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."""

from __future__ import absolute_import, division, print_function, unicode_literals

import collections
import json
import logging
import math
import os
import sys
from io import open

import torch
import torch.nn as nn
from torch.nn import CrossEntropyLoss
from torch.nn.parameter import Parameter

from .modeling_utils import (Conv1D, CONFIG_NAME, WEIGHTS_NAME, PretrainedConfig,
                             PreTrainedModel, prune_conv1d_layer, SequenceSummary)
from .modeling_bert import BertLayerNorm as LayerNorm

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"}
OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP = {"openai-gpt": "https://s3.amazonaws.com/models.huggingface.co/bert/openai-gpt-config.json"}


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))

    names = json.load(open(openai_checkpoint_folder_path + '/parameters_names.json', "r", encoding='utf-8'))
    shapes = json.load(open(openai_checkpoint_folder_path + '/params_shapes.json', "r", encoding='utf-8'))
    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):
                l = re.split(r'(\d+)', m_name)
            else:
                l = [m_name]
            if l[0] == 'g':
                pointer = getattr(pointer, 'weight')
            elif l[0] == 'b':
                pointer = getattr(pointer, 'bias')
            elif l[0] == 'w':
                pointer = getattr(pointer, 'weight')
            else:
                pointer = getattr(pointer, l[0])
            if len(l) >= 2:
                num = int(l[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 OpenAIGPTConfig(PretrainedConfig):
    """
    Configuration class to store the configuration of a `OpenAIGPTModel`.

    Args:
        vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `OpenAIGPTModel` or a configuration json file.
        n_special: The number of special tokens to learn during fine-tuning ('[SEP]', '[CLF]', ...)
        n_positions: Number of positional embeddings.
        n_ctx: Size of the causal mask (usually same as n_positions).
        n_embd: Dimensionality of the embeddings and hidden states.
        n_layer: Number of hidden layers in the Transformer encoder.
        n_head: Number of attention heads for each attention layer in
            the Transformer encoder.
        afn: The non-linear activation function (function or string) in the
            encoder and pooler. If string, "gelu", "relu" and "swish" are supported.
        resid_pdrop: The dropout probabilitiy for all fully connected
            layers in the embeddings, encoder, and pooler.
        attn_pdrop: The dropout ratio for the attention
            probabilities.
        embd_pdrop: The dropout ratio for the embeddings.
        layer_norm_epsilon: epsilon to use in the layer norm layers
        initializer_range: The sttdev of the truncated_normal_initializer for
            initializing all weight matrices.
        predict_special_tokens: should we predict special tokens (when the model has a LM head)
    """
    pretrained_config_archive_map = OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP

    def __init__(
        self,
        vocab_size_or_config_json_file=40478,
        n_positions=512,
        n_ctx=512,
        n_embd=768,
        n_layer=12,
        n_head=12,
        afn="gelu",
        resid_pdrop=0.1,
        embd_pdrop=0.1,
        attn_pdrop=0.1,
        layer_norm_epsilon=1e-5,
        initializer_range=0.02,
        predict_special_tokens=True,

        num_labels=1,
        summary_type='token_ids',
        summary_use_proj=True,
        summary_activation=None,
        summary_proj_to_labels=True,
        summary_first_dropout=0.1,
        **kwargs
    ):
        """Constructs OpenAIGPTConfig.
        """
        super(OpenAIGPTConfig, self).__init__(**kwargs)

        if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
                        and isinstance(vocab_size_or_config_json_file, unicode)):
            with open(vocab_size_or_config_json_file, "r", encoding="utf-8") as reader:
                json_config = json.loads(reader.read())
            for key, value in json_config.items():
                self.__dict__[key] = value
        elif isinstance(vocab_size_or_config_json_file, int):
            self.vocab_size = vocab_size_or_config_json_file
            self.n_ctx = n_ctx
            self.n_positions = n_positions
            self.n_embd = n_embd
            self.n_layer = n_layer
            self.n_head = n_head
            self.afn = afn
            self.resid_pdrop = resid_pdrop
            self.embd_pdrop = embd_pdrop
            self.attn_pdrop = attn_pdrop
            self.layer_norm_epsilon = layer_norm_epsilon
            self.initializer_range = initializer_range
            self.predict_special_tokens = predict_special_tokens

            self.num_labels = num_labels
            self.summary_type = summary_type
            self.summary_use_proj = summary_use_proj
            self.summary_activation = summary_activation
            self.summary_first_dropout = summary_first_dropout
            self.summary_proj_to_labels = summary_proj_to_labels
        else:
            raise ValueError(
                "First argument must be either a vocabulary size (int)"
                "or the path to a pretrained model config file (str)"
            )

