modeling_xlm.py

# coding=utf-8
# Copyright 2019-present, Facebook, Inc and the HuggingFace Inc. team.
#
# 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 XLM model.
"""
from __future__ import absolute_import, division, print_function, unicode_literals

import logging
import math

import itertools
import numpy as np

import torch
from torch import nn
from torch.nn import functional as F
from torch.nn import CrossEntropyLoss, MSELoss

from .modeling_utils import PreTrainedModel, prune_linear_layer, SequenceSummary, SQuADHead
from .configuration_xlm import XLMConfig
from .file_utils import add_start_docstrings

logger = logging.getLogger(__name__)

XLM_PRETRAINED_MODEL_ARCHIVE_MAP = {
    'xlm-mlm-en-2048': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-mlm-en-2048-pytorch_model.bin",
    'xlm-mlm-ende-1024': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-mlm-ende-1024-pytorch_model.bin",
    'xlm-mlm-enfr-1024': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-mlm-enfr-1024-pytorch_model.bin",
    'xlm-mlm-enro-1024': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-mlm-enro-1024-pytorch_model.bin",
    'xlm-mlm-tlm-xnli15-1024': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-mlm-tlm-xnli15-1024-pytorch_model.bin",
    'xlm-mlm-xnli15-1024': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-mlm-xnli15-1024-pytorch_model.bin",
    'xlm-clm-enfr-1024': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-clm-enfr-1024-pytorch_model.bin",
    'xlm-clm-ende-1024': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-clm-ende-1024-pytorch_model.bin",
    'xlm-mlm-17-1280': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-mlm-17-1280-pytorch_model.bin",
    'xlm-mlm-100-1280': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-mlm-100-1280-pytorch_model.bin",
}


def create_sinusoidal_embeddings(n_pos, dim, out):
    position_enc = np.array([
        [pos / np.power(10000, 2 * (j // 2) / dim) for j in range(dim)]
        for pos in range(n_pos)
    ])
    out[:, 0::2] = torch.FloatTensor(np.sin(position_enc[:, 0::2]))
    out[:, 1::2] = torch.FloatTensor(np.cos(position_enc[:, 1::2]))
    out.detach_()
    out.requires_grad = False


def gelu(x):
    """
    GELU activation
    https://arxiv.org/abs/1606.08415
    https://github.com/huggingface/pytorch-openai-transformer-lm/blob/master/model_pytorch.py#L14
    https://github.com/huggingface/pytorch-transformers/blob/master/modeling.py
    """
    # return 0.5 * x * (1 + torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3))))
    return 0.5 * x * (1.0 + torch.erf(x / math.sqrt(2.0)))


def get_masks(slen, lengths, causal, padding_mask=None):
    """
    Generate hidden states mask, and optionally an attention mask.
    """
    bs = lengths.size(0)
    if padding_mask is not None:
        mask = padding_mask
    else:
        assert lengths.max().item() <= slen
        alen = torch.arange(slen, dtype=torch.long, device=lengths.device)
        mask = alen < lengths[:, None]

    # attention mask is the same as mask, or triangular inferior attention (causal)
    if causal:
        attn_mask = alen[None, None, :].repeat(bs, slen, 1) <= alen[None, :, None]
    else:
        attn_mask = mask

    # sanity check
    assert mask.size() == (bs, slen)
    assert causal is False or attn_mask.size() == (bs, slen, slen)

    return mask, attn_mask


class MultiHeadAttention(nn.Module):

