modeling_gpt2.py

# coding=utf-8
# Copyright 2018 The OpenAI Team Authors and 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 OpenAI GPT-2 model."""


import logging
import os
import warnings

import torch
import torch.nn as nn
from torch.nn import CrossEntropyLoss

from .activations import ACT2FN
from .configuration_gpt2 import GPT2Config
from .file_utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_callable
from .modeling_utils import (
    Conv1D,
    PreTrainedModel,
    SequenceSummary,
    find_pruneable_heads_and_indices,
    prune_conv1d_layer,
)


logger = logging.getLogger(__name__)

_TOKENIZER_FOR_DOC = "GPT2Tokenizer"

GPT2_PRETRAINED_MODEL_ARCHIVE_LIST = [
    "gpt2",
    "gpt2-medium",
    "gpt2-large",
    "gpt2-xl",
    "distilgpt2",
    # See all GPT-2 models at https://huggingface.co/models?filter=gpt2
]


def load_tf_weights_in_gpt2(model, config, gpt2_checkpoint_path):
    """ Load tf checkpoints in a pytorch model
    """
    try:
        import re
        import tensorflow as tf
    except ImportError:
        logger.error(
            "Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
            "https://www.tensorflow.org/install/ for installation instructions."
        )
        raise
    tf_path = os.path.abspath(gpt2_checkpoint_path)
    logger.info("Converting TensorFlow checkpoint from {}".format(tf_path))
    # Load weights from TF model
    init_vars = tf.train.list_variables(tf_path)
    names = []
    arrays = []
    for name, shape in init_vars:
        logger.info("Loading TF weight {} with shape {}".format(name, shape))
        array = tf.train.load_variable(tf_path, name)
        names.append(name)
        arrays.append(array.squeeze())

    for name, array in zip(names, arrays):
        name = name[6:]  # skip "model/"
        name = name.split("/")
        pointer = model
        for m_name in name:
            if re.fullmatch(r"[A-Za-z]+\d+", m_name):
                scope_names = re.split(r"(\d+)", m_name)
            else:
                scope_names = [m_name]
            if scope_names[0] == "w" or scope_names[0] == "g":
                pointer = getattr(pointer, "weight")
            elif scope_names[0] == "b":
                pointer = getattr(pointer, "bias")
            elif scope_names[0] == "wpe" or scope_names[0] == "wte":
                pointer = getattr(pointer, scope_names[0])
                pointer = getattr(pointer, "weight")
            else:
                pointer = getattr(pointer, scope_names[0])
            if len(scope_names) >= 2:
                num = int(scope_names[1])
                pointer = pointer[num]
        try:
            assert pointer.shape == array.shape
        except AssertionError as e:
            e.args += (pointer.shape, array.shape)
            raise
        logger.info("Initialize PyTorch weight {}".format(name))
        pointer.data = torch.from_numpy(array)
    return model


class Attention(nn.Module):
    def __init__(self, nx, n_ctx, config, scale=False):
        super().__init__()

        n_state = nx  # in Attention: n_state=768 (nx=n_embd)
        # [switch nx => n_state from Block to Attention to keep identical to TF implem]
        assert n_state % config.n_head == 0
        self.register_buffer(
            "bias", torch.tril(torch.ones((n_ctx, n_ctx), dtype=torch.uint8)).view(1, 1, n_ctx, n_ctx)
        )
        self.register_buffer("masked_bias", torch.tensor(-1e4))
        self.n_head = config.n_head
        self.split_size = n_state
        self.scale = scale

        self.c_attn = Conv1D(n_state * 3, nx)
        self.c_proj = Conv1D(n_state, nx)
        self.attn_dropout = nn.Dropout(config.attn_pdrop)
        self.resid_dropout = nn.Dropout(config.resid_pdrop)
        self.pruned_heads = set()

    def prune_heads(self, heads):
        if len(heads) == 0:
            return
        heads, index = find_pruneable_heads_and_indices(
            heads, self.n_head, self.split_size // self.n_head, self.pruned_heads
        )
        index_attn = torch.cat([index, index + self.split_size, index + (2 * self.split_size)])