    @property
    def max_position_embeddings(self):
        return self.n_positions

    @property
    def hidden_size(self):
        return self.n_embd

    @property
    def num_attention_heads(self):
        return self.n_head

    @property
    def num_hidden_layers(self):
        return self.n_layer


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

    def prune_heads(self, heads):
        if len(heads) == 0:
            return
        mask = torch.ones(self.n_head, self.split_size // self.n_head)
        for head in 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)

    def _attn(self, q, k, v, 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 + -1e9 * (1 - b)

        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, 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, 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(MLP, self).__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(Block, self).__init__()
        nx = config.n_embd
        self.attn = Attention(nx, n_ctx, config, scale)
        self.ln_1 = LayerNorm(nx, eps=config.layer_norm_epsilon)
        self.mlp = MLP(4 * nx, config)
        self.ln_2 = LayerNorm(nx, eps=config.layer_norm_epsilon)

    def forward(self, x, head_mask=None):
        attn_outputs = self.attn(x, 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 dowloading 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__(self, *inputs, **kwargs):
        super(OpenAIGPTPreTrainedModel, self).__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, LayerNorm):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)


class OpenAIGPTModel(OpenAIGPTPreTrainedModel):
    """OpenAI GPT model ("Improving Language Understanding by Generative Pre-Training").

    OpenAI GPT uses a single embedding matrix to store the word and special embeddings.
    Special tokens embeddings are additional tokens that are not pre-trained, such as: [SEP], [CLS]...

    Special tokens need to be trained during the fine-tuning if you use them.
    The number of special embeddings can be controlled using the ``set_num_special_tokens(num_special_tokens)`` function.

    The embeddings are ordered as follow in the token embeddings matrix:

    ::

        [0,                                                         ----------------------
         ...                                                        -> word embeddings
         config.vocab_size - 1,                                     ______________________
         config.vocab_size,
         ...                                                        -> special embeddings
         config.vocab_size + n_special - 1]                  ______________________

    where ``total_tokens_embeddings``  is:

    ::

        total_tokens_embeddings = config.vocab_size + n_special

    You should use the associated indices to index the embeddings.

    Args:
        `config`: a OpenAIGPTConfig class instance with the configuration to build a new model
        `output_attentions`: If True, also output attentions weights computed by the model at each layer. Default: False
        `keep_multihead_output`: If True, saves output of the multi-head attention module with its gradient.
            This can be used to compute head importance metrics. Default: False


    Example::

        config = modeling_openai.OpenAIGPTConfig()
        model = modeling_openai.OpenAIGPTModel(config)
    """

    def __init__(self, config):
        super(OpenAIGPTModel, self).__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.apply(self.init_weights)

    def _resize_token_embeddings(self, new_num_tokens):
        self.tokens_embed = self._get_resized_embeddings(self.tokens_embed, new_num_tokens)
        return self.tokens_embed

    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)

    def forward(self, input_ids, position_ids=None, token_type_ids=None, head_mask=None):
        """
        Performs a model forward pass. **Can be called by calling the class directly, once it has been instantiated.**

        Args:
            `input_ids`: a ``torch.LongTensor`` of shape [batch_size, sequence_length] (or more generally [d_1, ..., d_n, sequence_length]
                were d_1 ... d_n are arbitrary dimensions) with the word BPE token indices selected in the range [0, total_tokens_embeddings[
            `position_ids`: an optional ``torch.LongTensor`` with the same shape as input_ids
                with the position indices (selected in the range [0, config.n_positions - 1[.
            `token_type_ids`: an optional ``torch.LongTensor`` with the same shape as input_ids
                You can use it to add a third type of embedding to each input token in the sequence
                (the previous two being the word and position embeddings).
                The input, position and token_type embeddings are summed inside the Transformer before the first
                self-attention block.
            `head_mask`: an optional ``torch.Tensor`` of shape [num_heads] or [num_layers, num_heads] with indices between 0 and 1.
                It's a mask to be used to nullify some heads of the transformer. 1.0 => head is fully masked, 0.0 => head is not masked.

        Returns:
            ``hidden_states``, a list of all the encoded-hidden-states in the model (length of the list is number
            of layers + 1 for the output of the embeddings)
            as ``torch.FloatTensor`` of size [batch_size, sequence_length, hidden_size]
            (or more generally [d_1, ..., d_n, hidden_size] were d_1 ... d_n are the dimension of input_ids)

        Example::

            # Already been converted into BPE token ids
            input_ids = torch.LongTensor([[31, 51, 99], [15, 5, 0]])

            hidden_states = model(input_ids)
            # or
            hidden_states = model.forward(input_ids)
        """
        if position_ids is None:
            # This was used when we had a single embedding matrice from position and token embeddings
            # start = self.config.vocab_size + self.config.n_special
            # end = start + input_ids.size(-1)
            # position_ids = torch.arange(start, end, dtype=torch.long, device=input_ids.device)
            position_ids = torch.arange(input_ids.size(-1), dtype=torch.long, device=input_ids.device)
            position_ids = position_ids.unsqueeze(0).expand_as(input_ids)

        # 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

        input_shape = input_ids.size()
        input_ids = input_ids.view(-1, input_ids.size(-1))
        position_ids = position_ids.view(-1, position_ids.size(-1))

        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, 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)


class OpenAIGPTLMHeadModel(OpenAIGPTPreTrainedModel):
    """OpenAI GPT model with a Language Modeling head ("Improving Language Understanding by Generative Pre-Training").

