modeling_openai.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 model."""


import json
import logging
import math
import os

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

from .configuration_openai import OpenAIGPTConfig
from .file_utils import add_start_docstrings, add_start_docstrings_to_callable
from .modeling_utils import Conv1D, PreTrainedModel, SequenceSummary, prune_conv1d_layer


logger = logging.getLogger(__name__)

OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP = {
    "openai-gpt": "https://s3.amazonaws.com/models.huggingface.co/bert/openai-gpt-pytorch_model.bin"
}


def load_tf_weights_in_openai_gpt(model, config, openai_checkpoint_folder_path):
    """ Load tf pre-trained weights in a pytorch model (from NumPy arrays here)
    """
    import re
    import numpy as np

    if ".ckpt" in openai_checkpoint_folder_path:
        openai_checkpoint_folder_path = os.path.dirname(openai_checkpoint_folder_path)

    logger.info("Loading weights from {}".format(openai_checkpoint_folder_path))

    with open(openai_checkpoint_folder_path + "/parameters_names.json", "r", encoding="utf-8") as names_handle:
        names = json.load(names_handle)
    with open(openai_checkpoint_folder_path + "/params_shapes.json", "r", encoding="utf-8") as shapes_handle:
        shapes = json.load(shapes_handle)
    offsets = np.cumsum([np.prod(shape) for shape in shapes])
    init_params = [np.load(openai_checkpoint_folder_path + "/params_{}.npy".format(n)) for n in range(10)]
    init_params = np.split(np.concatenate(init_params, 0), offsets)[:-1]
    init_params = [param.reshape(shape) for param, shape in zip(init_params, shapes)]

    # This was used when we had a single embedding matrix for positions and tokens
    # init_params[0] = np.concatenate([init_params[1], init_params[0]], 0)
    # del init_params[1]
    init_params = [arr.squeeze() for arr in init_params]

    try:
        assert model.tokens_embed.weight.shape == init_params[1].shape
        assert model.positions_embed.weight.shape == init_params[0].shape
    except AssertionError as e:
        e.args += (model.tokens_embed.weight.shape, init_params[1].shape)
        e.args += (model.positions_embed.weight.shape, init_params[0].shape)
        raise

    model.tokens_embed.weight.data = torch.from_numpy(init_params[1])
    model.positions_embed.weight.data = torch.from_numpy(init_params[0])
    names.pop(0)
    # Pop position and token embedding arrays
    init_params.pop(0)
    init_params.pop(0)

    for name, array in zip(names, init_params):  # names[1:n_transfer], init_params[1:n_transfer]):
        name = name[6:]  # skip "model/"
        assert name[-2:] == ":0"
        name = name[:-2]
        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] == "g":
                pointer = getattr(pointer, "weight")
            elif scope_names[0] == "b":
                pointer = getattr(pointer, "bias")
            elif scope_names[0] == "w":
                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
        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


def gelu(x):
    return 0.5 * x * (1 + torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3))))


def swish(x):
    return x * torch.sigmoid(x)


ACT_FNS = {"relu": nn.ReLU, "swish": swish, "gelu": gelu}


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)).view(1, 1, n_ctx, n_ctx))
        self.n_head = config.n_head
        self.split_size = n_state
        self.scale = scale

        self.output_attentions = config.output_attentions

        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
        mask = torch.ones(self.n_head, self.split_size // self.n_head)
        heads = set(heads) - self.pruned_heads
        for head in heads:
            head -= sum(1 if h < head else 0 for h in self.pruned_heads)
            mask[head] = 0
        mask = mask.view(-1).contiguous().eq(1)
        index = torch.arange(len(mask))[mask].long()
        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):
        w = torch.matmul(q, k)
        if self.scale:
            w = w / math.sqrt(v.size(-1))
        # w = w * self.bias + -1e9 * (1 - self.bias)  # TF implem method: mask_attn_weights
        # XD: self.b may be larger than w, so we need to crop it
        b = self.bias[:, :, : w.size(-2), : w.size(-1)]
        w = w * b + -1e4 * (1 - b)