        # Prune conv1d layers
        self.c_attn = prune_conv1d_layer(self.c_attn, index_attn, dim=1)
        self.c_proj = prune_conv1d_layer(self.c_proj, index, dim=0)

        # Update hyper params
        self.split_size = (self.split_size // self.n_head) * (self.n_head - len(heads))
        self.n_head = self.n_head - len(heads)
        self.pruned_heads = self.pruned_heads.union(heads)

    def _attn(self, q, k, v, attention_mask=None, head_mask=None, output_attentions=False):
        w = torch.matmul(q, k)
        if self.scale:
            w = w / (float(v.size(-1)) ** 0.5)
        nd, ns = w.size(-2), w.size(-1)
        mask = self.bias[:, :, ns - nd : ns, :ns]
        w = torch.where(mask.bool(), w, self.masked_bias.to(w.dtype))

        if attention_mask is not None:
            # Apply the attention mask
            w = w + attention_mask

        w = nn.Softmax(dim=-1)(w)
        w = self.attn_dropout(w)

        # Mask heads if we want to
        if head_mask is not None:
            w = w * head_mask

        outputs = [torch.matmul(w, v)]
        if output_attentions:
            outputs.append(w)
        return outputs

    def merge_heads(self, x):
        x = x.permute(0, 2, 1, 3).contiguous()
        new_x_shape = x.size()[:-2] + (x.size(-2) * x.size(-1),)
        return x.view(*new_x_shape)  # in Tensorflow implem: fct merge_states

    def split_heads(self, x, k=False):
        new_x_shape = x.size()[:-1] + (self.n_head, x.size(-1) // self.n_head)
        x = x.view(*new_x_shape)  # in Tensorflow implem: fct split_states
        if k:
            return x.permute(0, 2, 3, 1)  # (batch, head, head_features, seq_length)
        else:
            return x.permute(0, 2, 1, 3)  # (batch, head, seq_length, head_features)

    def forward(
        self, x, layer_past=None, attention_mask=None, head_mask=None, use_cache=False, output_attentions=False
    ):
        x = self.c_attn(x)
        query, key, value = x.split(self.split_size, dim=2)
        query = self.split_heads(query)
        key = self.split_heads(key, k=True)
        value = self.split_heads(value)
        if layer_past is not None:
            past_key, past_value = layer_past[0].transpose(-2, -1), layer_past[1]  # transpose back cf below
            key = torch.cat((past_key, key), dim=-1)
            value = torch.cat((past_value, value), dim=-2)

        if use_cache is True:
            present = torch.stack((key.transpose(-2, -1), value))  # transpose to have same shapes for stacking
        else:
            present = (None,)

        attn_outputs = self._attn(query, key, value, attention_mask, head_mask, output_attentions)
        a = attn_outputs[0]

        a = self.merge_heads(a)
        a = self.c_proj(a)
        a = self.resid_dropout(a)

        outputs = [a, present] + attn_outputs[1:]
        return outputs  # a, present, (attentions)


class MLP(nn.Module):
    def __init__(self, n_state, config):  # in MLP: n_state=3072 (4 * n_embd)
        super().__init__()
        nx = config.n_embd
        self.c_fc = Conv1D(n_state, nx)
        self.c_proj = Conv1D(nx, n_state)
        self.act = ACT2FN[config.activation_function]
        self.dropout = nn.Dropout(config.resid_pdrop)

    def forward(self, x):
        h = self.act(self.c_fc(x))
        h2 = self.c_proj(h)
        return self.dropout(h2)


class Block(nn.Module):
    def __init__(self, n_ctx, config, scale=False):
        super().__init__()
        nx = config.n_embd
        self.ln_1 = nn.LayerNorm(nx, eps=config.layer_norm_epsilon)
        self.attn = Attention(nx, n_ctx, config, scale)
        self.ln_2 = nn.LayerNorm(nx, eps=config.layer_norm_epsilon)
        self.mlp = MLP(4 * nx, config)

