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 copy
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 .file_utils import cached_path, CONFIG_NAME, WEIGHTS_NAME
from .modeling import BertLayerNorm as LayerNorm
from .modeling_gpt2 import prune_conv1d_layer

logger = logging.getLogger(__name__)

PRETRAINED_MODEL_ARCHIVE_MAP = {"openai-gpt": "https://s3.amazonaws.com/models.huggingface.co/bert/openai-gpt-pytorch_model.bin"}
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, openai_checkpoint_folder_path):
    """ Load tf pre-trained weights in a pytorch model (from NumPy arrays here)
    """
    import re
    import numpy as np
    print("Loading weights...")
    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
        print("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(object):
    """Configuration class to store the configuration of a `OpenAIGPTModel`.
    """

    def __init__(
        self,
        vocab_size_or_config_json_file=40478,
        n_special=0,
        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
    ):
        """Constructs OpenAIGPTConfig.

        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)
        """
        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_special = n_special
            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
        else:
            raise ValueError(
                "First argument must be either a vocabulary size (int)"
                "or the path to a pretrained model config file (str)"
            )

    @property
    def total_tokens_embeddings(self):
        return self.vocab_size + self.n_special

    @classmethod
    def from_dict(cls, json_object):
        """Constructs a `OpenAIGPTConfig` from a Python dictionary of parameters."""
        config = OpenAIGPTConfig(vocab_size_or_config_json_file=-1)
        for key, value in json_object.items():
            config.__dict__[key] = value
        return config

    @classmethod
    def from_json_file(cls, json_file):
        """Constructs a `OpenAIGPTConfig` from a json file of parameters."""
        with open(json_file, "r", encoding="utf-8") as reader:
            text = reader.read()
        return cls.from_dict(json.loads(text))

    def __repr__(self):
        return str(self.to_json_string())

    def to_dict(self):
        """Serializes this instance to a Python dictionary."""
        output = copy.deepcopy(self.__dict__)
        return output

    def to_json_string(self):
        """Serializes this instance to a JSON string."""
        return json.dumps(self.to_dict(), indent=2, sort_keys=True) + "\n"

    def to_json_file(self, json_file_path):
        """ Save this instance to a json file."""
        with open(json_file_path, "w", encoding='utf-8') as writer:
            writer.write(self.to_json_string())


class Conv1D(nn.Module):
    def __init__(self, nf, rf, nx):
        super(Conv1D, self).__init__()
        self.rf = rf
        self.nf = nf
        if rf == 1:  # faster 1x1 conv
            w = torch.empty(nx, nf)
            nn.init.normal_(w, std=0.02)
            self.weight = Parameter(w)
            self.bias = Parameter(torch.zeros(nf))
        else:  # was used to train LM
            raise NotImplementedError

    def forward(self, x):
        if self.rf == 1:
            size_out = x.size()[:-1] + (self.nf,)
            x = torch.addmm(self.bias, x.view(-1, x.size(-1)), self.weight)
            x = x.view(*size_out)
        else:
            raise NotImplementedError
        return x


class Attention(nn.Module):
    def __init__(self, nx, n_ctx, config, scale=False, output_attentions=False, keep_multihead_output=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 = output_attentions
        self.keep_multihead_output = keep_multihead_output
        self.multihead_output = None

        self.c_attn = Conv1D(n_state * 3, 1, nx)
        self.c_proj = Conv1D(n_state, 1, 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

        if self.output_attentions:
            return w, torch.matmul(w, v)
        return torch.matmul(w, v)

    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)

        a = self._attn(query, key, value, head_mask)
        if self.keep_multihead_output:
            self.multihead_output = a
            self.multihead_output.retain_grad()

        if self.output_attentions:
            attentions, a = a
        a = self.merge_heads(a)
        a = self.c_proj(a)
        a = self.resid_dropout(a)
        if self.output_attentions:
            return attentions, a
        return a