    OpenAI GPT use a single embedding matrix to store the word and special embeddings.
    Special tokens embeddings are additional tokens that are not pre-trained: [SEP], [CLS]...
    Special tokens need to be trained during the fine-tuning if you use them. The number of special embeddings
    can be controlled using the ``set_num_special_tokens(num_special_tokens)`` function.

    The embeddings are ordered as follow in the token embeddings matrix:

    ::

        [0,                                                         ----------------------
         ...                                                        -> word embeddings
         config.vocab_size - 1,                                     ______________________
         config.vocab_size,
         ...                                                        -> special embeddings
         config.vocab_size + config.n_special - 1]                  ______________________

    where ``total_tokens_embeddings`` can be obtained as ``config.total_tokens_embeddings`` and is:

    ::

        total_tokens_embeddings = config.vocab_size + config.n_special

    You should use the associated indices to index the embeddings.

    Args:
        `config`: a OpenAIGPTConfig class instance with the configuration to build a new model
        `output_attentions`: If True, also output attentions weights computed by the model at each layer. Default: False
        `keep_multihead_output`: If True, saves output of the multi-head attention module with its gradient.
            This can be used to compute head importance metrics. Default: False


    Example::

        config = modeling_openai.OpenAIGPTConfig()
        model = modeling_openai.OpenAIGPTLMHeadModel(config)
    """

    def __init__(self, config):
        super(OpenAIGPTLMHeadModel, self).__init__(config)
        self.transformer = OpenAIGPTModel(config)
        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)

        self.apply(self.init_weights)
        self.tie_weights()

    def tie_weights(self):
        """ Make sure we are sharing the input and output embeddings.
            Export to TorchScript can't handle parameter sharing so we are cloning them instead.
        """
        self._tie_or_clone_weights(self.lm_head,
                                   self.transformer.tokens_embed)

    def forward(self, input_ids, position_ids=None, token_type_ids=None, lm_labels=None, head_mask=None):
        """
        Performs a model forward pass. **Can be called by calling the class directly, once it has been instantiated.**

        Args:
            `input_ids`: a ``torch.LongTensor`` of shape [batch_size, sequence_length] (or more generally [d_1, ..., d_n, sequence_length]
                were d_1 ... d_n are arbitrary dimensions) with the word BPE token indices selected in the range [0, total_tokens_embeddings[
            `position_ids`: an optional ``torch.LongTensor`` with the same shape as input_ids
                with the position indices (selected in the range [0, config.n_positions - 1[.
            `token_type_ids`: an optional ``torch.LongTensor`` with the same shape as input_ids
                You can use it to add a third type of embedding to each input token in the sequence
                (the previous two being the word and position embeddings).
                The input, position and token_type embeddings are summed inside the Transformer before the first
                self-attention block.
            `lm_labels`: optional language modeling labels: ``torch.LongTensor`` of shape [batch_size, sequence_length]
                with indices selected in [-1, 0, ..., vocab_size]. All labels set to -1 are ignored (masked), the loss
                is only computed for the labels set in [0, ..., vocab_size]
            `head_mask`: an optional ``torch.Tensor`` of shape [num_heads] or [num_layers, num_heads] with indices between 0 and 1.
                It's a mask to be used to nullify some heads of the transformer. 1.0 => head is fully masked, 0.0 => head is not masked.

        Returns:
            if ``lm_labels`` is not ``None``, outputs the language modeling loss. Otherwise, outputs ``lm_logits``,
            the language modeling logits as a ``torch.FloatTensor`` of size [batch_size, sequence_length,
            total_tokens_embeddings] (or more generally [d_1, ..., d_n, total_tokens_embeddings] where d_1 ... d_n are
            the dimension of input_ids)

        Example::

            # Already been converted into BPE token ids
            input_ids = torch.LongTensor([[31, 51, 99], [15, 5, 0]])

            lm_logits = model(input_ids)
            # or
            lm_logits = model.forward(input_ids)
        """
        transformer_outputs = self.transformer(input_ids, position_ids, token_type_ids, head_mask)
        hidden_states = transformer_outputs[0]
        lm_logits = self.lm_head(hidden_states)

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


class OpenAIGPTDoubleHeadsModel(OpenAIGPTPreTrainedModel):
    """OpenAI GPT model with a Language Modeling and a Multiple Choice head ("Improving Language Understanding by Generative Pre-Training").

