modeling_tf_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.
""" TF 2.0 OpenAI GPT-2 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 numpy as np
import tensorflow as tf

from .modeling_tf_utils import (TFPreTrainedModel, TFConv1D, TFSharedEmbeddings,
                                TFSequenceSummary, shape_list)
from .configuration_gpt2 import GPT2Config
from .file_utils import add_start_docstrings
from .modeling_tf_pytorch_utils import load_pytorch_checkpoint_in_tf2_model

logger = logging.getLogger(__name__)

TF_GPT2_PRETRAINED_MODEL_ARCHIVE_MAP = {"gpt2": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-tf_model.h5",
                                     "gpt2-medium": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-medium-tf_model.h5",
                                     "gpt2-large": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-large-tf_model.h5"}


def load_gpt2_pt_weights_in_tf2(tf_model, pytorch_checkpoint_path):
    # build the network
    inputs_list = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
    tf_inputs = tf.constant(inputs_list)
    tfo = tf_model(tf_inputs, training=False)
    return load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path, tf_inputs=tf_inputs)


def gelu(x):
    """Gaussian Error Linear Unit.
    This is a smoother version of the RELU.
    Original paper: https://arxiv.org/abs/1606.08415
    Args:
        x: float Tensor to perform activation.
    Returns:
        `x` with the GELU activation applied.
    """
    cdf = 0.5 * (1.0 + tf.tanh(
        (np.sqrt(2 / np.pi) * (x + 0.044715 * tf.pow(x, 3)))))
    return x * cdf


class TFAttention(tf.keras.layers.Layer):
    def __init__(self, nx, n_ctx, config, scale=False, **kwargs):
        super(TFAttention, self).__init__(**kwargs)
        self.output_attentions = config.output_attentions

        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.n_ctx = n_ctx
        self.n_head = config.n_head
        self.split_size = n_state
        self.scale = scale

        self.c_attn = TFConv1D(n_state * 3, nx, name='c_attn')
        self.c_proj = TFConv1D(n_state, nx, name='c_proj')
        self.attn_dropout = tf.keras.layers.Dropout(config.attn_pdrop)
        self.resid_dropout = tf.keras.layers.Dropout(config.resid_pdrop)
        self.pruned_heads = set()

    def prune_heads(self, heads):
        pass

    @staticmethod
    def causal_attention_mask(nd, ns, dtype):
        """1's in the lower triangle, counting from the lower right corner.
        Same as tf.matrix_band_part(tf.ones([nd, ns]), -1, ns-nd), but doesn't produce garbage on TPUs.
        """
        i = tf.range(nd)[:,None]
        j = tf.range(ns)
        m = i >= j - ns + nd
        return tf.cast(m, dtype)

    def _attn(self, inputs, training=False):
        q, k, v, attention_mask, head_mask = inputs
        # q, k, v have shape [batch, heads, sequence, features]
        w = tf.matmul(q, k, transpose_b=True)
        if self.scale:
            dk = tf.cast(tf.shape(k)[-1], tf.float32) # scale attention_scores
            w = w / tf.math.sqrt(dk)

        # w has shape [batch, heads, dst_sequence, src_sequence], where information flows from src to dst.
        _, _, nd, ns = shape_list(w)
        b = self.causal_attention_mask(nd, ns, dtype=w.dtype)
        b = tf.reshape(b, [1, 1, nd, ns])
        w = w * b - 1e4 * (1 - b)

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

        w = tf.nn.softmax(w, axis=-1)
        w = self.attn_dropout(w, training=training)