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, 1, nx)
        self.c_proj = Conv1D(nx, 1, 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, output_attentions=False, keep_multihead_output=False):
        super(Block, self).__init__()
        nx = config.n_embd
        self.output_attentions = output_attentions
        self.attn = Attention(nx, n_ctx, config, scale, output_attentions, keep_multihead_output)
        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):
        a = self.attn(x, head_mask=head_mask)
        if self.output_attentions:
            attentions, a = a
        n = self.ln_1(x + a)
        m = self.mlp(n)
        h = self.ln_2(n + m)
        if self.output_attentions:
            return attentions, h
        return h


class OpenAIGPTLMHead(nn.Module):
    """ Language Model Head for the transformer """

    def __init__(self, model_embeddings_weights, config):
        super(OpenAIGPTLMHead, self).__init__()
        self.n_embd = config.n_embd
        self.vocab_size = config.vocab_size
        self.predict_special_tokens = config.predict_special_tokens
        embed_shape = model_embeddings_weights.shape
        self.decoder = nn.Linear(embed_shape[1], embed_shape[0], bias=False)
        self.set_embeddings_weights(model_embeddings_weights)

    def set_embeddings_weights(self, model_embeddings_weights, predict_special_tokens=True):
        self.predict_special_tokens = predict_special_tokens
        embed_shape = model_embeddings_weights.shape
        self.decoder.weight = model_embeddings_weights  # Tied weights

    def forward(self, hidden_state):
        lm_logits = self.decoder(hidden_state)
        if not self.predict_special_tokens:
            lm_logits = lm_logits[..., :self.vocab_size]
        return lm_logits


class OpenAIGPTMultipleChoiceHead(nn.Module):
    """ Classifier Head for the transformer """

    def __init__(self, config):
        super(OpenAIGPTMultipleChoiceHead, self).__init__()
        self.n_embd = config.n_embd
        self.dropout = nn.Dropout2d(config.resid_pdrop)  # To reproduce the noise_shape parameter of TF implementation
        self.linear = nn.Linear(config.n_embd, 1)

        nn.init.normal_(self.linear.weight, std=0.02)
        nn.init.normal_(self.linear.bias, 0)

    def forward(self, hidden_states, mc_token_ids):
        # Classification logits
        # hidden_state (bsz, num_choices, seq_length, hidden_size)
        # mc_token_ids (bsz, num_choices)
        mc_token_ids = mc_token_ids.unsqueeze(-1).unsqueeze(-1).expand(-1, -1, -1, hidden_states.size(-1))
        # (bsz, num_choices, 1, hidden_size)
        multiple_choice_h = hidden_states.gather(2, mc_token_ids).squeeze(2)
        # (bsz, num_choices, hidden_size)
        multiple_choice_h = self.dropout(multiple_choice_h.transpose(1, 2)).transpose(1, 2)
        multiple_choice_logits = self.linear(multiple_choice_h).squeeze(-1)
        # (bsz, num_choices)
        return multiple_choice_logits


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

    def __init__(self, config, *inputs, **kwargs):
        super(OpenAIGPTPreTrainedModel, self).__init__()
        if not isinstance(config, OpenAIGPTConfig):
            raise ValueError(
                "Parameter config in `{}(config)` should be an instance of class `OpenAIGPTConfig`. "
                "To create a model from a pretrained model use "
                "`model = {}.from_pretrained(PRETRAINED_MODEL_NAME)`".format(
                    self.__class__.__name__, self.__class__.__name__
                )
            )
        self.config = config

    def init_weights(self, module):
        """ Initialize the weights.
        """
        if isinstance(module, (nn.Linear, nn.Embedding)):
            # 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)
        elif isinstance(module, LayerNorm):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)
        if isinstance(module, nn.Linear) and module.bias is not None:
            module.bias.data.zero_()

    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path, num_special_tokens=None, *inputs, **kwargs):
        """
        Instantiate a OpenAIGPTPreTrainedModel from a pre-trained model file or a pytorch state dict.
        Download and cache the pre-trained model file if needed.

        Params:
            pretrained_model_name_or_path: either:
                - a str with the name of a pre-trained model to load selected in the list of:
                    . `openai-gpt`
                - a path or url to a pretrained model archive containing:
                    . `openai_gpt_config.json` a configuration file for the model
                    . `pytorch_model.bin` a PyTorch dump of a OpenAIGPTModel instance
                - a path or url to a pretrained model archive containing:
                    . `openai-gpt-config.json` a configuration file for the model
                    . a series of NumPy files containing OpenAI TensorFlow trained weights
            from_tf: should we load the weights from a locally saved TensorFlow checkpoint
            cache_dir: an optional path to a folder in which the pre-trained models will be cached.
            state_dict: an optional state dictionnary (collections.OrderedDict object) to use instead of pre-trained models
            *inputs, **kwargs: additional input for the specific OpenAI-GPT class
        """
        state_dict = kwargs.get('state_dict', None)
        kwargs.pop('state_dict', None)
        cache_dir = kwargs.get('cache_dir', None)
        kwargs.pop('cache_dir', None)
        from_tf = kwargs.get('from_tf', False)
        kwargs.pop('from_tf', None)