    def forward(
        self, x, layer_past=None, attention_mask=None, head_mask=None, use_cache=False, output_attentions=False,
    ):
        output_attn = self.attn(
            self.ln_1(x),
            layer_past=layer_past,
            attention_mask=attention_mask,
            head_mask=head_mask,
            use_cache=use_cache,
            output_attentions=output_attentions,
        )
        a = output_attn[0]  # output_attn: a, present, (attentions)

        x = x + a
        m = self.mlp(self.ln_2(x))
        x = x + m

        outputs = [x] + output_attn[1:]
        return outputs  # x, present, (attentions)


class GPT2PreTrainedModel(PreTrainedModel):
    """ An abstract class to handle weights initialization and
        a simple interface for downloading and loading pretrained models.
    """

    config_class = GPT2Config
    load_tf_weights = load_tf_weights_in_gpt2
    base_model_prefix = "transformer"

    def __init__(self, *inputs, **kwargs):
        super().__init__(*inputs, **kwargs)

    def _init_weights(self, module):
        """ Initialize the weights.
        """
        if isinstance(module, (nn.Linear, nn.Embedding, Conv1D)):
            # Slightly different from the TF version which uses truncated_normal for initialization
            # cf https://github.com/pytorch/pytorch/pull/5617
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
            if isinstance(module, (nn.Linear, Conv1D)) and module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.LayerNorm):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)


GPT2_START_DOCSTRING = r"""

    This model is a PyTorch `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`_ sub-class.
    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general
    usage and behavior.

    Parameters:
        config (:class:`~transformers.GPT2Config`): Model configuration class with all the parameters of the model.
            Initializing with a config file does not load the weights associated with the model, only the configuration.
            Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
"""

GPT2_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, input_ids_length)`):
            :obj:`input_ids_length` = ``sequence_length`` if ``past`` is ``None`` else ``past[0].shape[-2]`` (``sequence_length`` of input past key value states).
            Indices of input sequence tokens in the vocabulary.

            If `past` is used, only `input_ids` that do not have their past calculated should be passed as `input_ids`.

            Indices can be obtained using :class:`transformers.GPT2Tokenizer`.
            See :func:`transformers.PreTrainedTokenizer.encode` and
            :func:`transformers.PreTrainedTokenizer.encode_plus` for details.

            `What are input IDs? <../glossary.html#input-ids>`__

        past (:obj:`List[torch.FloatTensor]` of length :obj:`config.n_layers`):
            Contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
            (see `past` output below). Can be used to speed up sequential decoding.
            The `input_ids` which have their past given to this model should not be passed as `input_ids` as they have already been computed.
        attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
            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.

            `What are attention masks? <../glossary.html#attention-mask>`__
        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, input_ids_length)`, `optional`, defaults to :obj:`None`):
            `input_ids_length` = `sequence_length if `past` is None else 1
            Segment token indices to indicate first and second portions of the inputs.
            Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
            corresponds to a `sentence B` token
            `What are token type IDs? <../glossary.html#token-type-ids>`_
        position_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
            Indices of positions of each input sequence tokens in the position embeddings.
            Selected in the range ``[0, config.max_position_embeddings - 1]``.

            `What are position IDs? <../glossary.html#position-ids>`_
        head_mask (:obj:`torch.FloatTensor` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`, defaults to :obj:`None`):
            Mask to nullify selected heads of the self-attention modules.
            Mask values selected in ``[0, 1]``:
            :obj:`1` indicates the head is **not masked**, :obj:`0` indicates the head is **masked**.
        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
            than the model's internal embedding lookup matrix.
            If `past` is used, optionally only the last `inputs_embeds` have to be input (see `past`).
        use_cache (:obj:`bool`):
            If `use_cache` is True, `past` key value states are returned and can be used to speed up decoding (see `past`). Defaults to `True`.
        output_attentions (:obj:`bool`, `optional`, defaults to :obj:`None`):
            If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
"""