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

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

    def merge_heads(self, x):
        x = tf.transpose(x, [0, 2, 1, 3])
        x_shape = shape_list(x)
        new_x_shape = x_shape[:-2] + [x_shape[-2] * x_shape[-1]]
        return tf.reshape(x, new_x_shape)

    def split_heads(self, x):
        x_shape = shape_list(x)
        new_x_shape = x_shape[:-1] + [self.n_head, x_shape[-1] // self.n_head]
        x = tf.reshape(x, new_x_shape)
        return tf.transpose(x, (0, 2, 1, 3))  # (batch, head, seq_length, head_features)

    def call(self, inputs, training=False):
        x, layer_past, attention_mask, head_mask = inputs

        x = self.c_attn(x)
        query, key, value = tf.split(x, 3, axis=2)
        query = self.split_heads(query)
        key = self.split_heads(key)
        value = self.split_heads(value)
        if layer_past is not None:
            past_key, past_value = tf.unstack(layer_past, axis=1)
            key = tf.concat([past_key, key], axis=-2)
            value = tf.concat([past_value, value], axis=-2)
        present = tf.stack([key, value], axis=1)

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

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

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


class TFMLP(tf.keras.layers.Layer):
    def __init__(self, n_state, config, **kwargs):
        super(TFMLP, self).__init__(**kwargs)
        nx = config.n_embd
        self.c_fc = TFConv1D(n_state, nx, name='c_fc')
        self.c_proj = TFConv1D(nx, n_state, name='c_proj')
        self.act = gelu
        self.dropout = tf.keras.layers.Dropout(config.resid_pdrop)

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


class TFBlock(tf.keras.layers.Layer):
    def __init__(self, n_ctx, config, scale=False, **kwargs):
        super(TFBlock, self).__init__(**kwargs)
        nx = config.n_embd
        self.ln_1 = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_epsilon, name='ln_1')
        self.attn = TFAttention(nx, n_ctx, config, scale, name='attn')
        self.ln_2 = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_epsilon, name='ln_2')
        self.mlp = TFMLP(4 * nx, config, name='mlp')

    def call(self, inputs, training=False):
        x, layer_past, attention_mask, head_mask = inputs

        a = self.ln_1(x)
        output_attn = self.attn([a, layer_past, attention_mask, head_mask], training=training)
        a = output_attn[0]  # output_attn: a, present, (attentions)
        x = x + a

        m = self.ln_2(x)
        m = self.mlp(m, training=training)
        x = x + m

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


class TFGPT2MainLayer(tf.keras.layers.Layer):
    def __init__(self, config, *inputs, **kwargs):
        super(TFGPT2MainLayer, self).__init__(config, *inputs, **kwargs)
        self.output_hidden_states = config.output_hidden_states
        self.output_attentions = config.output_attentions
        self.num_hidden_layers = config.n_layer
        self.vocab_size = config.vocab_size
        self.n_embd = config.n_embd

        self.wte = TFSharedEmbeddings(config.vocab_size, config.hidden_size, name='wte')
        self.wpe = tf.keras.layers.Embedding(config.n_positions, config.n_embd, name='wpe')
        self.drop = tf.keras.layers.Dropout(config.embd_pdrop)
        self.h = [TFBlock(config.n_ctx,
                          config,
                          scale=True,
                          name='h_._{}'.format(i)) for i in range(config.n_layer)]
        self.ln_f = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_epsilon, name='ln_f')

    def _resize_token_embeddings(self, new_num_tokens):
        raise NotImplementedError

    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}
        """
        raise NotImplementedError

    def call(self, inputs, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, training=False):
        if isinstance(inputs, (tuple, list)):
            input_ids = inputs[0]
            past = inputs[1] if len(inputs) > 1 else past
            attention_mask = inputs[2] if len(inputs) > 2 else attention_mask
            token_type_ids = inputs[3] if len(inputs) > 3 else token_type_ids
            position_ids = inputs[4] if len(inputs) > 4 else position_ids
            head_mask = inputs[5] if len(inputs) > 5 else head_mask
            assert len(inputs) <= 6, "Too many inputs."
        elif isinstance(inputs, dict):
            input_ids = inputs.get('input_ids')
            past = inputs.get('past', past)
            attention_mask = inputs.get('attention_mask', attention_mask)
            token_type_ids = inputs.get('token_type_ids', token_type_ids)
            position_ids = inputs.get('position_ids', position_ids)
            head_mask = inputs.get('head_mask', head_mask)
            assert len(inputs) <= 6, "Too many inputs."
        else:
            input_ids = inputs

        if past is None:
            past_length = 0
            past = [None] * len(self.h)
        else:
            past_length = shape_list(past[0][0])[-2]
        if position_ids is None:
            position_ids = tf.range(past_length, shape_list(input_ids)[-1] + past_length, dtype=tf.int32)[tf.newaxis, :]