    NEW_ID = itertools.count()

    def __init__(self, n_heads, dim, config):
        super(MultiHeadAttention, self).__init__()
        self.layer_id = next(MultiHeadAttention.NEW_ID)
        self.output_attentions = config.output_attentions
        self.dim = dim
        self.n_heads = n_heads
        self.dropout = config.attention_dropout
        assert self.dim % self.n_heads == 0

        self.q_lin = nn.Linear(dim, dim)
        self.k_lin = nn.Linear(dim, dim)
        self.v_lin = nn.Linear(dim, dim)
        self.out_lin = nn.Linear(dim, dim)
        self.pruned_heads = set()

    def prune_heads(self, heads):
        attention_head_size = self.dim // self.n_heads
        if len(heads) == 0:
            return
        mask = torch.ones(self.n_heads, attention_head_size)
        heads = set(heads) - self.pruned_heads
        for head in heads:
            head -= sum(1 if h < head else 0 for h in self.pruned_heads)
            mask[head] = 0
        mask = mask.view(-1).contiguous().eq(1)
        index = torch.arange(len(mask))[mask].long()
        # Prune linear layers
        self.q_lin = prune_linear_layer(self.q_lin, index)
        self.k_lin = prune_linear_layer(self.k_lin, index)
        self.v_lin = prune_linear_layer(self.v_lin, index)
        self.out_lin = prune_linear_layer(self.out_lin, index, dim=1)
        # Update hyper params
        self.n_heads = self.n_heads - len(heads)
        self.dim = attention_head_size * self.n_heads
        self.pruned_heads = self.pruned_heads.union(heads)

    def forward(self, input, mask, kv=None, cache=None, head_mask=None):
        """
        Self-attention (if kv is None) or attention over source sentence (provided by kv).
        """
        # Input is (bs, qlen, dim)
        # Mask is (bs, klen) (non-causal) or (bs, klen, klen)
        bs, qlen, dim = input.size()
        if kv is None:
            klen = qlen if cache is None else cache['slen'] + qlen
        else:
            klen = kv.size(1)
        # assert dim == self.dim, 'Dimensions do not match: %s input vs %s configured' % (dim, self.dim)
        n_heads = self.n_heads
        dim_per_head = self.dim // n_heads
        mask_reshape = (bs, 1, qlen, klen) if mask.dim() == 3 else (bs, 1, 1, klen)

        def shape(x):
            """  projection """
            return x.view(bs, -1, self.n_heads, dim_per_head).transpose(1, 2)

        def unshape(x):
            """  compute context """
            return x.transpose(1, 2).contiguous().view(bs, -1, self.n_heads * dim_per_head)

        q = shape(self.q_lin(input))                                          # (bs, n_heads, qlen, dim_per_head)
        if kv is None:
            k = shape(self.k_lin(input))                                      # (bs, n_heads, qlen, dim_per_head)
            v = shape(self.v_lin(input))                                      # (bs, n_heads, qlen, dim_per_head)
        elif cache is None or self.layer_id not in cache:
            k = v = kv
            k = shape(self.k_lin(k))                                          # (bs, n_heads, qlen, dim_per_head)
            v = shape(self.v_lin(v))                                          # (bs, n_heads, qlen, dim_per_head)

        if cache is not None:
            if self.layer_id in cache:
                if kv is None:
                    k_, v_ = cache[self.layer_id]
                    k = torch.cat([k_, k], dim=2)                             # (bs, n_heads, klen, dim_per_head)
                    v = torch.cat([v_, v], dim=2)                             # (bs, n_heads, klen, dim_per_head)
                else:
                    k, v = cache[self.layer_id]
            cache[self.layer_id] = (k, v)

        q = q / math.sqrt(dim_per_head)                                       # (bs, n_heads, qlen, dim_per_head)
        scores = torch.matmul(q, k.transpose(2, 3))                           # (bs, n_heads, qlen, klen)
        mask = (mask == 0).view(mask_reshape).expand_as(scores)               # (bs, n_heads, qlen, klen)
        scores.masked_fill_(mask, -float('inf'))                              # (bs, n_heads, qlen, klen)

        weights = F.softmax(scores.float(), dim=-1).type_as(scores)           # (bs, n_heads, qlen, klen)
        weights = F.dropout(weights, p=self.dropout, training=self.training)  # (bs, n_heads, qlen, klen)