@add_start_docstrings(
    "The bare GPT2 Model transformer outputting raw hidden-states without any specific head on top.",
    GPT2_START_DOCSTRING,
)
class GPT2Model(GPT2PreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        self.wte = nn.Embedding(config.vocab_size, config.n_embd)
        self.wpe = nn.Embedding(config.n_positions, config.n_embd)
        self.drop = nn.Dropout(config.embd_pdrop)
        self.h = nn.ModuleList([Block(config.n_ctx, config, scale=True) for _ in range(config.n_layer)])
        self.ln_f = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)

        self.init_weights()

    def get_input_embeddings(self):
        return self.wte

    def set_input_embeddings(self, new_embeddings):
        self.wte = new_embeddings

    def _prune_heads(self, heads_to_prune):
        """ Prunes heads of the model.
            heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
        """
        for layer, heads in heads_to_prune.items():
            self.h[layer].attn.prune_heads(heads)

    @add_start_docstrings_to_callable(GPT2_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="gpt2")
    def forward(
        self,
        input_ids=None,
        past=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
    ):
        r"""
    Return:
        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.GPT2Config`) and inputs:
        last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
            Sequence of hidden-states at the last layer of the model.
            If `past` is used only the last hidden-state of the sequences of shape :obj:`(batch_size, 1, hidden_size)` is output.
        past (:obj:`List[torch.FloatTensor]` of length :obj:`config.n_layers` with each tensor of shape :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`):
            Contains pre-computed hidden-states (key and values in the attention blocks).
            Can be used (see `past` input) to speed up sequential decoding.
        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True``) is passed or when ``config.output_hidden_states=True``:
            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
            of shape :obj:`(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.

            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
            heads.
        """
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache

        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
        elif input_ids is not None:
            input_shape = input_ids.size()
            input_ids = input_ids.view(-1, input_shape[-1])
            batch_size = input_ids.shape[0]
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
            batch_size = inputs_embeds.shape[0]
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

        if token_type_ids is not None:
            token_type_ids = token_type_ids.view(-1, input_shape[-1])
        if position_ids is not None:
            position_ids = position_ids.view(-1, input_shape[-1])

        if past is None:
            past_length = 0
            past = [None] * len(self.h)
        else:
            past_length = past[0][0].size(-2)
        if position_ids is None:
            device = input_ids.device if input_ids is not None else inputs_embeds.device
            position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device)
            position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])

        # Attention mask.
        if attention_mask is not None:
            assert batch_size > 0, "batch_size has to be defined and > 0"
            attention_mask = attention_mask.view(batch_size, -1)
            # We create a 3D attention mask from a 2D tensor mask.
            # Sizes are [batch_size, 1, 1, to_seq_length]
            # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
            # this attention mask is more simple than the triangular masking of causal attention
            # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
            attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)

            # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
            # masked positions, this operation will create a tensor which is 0.0 for
            # positions we want to attend and -10000.0 for masked positions.
            # Since we are adding it to the raw scores before the softmax, this is
            # effectively the same as removing these entirely.
            attention_mask = attention_mask.to(dtype=next(self.parameters()).dtype)  # fp16 compatibility
            attention_mask = (1.0 - attention_mask) * -10000.0

        # Prepare head mask if needed
        # 1.0 in head_mask indicate we keep the head
        # attention_probs has shape bsz x n_heads x N x N
        # head_mask has shape n_layer x batch x n_heads x N x N
        head_mask = self.get_head_mask(head_mask, self.config.n_layer)

        if inputs_embeds is None:
            inputs_embeds = self.wte(input_ids)
        position_embeds = self.wpe(position_ids)
        if token_type_ids is not None:
            token_type_embeds = self.wte(token_type_ids)
        else:
            token_type_embeds = 0
        hidden_states = inputs_embeds + position_embeds + token_type_embeds
        hidden_states = self.drop(hidden_states)

        output_shape = input_shape + (hidden_states.size(-1),)

        presents = ()
        all_attentions = []
        all_hidden_states = ()
        for i, (block, layer_past) in enumerate(zip(self.h, past)):
            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states.view(*output_shape),)