        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 self.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)
        else:
            return x.permute(0, 2, 1, 3)

    def forward(self, x, attention_mask=None, head_mask=None):
        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)

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

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

        outputs = [a] + attn_outputs[1:]
        return outputs  # a, (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 = ACT_FNS[config.afn]
        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.attn = Attention(nx, n_ctx, config, scale)
        self.ln_1 = nn.LayerNorm(nx, eps=config.layer_norm_epsilon)
        self.mlp = MLP(4 * nx, config)
        self.ln_2 = nn.LayerNorm(nx, eps=config.layer_norm_epsilon)

    def forward(self, x, attention_mask=None, head_mask=None):
        attn_outputs = self.attn(x, attention_mask=attention_mask, head_mask=head_mask)
        a = attn_outputs[0]

        n = self.ln_1(x + a)
        m = self.mlp(n)
        h = self.ln_2(n + m)

        outputs = [h] + attn_outputs[1:]
        return outputs


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

    config_class = OpenAIGPTConfig
    pretrained_model_archive_map = OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP
    load_tf_weights = load_tf_weights_in_openai_gpt
    base_model_prefix = "transformer"

    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)


OPENAI_GPT_START_DOCSTRING = r"""    
    This model is a PyTorch `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.

    .. _`Improving Language Understanding by Generative Pre-Training`:
        https://openai.com/blog/language-unsupervised/

    .. _`torch.nn.Module`:
        https://pytorch.org/docs/stable/nn.html#module

    Parameters:
        config (:class:`~transformers.OpenAIGPTConfig`): 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.
"""

OPENAI_GPT_INPUTS_DOCSTRING = r"""    
    Args:
        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
            Indices of input sequence tokens in the vocabulary. 
            
            Indices can be obtained using :class:`transformers.OpenAIGPTTokenizer`.
            See :func:`transformers.PreTrainedTokenizer.encode` and
            :func:`transformers.PreTrainedTokenizer.encode_plus` for details.
            
            `What are input IDs? <../glossary.html#input-ids>`__
        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, sequence_length)`, `optional`, defaults to :obj:`None`): 
            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**.
        input_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
            Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
            than the model's internal embedding lookup matrix.
"""


@add_start_docstrings(
    "The bare OpenAI GPT transformer model outputting raw hidden-states without any specific head on top.",
    OPENAI_GPT_START_DOCSTRING,
)
class OpenAIGPTModel(OpenAIGPTPreTrainedModel):

    def __init__(self, config):
        super().__init__(config)
        self.output_attentions = config.output_attentions
        self.output_hidden_states = config.output_hidden_states

        self.tokens_embed = nn.Embedding(config.vocab_size, config.n_embd)
        self.positions_embed = 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.init_weights()

    def get_input_embeddings(self):
        return self.tokens_embed

    def set_input_embeddings(self, new_embeddings):
        self.tokens_embed = 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(OPENAI_GPT_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
    ):
        r"""
    Return:
        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (config) 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.
        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned 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 ``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::

        tokenizer = OpenAIGPTTokenizer.from_pretrained('openai-gpt')
        model = OpenAIGPTModel.from_pretrained('openai-gpt')
        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
        outputs = model(input_ids)
        last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple

        """
        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])
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

        if position_ids is None:
            # Code is different from when we had a single embedding matrice from position and token embeddings
            device = input_ids.device if input_ids is not None else inputs_embeds.device
            position_ids = torch.arange(input_shape[-1], dtype=torch.long, device=device)
            position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])

