses.beams, key=lambda x: x[0])
            for j in range(output_num_return_sequences_per_batch):
                effective_batch_idx = output_num_return_sequences_per_batch * i + j
                best_hyp = sorted_hyps.pop()[1]
                sent_lengths[effective_batch_idx] = len(best_hyp)
                best.append(best_hyp)

        # shorter batches are filled with pad_token
        if sent_lengths.min().item() != sent_lengths.max().item():
            assert pad_token_id is not None, "`Pad_token_id` has to be defined"
            sent_max_len = min(sent_lengths.max().item() + 1, max_length)
            decoded = input_ids.new(output_batch_size, sent_max_len).fill_(pad_token_id)

            # fill with hypothesis and eos_token_id if necessary
            for i, hypo in enumerate(best):
                decoded[i, : sent_lengths[i]] = hypo
                if sent_lengths[i] < max_length:
                    decoded[i, sent_lengths[i]] = eos_token_id
        else:
            # none of the hypotheses have an eos_token
            assert (len(hypo) == max_length for hypo in best)
            decoded = torch.stack(best).type(torch.long).to(next(self.parameters()).device)

        return decoded

    # force one of token_ids to be generated by setting prob of all other tokens to 0.
    def _force_token_ids_generation(self, scores, token_ids):
        if isinstance(token_ids, int):
            token_ids = [token_ids]
        all_but_token_ids_mask = torch.tensor(
            [x for x in range(self.config.vocab_size) if x not in token_ids],
            dtype=torch.long,
            device=next(self.parameters()).device,
        )
        assert len(scores.shape) == 2, "scores should be of rank 2 with shape: [batch_size, vocab_size]"
        scores[:, all_but_token_ids_mask] = -float("inf")

    @staticmethod
    def _reorder_cache(past, beam_idx):
        reordered_past = []
        for layer_past in past:
            # get the correct batch idx from layer past batch dim
            # batch dim of `past` and `mems` is at 2nd position
            reordered_layer_past = [layer_past[:, i].unsqueeze(1).clone().detach() for i in beam_idx]
            reordered_layer_past = torch.cat(reordered_layer_past, dim=1)
            # check that shape matches
            assert reordered_layer_past.shape == layer_past.shape
            reordered_past.append(reordered_layer_past)
        past = tuple(reordered_past)
        return past


def calc_banned_tokens(prev_input_ids, num_hypos, no_repeat_ngram_size, cur_len):
    # Copied from fairseq for no_repeat_ngram in beam_search"""
    if cur_len + 1 < no_repeat_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 = [{} 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(no_repeat_ngram_size)]):
            prev_ngram_tuple = tuple(ngram[:-1])
            generated_ngram[prev_ngram_tuple] = generated_ngram.get(prev_ngram_tuple, []) + [ngram[-1]]

    def _get_generated_ngrams(hypo_idx):
        # Before decoding the next token, prevent decoding of ngrams that have already appeared
        start_idx = cur_len + 1 - no_repeat_ngram_size
        ngram_idx = tuple(prev_input_ids[hypo_idx, start_idx:cur_len].tolist())
        return generated_ngrams[hypo_idx].get(ngram_idx, [])

    banned_tokens = [_get_generated_ngrams(hypo_idx) for hypo_idx in range(num_hypos)]
    return banned_tokens


def top_k_top_p_filtering(logits, top_k=0, top_p=1.0, filter_value=-float("Inf"), min_tokens_to_keep=1):
    """ Filter a distribution of logits using top-k and/or nucleus (top-p) filtering
        Args:
            logits: logits distribution shape (batch size, vocabulary size)
            if top_k > 0: keep only top k tokens with highest probability (top-k filtering).
            if top_p < 1.0: keep the top tokens with cumulative probability >= top_p (nucleus filtering).
                Nucleus filtering is described in Holtzman et al. (http://arxiv.org/abs/1904.09751)
            Make sure we keep at least min_tokens_to_keep per batch example in the output
        From: https://gist.github.com/thomwolf/1a5a29f6962089e871b94cbd09daf317
    """
    if top_k > 0:
        top_k = min(max(top_k, min_tokens_to_keep), logits.size(-1))  # Safety check
        # Remove all tokens with a probability less than the last token of the top-k
        indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
        logits[indices_to_remove] = filter_value

    if top_p < 1.0:
        sorted_logits, sorted_indices = torch.sort(logits, descending=True)
        cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)