            outputs = block(
                hidden_states,
                layer_past=layer_past,
                attention_mask=attention_mask,
                head_mask=head_mask[i],
                use_cache=use_cache,
                output_attentions=output_attentions,
            )

            hidden_states, present = outputs[:2]
            if use_cache is True:
                presents = presents + (present,)

            if output_attentions:
                all_attentions.append(outputs[2])

        hidden_states = self.ln_f(hidden_states)

        hidden_states = hidden_states.view(*output_shape)
        # Add last hidden state
        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        outputs = (hidden_states,)
        if use_cache is True:
            outputs = outputs + (presents,)
        if output_hidden_states:
            outputs = outputs + (all_hidden_states,)
        if output_attentions:
            # let the number of heads free (-1) so we can extract attention even after head pruning
            attention_output_shape = input_shape[:-1] + (-1,) + all_attentions[0].shape[-2:]
            all_attentions = tuple(t.view(*attention_output_shape) for t in all_attentions)
            outputs = outputs + (all_attentions,)
        return outputs  # last hidden state, (presents), (all hidden_states), (attentions)


@add_start_docstrings(
    """The GPT2 Model transformer with a language modeling head on top
    (linear layer with weights tied to the input embeddings). """,
    GPT2_START_DOCSTRING,
)
class GPT2LMHeadModel(GPT2PreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.transformer = GPT2Model(config)
        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)

        self.init_weights()

    def get_output_embeddings(self):
        return self.lm_head

    def prepare_inputs_for_generation(self, input_ids, past, **kwargs):
        # only last token for inputs_ids if past is defined in kwargs
        if past:
            input_ids = input_ids[:, -1].unsqueeze(-1)

        return {"input_ids": input_ids, "past": past, "use_cache": kwargs["use_cache"]}

    @add_start_docstrings_to_callable(GPT2_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="gpt2")
    def forward(
        self,
        input_ids=None,
        past=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        labels=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
    ):
        r"""
        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
            Labels for language modeling.
            Note that the labels **are shifted** inside the model, i.e. you can set ``labels = input_ids``
            Indices are selected in ``[-100, 0, ..., config.vocab_size]``
            All labels set to ``-100`` are ignored (masked), the loss is only
            computed for labels in ``[0, ..., config.vocab_size]``

    Return:
        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.GPT2Config`) and inputs:
        loss (:obj:`torch.FloatTensor` of shape `(1,)`, `optional`, returned when ``labels`` is provided)
            Language modeling loss.
        prediction_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
        past (:obj:`List[torch.FloatTensor]` of length :obj:`config.n_layers` with each tensor of shape :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`):
            Contains pre-computed hidden-states (key and values in the attention blocks).
            Can be used (see `past` input) to speed up sequential decoding.
        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
            of shape :obj:`(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.

            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
            heads.
        """
        transformer_outputs = self.transformer(
            input_ids,
            past=past,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
        )
        hidden_states = transformer_outputs[0]

        lm_logits = self.lm_head(hidden_states)

        outputs = (lm_logits,) + transformer_outputs[1:]
        if labels is not None:
            # Shift so that tokens < n predict n
            shift_logits = lm_logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
            outputs = (loss,) + outputs

        return outputs  # (loss), lm_logits, presents, (all hidden_states), (attentions)


@add_start_docstrings(
    """The GPT2 Model transformer with a language modeling and a multiple-choice classification
    head on top e.g. for RocStories/SWAG tasks. The two heads are two linear layers.
    The language modeling head has its weights tied to the input embeddings,
    the classification head takes as input the input of a specified classification token index in the input sequence).
""",
    GPT2_START_DOCSTRING,
)
class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        config.num_labels = 1
        self.transformer = GPT2Model(config)
        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
        self.multiple_choice_head = SequenceSummary(config)

        self.init_weights()