    OpenAI GPT use a single embedding matrix to store the word and special embeddings.
    Special tokens embeddings are additional tokens that are not pre-trained: [SEP], [CLS]...
    Special tokens need to be trained during the fine-tuning if you use them.
    The number of special embeddings can be controlled using the ``set_num_special_tokens(num_special_tokens)``
    function.

    The embeddings are ordered as follow in the token embeddings matrix:

    ::

        [0,                                                         ----------------------
         ...                                                        -> word embeddings
         config.vocab_size - 1,                                     ______________________
         config.vocab_size,
         ...                                                        -> special embeddings
         config.vocab_size + n_special - 1]                  ______________________

    where ``total_tokens_embeddings`` is:

    ::

        total_tokens_embeddings = config.vocab_size + .n_special

    You should use the associate indices to index the embeddings.

    Args:
        `config`: a OpenAIGPTConfig class instance with the configuration to build a new model
        `output_attentions`: If True, also output attentions weights computed by the model at each layer. Default: False
        `keep_multihead_output`: If True, saves output of the multi-head attention module with its gradient.
            This can be used to compute head importance metrics. Default: False

    Example::

        config = modeling_openai.OpenAIGPTConfig()
        model = modeling_openai.OpenAIGPTDoubleHeadsModel(config)
    """

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

        self.transformer = OpenAIGPTModel(config)
        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
        self.multiple_choice_head = SequenceSummary(config)

        self.apply(self.init_weights)
        self.tie_weights()

    def tie_weights(self):
        """ Make sure we are sharing the input and output embeddings.
            Export to TorchScript can't handle parameter sharing so we are cloning them instead.
        """
        self._tie_or_clone_weights(self.lm_head,
                                   self.transformer.tokens_embed)

    def forward(self, input_ids, mc_token_ids=None, lm_labels=None, mc_labels=None, token_type_ids=None,
                position_ids=None, head_mask=None):
        """
        Performs a model forward pass. **Can be called by calling the class directly, once it has been instantiated.**

        Args:
            `input_ids`: a ``torch.LongTensor`` of shape [batch_size, num_choices, sequence_length] with the BPE token
                indices selected in the range [0, total_tokens_embeddings[
            `mc_token_ids`: a ``torch.LongTensor`` of shape [batch_size, num_choices] with the index of the token from
                which we should take the hidden state to feed the multiple choice classifier (usually last token of the sequence)
            `position_ids`: an optional ``torch.LongTensor`` with the same shape as input_ids
                with the position indices (selected in the range [0, config.n_positions - 1[.
            `token_type_ids`: an optional ``torch.LongTensor`` with the same shape as input_ids
                You can use it to add a third type of embedding to each input token in the sequence
                (the previous two being the word and position embeddings).
                The input, position and token_type embeddings are summed inside the Transformer before the first
                self-attention block.
            `lm_labels`: optional language modeling labels: ``torch.LongTensor`` of shape [batch_size, num_choices, sequence_length]
                with indices selected in [-1, 0, ..., total_tokens_embeddings]. All labels set to -1 are ignored (masked), the loss
                is only computed for the labels set in [0, ..., total_tokens_embeddings]
            `multiple_choice_labels`: optional multiple choice labels: ``torch.LongTensor`` of shape [batch_size]
                with indices selected in [0, ..., num_choices].
            `head_mask`: an optional ``torch.Tensor`` of shape [num_heads] or [num_layers, num_heads] with indices between 0 and 1.
                It's a mask to be used to nullify some heads of the transformer. 1.0 => head is fully masked, 0.0 => head is not masked.

        Returns:
            if ``lm_labels`` and ``multiple_choice_labels`` are not ``None``, outputs a tuple of losses with the
            language modeling loss and the multiple choice loss. Otherwise, returns a
            ``tuple(lm_logits, multiple_choice_logits)``.

                ``lm_logits`` are the language modeling logits as a ``torch.FloatTensor`` of size
                [batch_size, num_choices, sequence_length, total_tokens_embeddings]

                ``multiple_choice_logits``: the multiple choice logits as a ``torch.FloatTensor`` of
                size [batch_size, num_choices]

        Example::

            # Already been converted into BPE token ids
            input_ids = torch.LongTensor([[[31, 51, 99], [15, 5, 0]]])  # (bsz, number of choice, seq length)
            mc_token_ids = torch.LongTensor([[2], [1]]) # (bsz, number of choice)

            lm_logits, multiple_choice_logits = model(input_ids, mc_token_ids)
            # or
            lm_logits, multiple_choice_logits = model.forward(input_ids, mc_token_ids)
        """
        transformer_outputs = self.transformer(input_ids, position_ids, token_type_ids, head_mask)
        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(ignore_index=-1)
            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)