        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[:, tf.newaxis, tf.newaxis, :]

            # 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 = tf.cast(attention_mask, tf.float32)
            attention_mask = (1.0 - attention_mask) * -10000.0
        else:
            attention_mask = None

        # Prepare head mask if needed
        # 1.0 in head_mask indicate we keep the head
        # attention_probs has shape bsz x n_heads x N x N
        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
        if not head_mask is None:
            raise NotImplementedError
        else:
            head_mask = [None] * self.num_hidden_layers
            # head_mask = tf.constant([0] * self.num_hidden_layers)

        input_shape = shape_list(input_ids)
        input_ids = tf.reshape(input_ids, [-1, input_shape[-1]])
        position_ids = tf.reshape(position_ids, [-1, shape_list(position_ids)[-1]])

        inputs_embeds = self.wte(input_ids, mode='embedding')
        position_embeds = self.wpe(position_ids)
        if token_type_ids is not None:
            token_type_ids = tf.reshape(token_type_ids, [-1, shape_list(token_type_ids)[-1]])
            token_type_embeds = self.wte(token_type_ids, mode='embedding')
        else:
            token_type_embeds = 0
        hidden_states = inputs_embeds + position_embeds + token_type_embeds
        hidden_states = self.drop(hidden_states, training=training)

        output_shape = input_shape + [shape_list(hidden_states)[-1]]

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

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

            hidden_states, present = outputs[:2]
            presents = presents + (present,)

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

        hidden_states = self.ln_f(hidden_states)

        hidden_states = tf.reshape(hidden_states, output_shape)
        # Add last hidden state
        if self.output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        outputs = (hidden_states, presents)
        if self.output_hidden_states:
            outputs = outputs + (all_hidden_states,)
        if self.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] + shape_list(all_attentions[0])[-2:]
            all_attentions = tuple(tf.reshape(t, attention_output_shape) for t in all_attentions)
            outputs = outputs + (all_attentions,)
        return outputs  # last hidden state, presents, (all hidden_states), (attentions)


class TFGPT2PreTrainedModel(TFPreTrainedModel):
    """ An abstract class to handle weights initialization and
        a simple interface for dowloading and loading pretrained models.
    """
    config_class = GPT2Config
    pretrained_model_archive_map = TF_GPT2_PRETRAINED_MODEL_ARCHIVE_MAP
    load_pt_weights = load_gpt2_pt_weights_in_tf2
    base_model_prefix = "transformer"


GPT2_START_DOCSTRING = r"""    OpenAI GPT-2 model was proposed in
    `Language Models are Unsupervised Multitask Learners`_
    by Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever**.
    It's a causal (unidirectional) transformer pre-trained using  language modeling on a very large
    corpus of ~40 GB of text data.

    This model is a tf.keras.Model `tf.keras.Model`_ sub-class. Use it as a regular TF 2.0 Keras Model and
    refer to the TF 2.0 documentation for all matter related to general usage and behavior.