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

        context = torch.matmul(weights, v)                                    # (bs, n_heads, qlen, dim_per_head)
        context = unshape(context)                                            # (bs, qlen, dim)

        outputs = (self.out_lin(context),)
        if self.output_attentions:
            outputs = outputs + (weights,)
        return outputs


class TransformerFFN(nn.Module):

    def __init__(self, in_dim, dim_hidden, out_dim, config):
        super(TransformerFFN, self).__init__()
        self.dropout = config.dropout
        self.lin1 = nn.Linear(in_dim, dim_hidden)
        self.lin2 = nn.Linear(dim_hidden, out_dim)
        self.act = gelu if config.gelu_activation else F.relu

    def forward(self, input):
        x = self.lin1(input)
        x = self.act(x)
        x = self.lin2(x)
        x = F.dropout(x, p=self.dropout, training=self.training)
        return x


class XLMPreTrainedModel(PreTrainedModel):
    """ An abstract class to handle weights initialization and
        a simple interface for dowloading and loading pretrained models.
    """
    config_class = XLMConfig
    pretrained_model_archive_map = XLM_PRETRAINED_MODEL_ARCHIVE_MAP
    load_tf_weights = None
    base_model_prefix = "transformer"

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

    def _init_weights(self, module):
        """ Initialize the weights. """
        if isinstance(module, nn.Embedding):
            if self.config is not None and self.config.embed_init_std is not None:
                nn.init.normal_(module.weight, mean=0, std=self.config.embed_init_std)
        if isinstance(module, nn.Linear):
            if self.config is not None and self.config.init_std is not None:
                nn.init.normal_(module.weight, mean=0, std=self.config.init_std)
                if hasattr(module, 'bias') and module.bias is not None:
                    nn.init.constant_(module.bias, 0.)
        if isinstance(module, nn.LayerNorm):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)


XLM_START_DOCSTRING = r"""    The XLM model was proposed in
    `Cross-lingual Language Model Pretraining`_
    by Guillaume Lample*, Alexis Conneau*. It's a transformer pre-trained using one of the following objectives:

        - a causal language modeling (CLM) objective (next token prediction),
        - a masked language modeling (MLM) objective (Bert-like), or
        - a Translation Language Modeling (TLM) object (extension of Bert's MLM to multiple language inputs)

    Original code can be found `here`_.

    This model is a PyTorch `torch.nn.Module`_ sub-class. Use it as a regular PyTorch Module and
    refer to the PyTorch documentation for all matter related to general usage and behavior.

    .. _`Cross-lingual Language Model Pretraining`:
        https://arxiv.org/abs/1901.07291

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

    .. _`here`:
        https://github.com/facebookresearch/XLM

    Parameters:
        config (:class:`~pytorch_transformers.XLMConfig`): 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.
"""

XLM_INPUTS_DOCSTRING = r"""
    Inputs:
        **input_ids**: ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
            Indices of input sequence tokens in the vocabulary.

            XLM 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.XLMTokenizer`.
            See :func:`pytorch_transformers.PreTrainedTokenizer.encode` and
            :func:`pytorch_transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
        **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]``.
        **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).
        **langs**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
            A parallel sequence of tokens to be used to indicate the language of each token in the input.
            Indices are languages ids which can be obtained from the language names by using two conversion mappings
            provided in the configuration of the model (only provided for multilingual models).
            More precisely, the `language name -> language id` mapping is in `model.config.lang2id` (dict str -> int) and
            the `language id -> language name` mapping is `model.config.id2lang` (dict int -> str).
        **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.
        **lengths**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
            Length of each sentence that can be used to avoid performing attention on padding token indices.
            You can also use `attention_mask` for the same result (see above), kept here for compatbility.
            Indices selected in ``[0, ..., input_ids.size(-1)]``:
        **cache**:
            dictionary with ``torch.FloatTensor`` that contains pre-computed
            hidden-states (key and values in the attention blocks) as computed by the model
            (see `cache` output below). Can be used to speed up sequential decoding.
            The dictionary object will be modified in-place during the forward pass to add newly computed hidden-states.
        **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 XLM Model transformer outputing raw hidden-states without any specific head on top.",
                      XLM_START_DOCSTRING, XLM_INPUTS_DOCSTRING)
class XLMModel(XLMPreTrainedModel):
    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.
        **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 = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
        model = XLMModel.from_pretrained('xlm-mlm-en-2048')
        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