        # Attention mask.
        if attention_mask is not None:
            # 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
        if head_mask is not None:
            if head_mask.dim() == 1:
                head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
                head_mask = head_mask.expand(self.config.n_layer, -1, -1, -1, -1)
            elif head_mask.dim() == 2:
                head_mask = (
                    head_mask.unsqueeze(1).unsqueeze(-1).unsqueeze(-1)
                )  # We can specify head_mask for each layer
            head_mask = head_mask.to(
                dtype=next(self.parameters()).dtype
            )  # switch to fload if need + fp16 compatibility
        else:
            head_mask = [None] * self.config.n_layer

        if inputs_embeds is None:
            inputs_embeds = self.tokens_embed(input_ids)
        position_embeds = self.positions_embed(position_ids)
        if token_type_ids is not None:
            token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1))
            token_type_embeds = self.tokens_embed(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),)

        all_attentions = ()
        all_hidden_states = ()
        for i, block in enumerate(self.h):
            if self.output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states.view(*output_shape),)

            outputs = block(hidden_states, attention_mask, head_mask[i])
            hidden_states = outputs[0]
            if self.output_attentions:
                all_attentions = all_attentions + (outputs[1],)

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

        outputs = (hidden_states.view(*output_shape),)
        if self.output_hidden_states:
            outputs = outputs + (all_hidden_states,)
        if self.output_attentions:
            outputs = outputs + (all_attentions,)
        return outputs  # last hidden state, (all hidden states), (all attentions)


@add_start_docstrings(
    """OpenAI GPT Model transformer with a language modeling head on top
    (linear layer with weights tied to the input embeddings). """,
    OPENAI_GPT_START_DOCSTRING,
)
class OpenAIGPTLMHeadModel(OpenAIGPTPreTrainedModel):

    def __init__(self, config):
        super().__init__(config)
        self.transformer = OpenAIGPTModel(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

    @add_start_docstrings_to_callable(OPENAI_GPT_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        labels=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 ``lm_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 (:obj:`~transformers.OpenAIGPTConfig`) 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. The token ids which have their past given to this model
            should not be passed as input ids as they have already been computed.
        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned 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 ``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::

        tokenizer = OpenAIGPTTokenizer.from_pretrained('openai-gpt')
        model = OpenAIGPTLMHeadModel.from_pretrained('openai-gpt')
        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
        outputs = model(input_ids, labels=input_ids)
        loss, logits = outputs[:2]

    """
        transformer_outputs = self.transformer(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
        )
        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, (all hidden states), (all attentions)


@add_start_docstrings(
    """OpenAI GPT 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).
""",
    OPENAI_GPT_START_DOCSTRING,
)
class OpenAIGPTDoubleHeadsModel(OpenAIGPTPreTrainedModel):

    def __init__(self, config):
        super().__init__(config)

        config.num_labels = 1
        self.transformer = OpenAIGPTModel(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(OPENAI_GPT_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        mc_token_ids=None,
        lm_labels=None,
        mc_labels=None,
    ):
        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[``.
        lm_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 ``lm_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)

    Return:
        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:obj:`~transformers.OpenAIGPTConfig`) and inputs:
        lm_loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when ``lm_labels`` is provided):
            Language modeling loss.
        mc_loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`multiple_choice_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. The token ids which have their past given to this model
            should not be passed as input ids as they have already been computed.
        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned 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 ``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::

        tokenizer = OpenAIGPTTokenizer.from_pretrained('openai-gpt')
        model = OpenAIGPTDoubleHeadsModel.from_pretrained('openai-gpt')
        tokenizer.add_special_tokens({'cls_token': '[CLS]'})  # Add a [CLS] to the vocabulary (we should train it also!)
        model.resize_token_embeddings(len(tokenizer))

        choices = ["Hello, my dog is cute [CLS]", "Hello, my cat is cute [CLS]"]
        input_ids = torch.tensor([tokenizer.encode(s) for s in choices]).unsqueeze(0)  # Batch size 1, 2 choices
        mc_token_ids = torch.tensor([input_ids.size(-1)-1, input_ids.size(-1)-1]).unsqueeze(0)  # Batch size 1

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

    """
        transformer_outputs = self.transformer(
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
        )
        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 lm_labels is not None:
            shift_logits = lm_logits[..., :-1, :].contiguous()
            shift_labels = lm_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, (all hidden_states), (attentions)