        if pretrained_model_name_or_path in PRETRAINED_MODEL_ARCHIVE_MAP:
            archive_file = PRETRAINED_MODEL_ARCHIVE_MAP[pretrained_model_name_or_path]
            config_file = PRETRAINED_CONFIG_ARCHIVE_MAP[pretrained_model_name_or_path]
        else:
            archive_file = os.path.join(pretrained_model_name_or_path, WEIGHTS_NAME)
            config_file = os.path.join(pretrained_model_name_or_path, CONFIG_NAME)
        # redirect to the cache, if necessary
        try:
            resolved_archive_file = cached_path(archive_file, cache_dir=cache_dir)
        except EnvironmentError:
            if pretrained_model_name_or_path in PRETRAINED_MODEL_ARCHIVE_MAP:
                logger.error(
                    "Couldn't reach server at '{}' to download pretrained weights.".format(
                        archive_file))
            else:
                logger.error(
                    "Model name '{}' was not found in model name list ({}). "
                    "We assumed '{}' was a path or url but couldn't find file {} "
                    "at this path or url.".format(
                        pretrained_model_name_or_path, ", ".join(PRETRAINED_MODEL_ARCHIVE_MAP.keys()), pretrained_model_name_or_path,
                        archive_file
                    )
                )
            return None
        try:
            resolved_config_file = cached_path(config_file, cache_dir=cache_dir)
        except EnvironmentError:
            if pretrained_model_name_or_path in PRETRAINED_CONFIG_ARCHIVE_MAP:
                logger.error(
                    "Couldn't reach server at '{}' to download pretrained model configuration file.".format(
                        config_file))
            else:
                logger.error(
                    "Model name '{}' was not found in model name list ({}). "
                    "We assumed '{}' was a path or url but couldn't find file {} "
                    "at this path or url.".format(
                        pretrained_model_name_or_path, ", ".join(PRETRAINED_CONFIG_ARCHIVE_MAP.keys()), pretrained_model_name_or_path,
                        config_file
                    )
                )
            return None
        if resolved_archive_file == archive_file and resolved_config_file == config_file:
            logger.info("loading weights file {}".format(archive_file))
            logger.info("loading configuration file {}".format(config_file))
        else:
            logger.info("loading weights file {} from cache at {}".format(
                archive_file, resolved_archive_file))
            logger.info("loading configuration file {} from cache at {}".format(
                config_file, resolved_config_file))
        # Load config
        config = OpenAIGPTConfig.from_json_file(resolved_config_file)
        logger.info("Model config {}".format(config))
        # Instantiate model.
        model = cls(config, *inputs, **kwargs)
        if state_dict is None and not from_tf:
            state_dict = torch.load(resolved_archive_file, map_location='cpu')
        if from_tf:
            # Directly load from a TensorFlow checkpoint (stored as NumPy array)
            return load_tf_weights_in_openai_gpt(model, resolved_archive_file)

        old_keys = []
        new_keys = []
        for key in state_dict.keys():
            new_key = None
            if key.endswith(".g"):
                new_key = key[:-2] + ".weight"
            elif key.endswith(".b"):
                new_key = key[:-2] + ".bias"
            elif key.endswith(".w"):
                new_key = key[:-2] + ".weight"
            if new_key:
                old_keys.append(key)
                new_keys.append(new_key)
        for old_key, new_key in zip(old_keys, new_keys):
            state_dict[new_key] = state_dict.pop(old_key)

        missing_keys = []
        unexpected_keys = []
        error_msgs = []
        # copy state_dict so _load_from_state_dict can modify it
        metadata = getattr(state_dict, "_metadata", None)
        state_dict = state_dict.copy()
        if metadata is not None:
            state_dict._metadata = metadata

        def load(module, prefix=""):
            local_metadata = {} if metadata is None else metadata.get(prefix[:-1], {})
            module._load_from_state_dict(
                state_dict, prefix, local_metadata, True, missing_keys, unexpected_keys, error_msgs
            )
            for name, child in module._modules.items():
                if child is not None:
                    load(child, prefix + name + ".")