    """
    ATTRIBUTES = ['encoder', 'eos_index', 'pad_index',  # 'with_output', 
                  'n_langs', 'use_lang_emb', 'n_words', 'dim', 'n_layers', 'n_heads', 
                  'hidden_dim', 'dropout', 'attention_dropout', 'asm',
                  'asm_cutoffs', 'asm_div_value']

    def __init__(self, config):  #, dico, is_encoder, with_output):
        super(XLMModel, self).__init__(config)
        self.output_attentions = config.output_attentions
        self.output_hidden_states = config.output_hidden_states

        # encoder / decoder, output layer
        self.is_encoder = config.is_encoder
        self.is_decoder = not config.is_encoder
        if self.is_decoder:
            raise NotImplementedError("Currently XLM can only be used as an encoder")
        # self.with_output = with_output
        self.causal = config.causal

        # dictionary / languages
        self.n_langs = config.n_langs
        self.use_lang_emb = config.use_lang_emb
        self.n_words = config.n_words
        self.eos_index = config.eos_index
        self.pad_index = config.pad_index
        # self.dico = dico
        # self.id2lang = config.id2lang
        # self.lang2id = config.lang2id
        # assert len(self.dico) == self.n_words
        # assert len(self.id2lang) == len(self.lang2id) == self.n_langs

        # model parameters
        self.dim = config.emb_dim       # 512 by default
        self.hidden_dim = self.dim * 4  # 2048 by default
        self.n_heads = config.n_heads   # 8 by default
        self.n_layers = config.n_layers
        self.dropout = config.dropout
        self.attention_dropout = config.attention_dropout
        assert self.dim % self.n_heads == 0, 'transformer dim must be a multiple of n_heads'

        # embeddings
        self.position_embeddings = nn.Embedding(config.max_position_embeddings, self.dim)
        if config.sinusoidal_embeddings:
            create_sinusoidal_embeddings(config.max_position_embeddings, self.dim, out=self.position_embeddings.weight)
        if config.n_langs > 1 and config.use_lang_emb:
            self.lang_embeddings = nn.Embedding(self.n_langs, self.dim)
        self.embeddings = nn.Embedding(self.n_words, self.dim, padding_idx=self.pad_index)
        self.layer_norm_emb = nn.LayerNorm(self.dim, eps=config.layer_norm_eps)

        # transformer layers
        self.attentions = nn.ModuleList()
        self.layer_norm1 = nn.ModuleList()
        self.ffns = nn.ModuleList()
        self.layer_norm2 = nn.ModuleList()
        # if self.is_decoder:
        #     self.layer_norm15 = nn.ModuleList()
        #     self.encoder_attn = nn.ModuleList()

        for _ in range(self.n_layers):
            self.attentions.append(MultiHeadAttention(self.n_heads, self.dim, config=config))
            self.layer_norm1.append(nn.LayerNorm(self.dim, eps=config.layer_norm_eps))
            # if self.is_decoder:
            #     self.layer_norm15.append(nn.LayerNorm(self.dim, eps=config.layer_norm_eps))
            #     self.encoder_attn.append(MultiHeadAttention(self.n_heads, self.dim, dropout=self.attention_dropout))
            self.ffns.append(TransformerFFN(self.dim, self.hidden_dim, self.dim, config=config))
            self.layer_norm2.append(nn.LayerNorm(self.dim, eps=config.layer_norm_eps))

        if hasattr(config, "pruned_heads"):
            pruned_heads = config.pruned_heads.copy().items()
            config.pruned_heads = {}
            for layer, heads in pruned_heads:
                if self.attentions[int(layer)].n_heads == config.n_heads:
                    self.prune_heads({int(layer): list(map(int, heads))})

        self.init_weights()