        # Remove tokens with cumulative probability above the threshold (token with 0 are kept)
        sorted_indices_to_remove = cumulative_probs > top_p
        if 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[..., :min_tokens_to_keep] = 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)
        logits[indices_to_remove] = filter_value
    return logits


class BeamHypotheses(object):
    def __init__(self, num_beams, max_length, length_penalty, early_stopping):
        """
        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, sum_logprobs):
        """
        Add a new hypothesis to the list.
        """
        score = sum_logprobs / len(hyp) ** 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_scores = sorted([(s, idx) for idx, (s, _) in enumerate(self.beams)])
                del self.beams[sorted_scores[0][1]]
                self.worst_score = sorted_scores[1][0]
            else:
                self.worst_score = min(score, self.worst_score)

    def is_done(self, best_sum_logprobs, cur_len=None):
        """
        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:
            if cur_len is None:
                cur_len = self.max_length
            cur_score = best_sum_logprobs / cur_len ** self.length_penalty
            ret = self.worst_score >= cur_score
            return ret


class Conv1D(nn.Module):
    def __init__(self, nf, nx):
        """ Conv1D 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
        """
        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. """

    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.hidden_size, 1)

    def forward(self, hidden_states, p_mask=None):
        """ Args:
            **p_mask**: (`optional`) ``torch.FloatTensor`` of shape `(batch_size, seq_len)`
                invalid position mask such as query and special symbols (PAD, SEP, CLS)
                1.0 means token should be masked.
        """
        x = self.dense(hidden_states).squeeze(-1)

        if p_mask is not None:
            if next(self.parameters()).dtype == 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 and start token hidden state.
    """

    def __init__(self, config):
        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, start_states=None, start_positions=None, p_mask=None):
        """ Args:
            One of ``start_states``, ``start_positions`` should be not None.
            If both are set, ``start_positions`` overrides ``start_states``.

            **start_states**: ``torch.LongTensor`` of shape identical to hidden_states
                hidden states of the first tokens for the labeled span.
            **start_positions**: ``torch.LongTensor`` of shape ``(batch_size,)``
                position of the first token for the labeled span:
            **p_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, seq_len)``
                Mask of invalid position such as query and special symbols (PAD, SEP, CLS)
                1.0 means token should be masked.
        """
        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 next(self.parameters()).dtype == 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. """

    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, start_states=None, start_positions=None, cls_index=None):
        """
        Args:
            One of ``start_states``, ``start_positions`` should be not None.
            If both are set, ``start_positions`` overrides ``start_states``.

            **start_states**: ``torch.LongTensor`` of shape identical to ``hidden_states``.
                hidden states of the first tokens for the labeled span.
            **start_positions**: ``torch.LongTensor`` of shape ``(batch_size,)``
                position of the first token for the labeled span.
            **cls_index**: torch.LongTensor of shape ``(batch_size,)``
                position of the CLS token. If None, take the last token.

            note(Original repo):
                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


class SQuADHead(nn.Module):
    r""" A SQuAD head inspired by XLNet.

    Parameters:
        config (:class:`~transformers.XLNetConfig`): Model configuration class with all the parameters of the model.

    Inputs:
        **hidden_states**: ``torch.FloatTensor`` of shape ``(batch_size, seq_len, hidden_size)``
            hidden states of sequence tokens
        **start_positions**: ``torch.LongTensor`` of shape ``(batch_size,)``
            position of the first token for the labeled span.
        **end_positions**: ``torch.LongTensor`` of shape ``(batch_size,)``
            position of the last token for the labeled span.
        **cls_index**: torch.LongTensor of shape ``(batch_size,)``
            position of the CLS token. If None, take the last token.
        **is_impossible**: ``torch.LongTensor`` of shape ``(batch_size,)``
            Whether the question has a possible answer in the paragraph or not.
        **p_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, seq_len)``
            Mask of invalid position such as query and special symbols (PAD, SEP, CLS)
            1.0 means token should be masked.