        start_model = model
        if hasattr(model, "transformer") and all(not s.startswith('transformer.') for s in state_dict.keys()):
            start_model = model.transformer
        load(start_model, prefix="")

        if len(missing_keys) > 0:
            logger.info(
                "Weights of {} not initialized from pretrained model: {}".format(model.__class__.__name__, missing_keys)
            )
        if len(unexpected_keys) > 0:
            logger.info(
                "Weights from pretrained model not used in {}: {}".format(model.__class__.__name__, unexpected_keys)
            )
        if len(error_msgs) > 0:
            raise RuntimeError(
                "Error(s) in loading state_dict for {}:\n\t{}".format(model.__class__.__name__, "\n\t".join(error_msgs))
            )

        # Add additional embeddings for special tokens if needed
        # This step also make sure we are still sharing the output and input embeddings after loading weights
        model.set_num_special_tokens(num_special_tokens if num_special_tokens is not None else config.n_special)
        return model


class OpenAIGPTModel(OpenAIGPTPreTrainedModel):
    """OpenAI GPT model ("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 controled using the `set_num_special_tokens(num_special_tokens)` function.

    The embeddings are ordered as follow in the token embeddings matrice:
        [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 associate indices to index the embeddings.

    Params:
        `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

    Inputs:
        `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.

    Outputs:
        `hidden_states`: a list of all the encoded-hidden-states in the model (length of the list: 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 usage:
    ```python
    # Already been converted into BPE token ids
    input_ids = torch.LongTensor([[31, 51, 99], [15, 5, 0]])

    config = modeling_openai.OpenAIGPTConfig()

    model = modeling_openai.OpenAIGPTModel(config)
    hidden_states = model(input_ids)
    ```
    """

    def __init__(self, config, output_attentions=False, keep_multihead_output=False):
        super(OpenAIGPTModel, self).__init__(config)
        self.output_attentions = output_attentions
        self.tokens_embed = nn.Embedding(config.total_tokens_embeddings, config.n_embd)
        self.positions_embed = nn.Embedding(config.n_positions, config.n_embd)
        self.drop = nn.Dropout(config.embd_pdrop)
        block = Block(config.n_ctx, config, scale=True, output_attentions=output_attentions,
                                                        keep_multihead_output=keep_multihead_output)
        self.h = nn.ModuleList([copy.deepcopy(block) for _ in range(config.n_layer)])

        self.apply(self.init_weights)

    def set_num_special_tokens(self, num_special_tokens):
        " Update input embeddings with new embedding matrice if needed "
        if self.config.n_special == num_special_tokens:
            return
        # Update config
        self.config.n_special = num_special_tokens
        # Build new embeddings and initialize all new embeddings (in particular the special tokens)
        old_embed = self.tokens_embed
        self.tokens_embed = nn.Embedding(self.config.total_tokens_embeddings, self.config.n_embd)
        self.tokens_embed.to(old_embed.weight.device)
        self.init_weights(self.tokens_embed)
        # Copy word embeddings from the previous weights
        self.tokens_embed.weight.data[:self.config.vocab_size, :] = old_embed.weight.data[:self.config.vocab_size, :]

    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 get_multihead_outputs(self):
        """ Gather all multi-head outputs.
            Return: list (layers) of multihead module outputs with gradients
        """
        return [h.attn.multihead_output for h in self.h]

    def forward(self, input_ids, position_ids=None, token_type_ids=None, head_mask=None):
        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_as(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 = [hidden_states.view(*output_shape)]
        for i, block in enumerate(self.h):
            outputs = block(hidden_states, head_mask[i])
            if self.output_attentions:
                attentions, hidden_states = outputs
                all_attentions.append(attentions)
            else:
                hidden_states = outputs
            all_hidden_states.append(hidden_states.view(*output_shape))

        if self.output_attentions:
            return all_attentions, all_hidden_states
        return all_hidden_states


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 controled using the `set_num_special_tokens(num_special_tokens)` function.

    The embeddings are ordered as follow in the token embeddings matrice:
        [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 associate indices to index the embeddings.

    Params:
        `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

    Inputs:
        `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.