    .. _`Language Models are Unsupervised Multitask Learners`:
        https://openai.com/blog/better-language-models/

    .. _`tf.keras.Model`:
        https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/Model

    Important note on the model inputs:
        The inputs of the TF 2.0 models are slightly different from the PyTorch ones since
        TF 2.0 Keras doesn't accept named arguments with defaults values for input Tensor.
        More precisely, input Tensors are gathered in the first arguments of the model call function: `model(inputs)`.
        There are three possibilities to gather and feed the inputs to the model:

        - a single Tensor with input_ids only and nothing else: `model(inputs_ids)
        - a list of varying length with one or several input Tensors IN THE ORDER given in the docstring:
            `model([input_ids, attention_mask])` or `model([input_ids, attention_mask, token_type_ids])`
        - a dictionary with one or several input Tensors associaed to the input names given in the docstring:
            `model({'input_ids': input_ids, 'token_type_ids': token_type_ids})`

    Parameters:
        config (:class:`~pytorch_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:`~pytorch_transformers.PreTrainedModel.from_pretrained` method to load the model weights.
"""

GPT2_INPUTS_DOCSTRING = r"""    Inputs:
        **input_ids**: ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
            Indices of input sequence tokens in the vocabulary.
            GPT-2 is a model with absolute position embeddings so it's usually advised to pad the inputs on
            the right rather than the left.
            Indices can be obtained using :class:`pytorch_transformers.BPT2Tokenizer`.
            See :func:`pytorch_transformers.PreTrainedTokenizer.encode` and
            :func:`pytorch_transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
        **past**:
            list of ``torch.FloatTensor`` (one for each layer):
            that 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.
        **attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``:
            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.
        **token_type_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
            A parallel sequence of tokens (can be used to indicate various portions of the inputs).
            The embeddings from these tokens will be summed with the respective token embeddings.
            Indices are selected in the vocabulary (unlike BERT which has a specific vocabulary for segment indices).
        **position_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
            Indices of positions of each input sequence tokens in the position embeddings.
            Selected in the range ``[0, config.max_position_embeddings - 1]``.
        **head_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
            Mask to nullify selected heads of the self-attention modules.
            Mask values selected in ``[0, 1]``:
            ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
"""

@add_start_docstrings("The bare GPT2 Model transformer outputing raw hidden-states without any specific head on top.",
                      GPT2_START_DOCSTRING, GPT2_INPUTS_DOCSTRING)
class TFGPT2Model(TFGPT2PreTrainedModel):
    r"""
    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
        **last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``
            Sequence of hidden-states at the last layer of the model.
        **past**:
            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
            that 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**: (`optional`, returned when ``config.output_hidden_states=True``)
            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
            of shape ``(batch_size, sequence_length, hidden_size)``:
            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
            list of ``torch.FloatTensor`` (one for each layer) of shape ``(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 = GPT2Tokenizer.from_pretrained('gpt2')
        model = GPT2Model.from_pretrained('gpt2')
        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).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

    """
    def __init__(self, config, *inputs, **kwargs):
        super(TFGPT2Model, self).__init__(config, *inputs, **kwargs)
        self.transformer = TFGPT2MainLayer(config, name='transformer')

    def call(self, inputs, **kwargs):
        outputs = self.transformer(inputs, **kwargs)
        return outputs


@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, GPT2_INPUTS_DOCSTRING)
class TFGPT2LMHeadModel(TFGPT2PreTrainedModel):
    r"""
    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
        **prediction_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
        **past**:
            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
            that 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**: (`optional`, returned when ``config.output_hidden_states=True``)
            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
            of shape ``(batch_size, sequence_length, hidden_size)``:
            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
            list of ``torch.FloatTensor`` (one for each layer) of shape ``(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 pytorch_transformers import GPT2Tokenizer, GPT2LMHeadModel

        tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
        model = GPT2LMHeadModel.from_pretrained('gpt2')

        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
        outputs = model(input_ids)
        logits = outputs[:2]

    """
    def __init__(self, config, *inputs, **kwargs):
        super(TFGPT2LMHeadModel, self).__init__(config, *inputs, **kwargs)
        self.transformer = TFGPT2MainLayer(config, name='transformer')

    def call(self, inputs, **kwargs):
        transformer_outputs = self.transformer(inputs, **kwargs)
        hidden_states = transformer_outputs[0]

        lm_logits = self.transformer.wte(hidden_states, mode="linear")

        outputs = (lm_logits,) + transformer_outputs[1:]

        return outputs  # 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, GPT2_INPUTS_DOCSTRING)
class TFGPT2DoubleHeadsModel(TFGPT2PreTrainedModel):
    r"""
        **mc_token_ids**: (`optional`, default to index of the last token of the input) ``torch.LongTensor`` of shape ``(batch_size, num_choices)``:
            Index of the classification token in each input sequence.
            Selected in the range ``[0, input_ids.size(-1) - 1[``.