    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
        **loss**: (`optional`, returned if both ``start_positions`` and ``end_positions`` are provided) ``torch.FloatTensor`` of shape ``(1,)``:
            Classification loss as the sum of start token, end token (and is_impossible if provided) classification losses.
        **start_top_log_probs**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
            ``torch.FloatTensor`` of shape ``(batch_size, config.start_n_top)``
            Log probabilities for the top config.start_n_top start token possibilities (beam-search).
        **start_top_index**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
            ``torch.LongTensor`` of shape ``(batch_size, config.start_n_top)``
            Indices for the top config.start_n_top start token possibilities (beam-search).
        **end_top_log_probs**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
            ``torch.FloatTensor`` of shape ``(batch_size, config.start_n_top * config.end_n_top)``
            Log probabilities for the top ``config.start_n_top * config.end_n_top`` end token possibilities (beam-search).
        **end_top_index**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
            ``torch.LongTensor`` of shape ``(batch_size, config.start_n_top * config.end_n_top)``
            Indices for the top ``config.start_n_top * config.end_n_top`` end token possibilities (beam-search).
        **cls_logits**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
            ``torch.FloatTensor`` of shape ``(batch_size,)``
            Log probabilities for the ``is_impossible`` label of the answers.
    """

    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)

    def forward(
        self, hidden_states, start_positions=None, end_positions=None, cls_index=None, is_impossible=None, p_mask=None,
    ):
        outputs = ()

        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

            outputs = (total_loss,) + outputs

        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)

            outputs = (start_top_log_probs, start_top_index, end_top_log_probs, end_top_index, cls_logits,) + outputs

        # return start_top_log_probs, start_top_index, end_top_log_probs, end_top_index, cls_logits
        # or (if labels are provided) (total_loss,)
        return outputs


class SequenceSummary(nn.Module):
    r""" Compute a single vector summary of a sequence hidden states according to various possibilities:
        Args of the config class:
            summary_type:
                - 'last' => [default] take the last token hidden state (like XLNet)
                - 'first' => take the first token hidden state (like Bert)
                - 'mean' => take the mean of all tokens hidden states
                - 'cls_index' => supply a Tensor of classification token position (GPT/GPT-2)
                - 'attn' => Not implemented now, use multi-head attention
            summary_use_proj: Add a projection after the vector extraction
            summary_proj_to_labels: If True, the projection outputs to config.num_labels classes (otherwise to hidden_size). Default: False.
            summary_activation: 'tanh' or another string => add an activation to the output, Other => no activation. Default
            summary_first_dropout: Add a dropout before the projection and activation
            summary_last_dropout: Add a dropout 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 = (
            get_activation(activation_string) if activation_string else Identity()
        )  # type: typing.Callable

        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, cls_index=None):
        """ hidden_states: float Tensor in shape [bsz, ..., seq_len, hidden_size], the hidden-states of the last layer.
            cls_index: [optional] position of the classification token if summary_type == 'cls_index',
                shape (bsz,) or more generally (bsz, ...) where ... are optional leading dimensions of hidden_states.
                if summary_type == 'cls_index' and cls_index is None:
                    we take the last token of the sequence as classification token
        """
        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 create_position_ids_from_input_ids(input_ids, padding_idx):
    """ Replace non-padding symbols with their position numbers. Position numbers begin at
    padding_idx+1. Padding symbols are ignored. This is modified from fairseq's
    `utils.make_positions`.

    :param torch.Tensor x:
    :return torch.Tensor:
    """
    # The series of casts and type-conversions here are carefully balanced to both work with ONNX export and XLA.
    mask = input_ids.ne(padding_idx).int()
    incremental_indicies = torch.cumsum(mask, dim=1).type_as(mask) * mask
    return incremental_indicies.long() + padding_idx


def prune_linear_layer(layer, index, dim=0):
    """ Prune a linear layer (a model parameters) to keep only entries in index.
        Return the pruned layer as a new layer with requires_grad=True.
        Used to remove heads.
    """
    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, index, dim=1):
    """ Prune a Conv1D layer (a model parameters) to keep only entries in index.
        A Conv1D work as a Linear layer (see e.g. BERT) but the weights are transposed.
        Return the pruned layer as a new layer with requires_grad=True.
        Used to remove heads.
    """
    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, index, dim=None):
    """ Prune a Conv1D or nn.Linear layer (a model parameters) to keep only entries in index.
        Return the pruned layer as a new layer with requires_grad=True.
        Used to remove heads.
    """
    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__))