    Outputs:
        if `lm_labels` is not `None`:
            Outputs the language modeling loss.
        else:
            `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] were d_1 ... d_n are the dimension of input_ids)

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

    config = modeling_openai.OpenAIGPTConfig()

    model = modeling_openai.OpenAIGPTLMHeadModel(config)
    lm_logits = model(input_ids)
    ```
    """

    def __init__(self, config, output_attentions=False, keep_multihead_output=False):
        super(OpenAIGPTLMHeadModel, self).__init__(config)
        self.transformer = OpenAIGPTModel(config, output_attentions=output_attentions,
                                             keep_multihead_output=keep_multihead_output)
        self.lm_head = OpenAIGPTLMHead(self.transformer.tokens_embed.weight, config)
        self.apply(self.init_weights)

    def set_num_special_tokens(self, num_special_tokens, predict_special_tokens=True):
        """ Update input and output embeddings with new embedding matrice
            Make sure we are sharing the embeddings
        """
        self.config.predict_special_tokens = self.transformer.config.predict_special_tokens = predict_special_tokens
        self.transformer.set_num_special_tokens(num_special_tokens)
        self.lm_head.set_embeddings_weights(self.transformer.tokens_embed.weight, predict_special_tokens=predict_special_tokens)

    def forward(self, input_ids, position_ids=None, token_type_ids=None, lm_labels=None, head_mask=None):
        hidden_states = self.transformer(input_ids, position_ids, token_type_ids, head_mask)
        if self.transformer.output_attentions:
            all_attentions, hidden_states = hidden_states
        hidden_states = hidden_states[-1]

        lm_logits = self.lm_head(hidden_states)
        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))
            return loss
        if self.transformer.output_attentions:
            return all_attentions, lm_logits
        return lm_logits


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 controled using the `set_num_special_tokens(num_special_tokens)` function.

    The embeddings are ordered as follow in the token embeddings matrice:
        [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 associate indices to index the embeddings.

    Params:
        `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

    Inputs:
        `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.

    Outputs:
        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.
        else: a tuple with
            `lm_logits`: 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 usage:
    ```python
    # 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)

    config = modeling_openai.OpenAIGPTConfig()

    model = modeling_openai.OpenAIGPTDoubleHeadsModel(config)
    lm_logits, multiple_choice_logits = model(input_ids, mc_token_ids)
    ```
    """

    def __init__(self, config, output_attentions=False, keep_multihead_output=False):
        super(OpenAIGPTDoubleHeadsModel, self).__init__(config)
        self.transformer = OpenAIGPTModel(config, output_attentions=output_attentions,
                                             keep_multihead_output=keep_multihead_output)
        self.lm_head = OpenAIGPTLMHead(self.transformer.tokens_embed.weight, config)
        self.multiple_choice_head = OpenAIGPTMultipleChoiceHead(config)
        self.apply(self.init_weights)

    def set_num_special_tokens(self, num_special_tokens, predict_special_tokens=True):
        """ Update input and output embeddings with new embedding matrice
            Make sure we are sharing the embeddings
        """
        self.config.predict_special_tokens = self.transformer.config.predict_special_tokens = predict_special_tokens
        self.transformer.set_num_special_tokens(num_special_tokens)
        self.lm_head.set_embeddings_weights(self.transformer.tokens_embed.weight, predict_special_tokens=predict_special_tokens)

    def forward(self, input_ids, mc_token_ids, lm_labels=None, mc_labels=None, token_type_ids=None,
                position_ids=None, head_mask=None):
        hidden_states = self.transformer(input_ids, position_ids, token_type_ids, head_mask)
        if self.transformer.output_attentions:
            all_attentions, hidden_states = hidden_states
        hidden_states = hidden_states[-1]

        lm_logits = self.lm_head(hidden_states)
        mc_logits = self.multiple_choice_head(hidden_states, mc_token_ids)
        losses = []
        if lm_labels is not None:
            shift_logits = lm_logits[..., :-1, :].contiguous()
            shift_labels = lm_labels[..., 1:].contiguous()
            loss_fct = CrossEntropyLoss(ignore_index=-1)
            losses.append(loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)))
        if mc_labels is not None:
            loss_fct = CrossEntropyLoss()
            losses.append(loss_fct(mc_logits.view(-1, mc_logits.size(-1)), mc_labels.view(-1)))
        if losses:
            return losses
        if self.transformer.output_attentions:
            return all_attentions, lm_logits, mc_logits
        return lm_logits, mc_logits