    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
        **lm_prediction_scores**: ``torch.FloatTensor`` of shape ``(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**: ``torch.FloatTensor`` of shape ``(batch_size, num_choices)``
            Prediction scores of the multiplechoice classification head (scores for each choice before SoftMax).
        **past**:
            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
            that 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**: (`optional`, returned when ``config.output_hidden_states=True``)
            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
            of shape ``(batch_size, sequence_length, hidden_size)``:
            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
            list of ``torch.FloatTensor`` (one for each layer) of shape ``(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 pytorch_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!)
        tokenizer.add_special_tokens({'cls_token': '[CLS]'})
        model.resize_token_embeddings(len(tokenizer))  # Update the model embeddings with the new vocabulary size
        print(tokenizer.cls_token_id, len(tokenizer))  # The newly token the last token of the vocabulary
        
        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]

    """
    def __init__(self, config, *inputs, **kwargs):
        super(TFGPT2DoubleHeadsModel, self).__init__(config, *inputs, **kwargs)
        self.transformer = TFGPT2MainLayer(config, name='transformer')
        self.multiple_choice_head = TFSequenceSummary(config, name='multiple_choice_head')

    def call(self, inputs, past=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, mc_token_ids=None, training=False):
        if isinstance(inputs, (tuple, list)):
            input_ids = inputs[0]
            past = inputs[1] if len(inputs) > 1 else past
            attention_mask = inputs[2] if len(inputs) > 2 else attention_mask
            token_type_ids = inputs[3] if len(inputs) > 3 else token_type_ids
            position_ids = inputs[4] if len(inputs) > 4 else position_ids
            head_mask = inputs[5] if len(inputs) > 5 else head_mask
            mc_token_ids = inputs[6] if len(inputs) > 6 else mc_token_ids
            assert len(inputs) <= 7, "Too many inputs."
        elif isinstance(inputs, dict):
            input_ids = inputs.get('input_ids')
            past = inputs.get('past', past)
            attention_mask = inputs.get('attention_mask', attention_mask)
            token_type_ids = inputs.get('token_type_ids', token_type_ids)
            position_ids = inputs.get('position_ids', position_ids)
            head_mask = inputs.get('head_mask', head_mask)
            mc_token_ids = inputs.get('mc_token_ids', mc_token_ids)
            assert len(inputs) <= 7, "Too many inputs."
        else:
            input_ids = inputs

        input_shapes = shape_list(input_ids)

        seq_length = input_shapes[-1]

        flat_input_ids = tf.reshape(input_ids, (-1, seq_length))
        flat_attention_mask = tf.reshape(attention_mask, (-1, seq_length)) if attention_mask is not None else None
        flat_token_type_ids = tf.reshape(token_type_ids, (-1, seq_length)) if token_type_ids is not None else None
        flat_position_ids = tf.reshape(position_ids, (-1, seq_length)) if position_ids is not None else None

        flat_inputs = [flat_input_ids, past, flat_attention_mask, flat_token_type_ids, flat_position_ids, head_mask]

        transformer_outputs = self.transformer(flat_inputs, training=training)
        hidden_states = transformer_outputs[0]

        hidden_states = tf.reshape(hidden_states, input_shapes + shape_list(hidden_states)[-1:])

        lm_logits = self.transformer.wte(hidden_states, mode="linear")
        mc_logits = self.multiple_choice_head([hidden_states, mc_token_ids], training=training)

        mc_logits = tf.squeeze(mc_logits, axis=-1)

        outputs = (lm_logits, mc_logits) + transformer_outputs[1:]

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