    def get_output_embeddings(self):
        return self.lm_head

    @add_start_docstrings_to_callable(GPT2_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids=None,
        past=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        mc_token_ids=None,
        labels=None,
        mc_labels=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        **kwargs
    ):
        r"""
        mc_token_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, num_choices)`, `optional`, default to index of the last token of the input)
            Index of the classification token in each input sequence.
            Selected in the range ``[0, input_ids.size(-1) - 1[``.
        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`)
            Labels for language modeling.
            Note that the labels **are shifted** inside the model, i.e. you can set ``labels = input_ids``
            Indices are selected in ``[-1, 0, ..., config.vocab_size]``
            All labels set to ``-100`` are ignored (masked), the loss is only
            computed for labels in ``[0, ..., config.vocab_size]``
        mc_labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size)`, `optional`, defaults to :obj:`None`)
            Labels for computing the multiple choice classification loss.
            Indices should be in ``[0, ..., num_choices]`` where `num_choices` is the size of the second dimension
            of the input tensors. (see `input_ids` above)
        kwargs (:obj:`Dict[str, any]`, optional, defaults to `{}`):
            Used to hide legacy arguments that have been deprecated.

    Return:
        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.GPT2Config`) and inputs:
        lm_loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when ``labels`` is provided):
            Language modeling loss.
        mc_loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`mc_labels` is provided):
            Multiple choice classification loss.
        lm_prediction_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, num_choices, sequence_length, config.vocab_size)`):
            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
        mc_prediction_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, num_choices)`):
            Prediction scores of the multiple choice classification head (scores for each choice before SoftMax).
        past (:obj:`List[torch.FloatTensor]` of length :obj:`config.n_layers` with each tensor of shape :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`):
            Contains pre-computed hidden-states (key and values in the attention blocks).
            Can be used (see `past` input) to speed up sequential decoding.
        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
            of shape :obj:`(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.

            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
            heads.

    Examples::

        >>> import torch
        >>> from transformers import GPT2Tokenizer, GPT2DoubleHeadsModel

        >>> tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
        >>> model = GPT2DoubleHeadsModel.from_pretrained('gpt2')

        >>> # Add a [CLS] to the vocabulary (we should train it also!)
        >>> num_added_tokens = tokenizer.add_special_tokens({'cls_token': '[CLS]'})

        >>> embedding_layer = model.resize_token_embeddings(len(tokenizer))  # Update the model embeddings with the new vocabulary size

        >>> choices = ["Hello, my dog is cute [CLS]", "Hello, my cat is cute [CLS]"]
        >>> encoded_choices = [tokenizer.encode(s) for s in choices]
        >>> cls_token_location = [tokens.index(tokenizer.cls_token_id) for tokens in encoded_choices]

        >>> input_ids = torch.tensor(encoded_choices).unsqueeze(0)  # Batch size: 1, number of choices: 2
        >>> mc_token_ids = torch.tensor([cls_token_location])  # Batch size: 1

        >>> outputs = model(input_ids, mc_token_ids=mc_token_ids)
        >>> lm_prediction_scores, mc_prediction_scores = outputs[:2]

        """
        if "lm_labels" in kwargs:
            warnings.warn(
                "The `lm_labels` argument is deprecated and will be removed in a future version, use `labels` instead.",
                DeprecationWarning,
            )
            labels = kwargs.pop("lm_labels")
        assert kwargs == {}, f"Unexpected keyword arguments: {list(kwargs.keys())}."

        transformer_outputs = self.transformer(
            input_ids,
            past=past,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
        )

        hidden_states = transformer_outputs[0]

        lm_logits = self.lm_head(hidden_states)
        mc_logits = self.multiple_choice_head(hidden_states, mc_token_ids).squeeze(-1)

        outputs = (lm_logits, mc_logits) + transformer_outputs[1:]
        if mc_labels is not None:
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(mc_logits.view(-1, mc_logits.size(-1)), mc_labels.view(-1))
            outputs = (loss,) + outputs
        if labels is not None:
            shift_logits = lm_logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
            outputs = (loss,) + outputs

        return outputs  # (lm loss), (mc loss), lm logits, mc logits, presents, (all hidden_states), (attentions)