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

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

    def forward(self, input_ids, lengths=None, position_ids=None, langs=None,
                token_type_ids=None, attention_mask=None, cache=None, head_mask=None):  # src_enc=None, src_len=None, 
        if lengths is None:
            lengths = (input_ids != self.pad_index).sum(dim=1).long()
        # mask = input_ids != self.pad_index

        # check inputs
        bs, slen = input_ids.size()
        assert lengths.size(0) == bs
        assert lengths.max().item() <= slen
        # input_ids = input_ids.transpose(0, 1)  # batch size as dimension 0
        # assert (src_enc is None) == (src_len is None)
        # if src_enc is not None:
        #     assert self.is_decoder
        #     assert src_enc.size(0) == bs

        # generate masks
        mask, attn_mask = get_masks(slen, lengths, self.causal, padding_mask=attention_mask)
        # if self.is_decoder and src_enc is not None:
        #     src_mask = torch.arange(src_len.max(), dtype=torch.long, device=lengths.device) < src_len[:, None]

        # position_ids
        if position_ids is None:
            position_ids = input_ids.new((slen,)).long()
            position_ids = torch.arange(slen, out=position_ids).unsqueeze(0)
        else:
            assert position_ids.size() == (bs, slen)  # (slen, bs)
            # position_ids = position_ids.transpose(0, 1)

        # langs
        if langs is not None:
            assert langs.size() == (bs, slen)  # (slen, bs)
            # langs = langs.transpose(0, 1)

        # 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 qlen x klen]
        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.n_layers, -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.n_layers

        # do not recompute cached elements
        if cache is not None:
            _slen = slen - cache['slen']
            input_ids = input_ids[:, -_slen:]
            position_ids = position_ids[:, -_slen:]
            if langs is not None:
                langs = langs[:, -_slen:]
            mask = mask[:, -_slen:]
            attn_mask = attn_mask[:, -_slen:]

        # embeddings
        tensor = self.embeddings(input_ids)
        tensor = tensor + self.position_embeddings(position_ids).expand_as(tensor)
        if langs is not None and self.use_lang_emb:
            tensor = tensor + self.lang_embeddings(langs)
        if token_type_ids is not None:
            tensor = tensor + self.embeddings(token_type_ids)
        tensor = self.layer_norm_emb(tensor)
        tensor = F.dropout(tensor, p=self.dropout, training=self.training)
        tensor *= mask.unsqueeze(-1).to(tensor.dtype)

        # transformer layers
        hidden_states = ()
        attentions = ()
        for i in range(self.n_layers):
            if self.output_hidden_states:
                hidden_states = hidden_states + (tensor,)

            # self attention
            attn_outputs = self.attentions[i](tensor, attn_mask, cache=cache, head_mask=head_mask[i])
            attn = attn_outputs[0]
            if self.output_attentions:
                attentions = attentions + (attn_outputs[1],)
            attn = F.dropout(attn, p=self.dropout, training=self.training)
            tensor = tensor + attn
            tensor = self.layer_norm1[i](tensor)

            # encoder attention (for decoder only)
            # if self.is_decoder and src_enc is not None:
            #     attn = self.encoder_attn[i](tensor, src_mask, kv=src_enc, cache=cache)
            #     attn = F.dropout(attn, p=self.dropout, training=self.training)
            #     tensor = tensor + attn
            #     tensor = self.layer_norm15[i](tensor)

            # FFN
            tensor = tensor + self.ffns[i](tensor)
            tensor = self.layer_norm2[i](tensor)
            tensor *= mask.unsqueeze(-1).to(tensor.dtype)

        # Add last hidden state
        if self.output_hidden_states:
            hidden_states = hidden_states + (tensor,)

        # update cache length
        if cache is not None:
            cache['slen'] += tensor.size(1)

        # move back sequence length to dimension 0
        # tensor = tensor.transpose(0, 1)

        outputs = (tensor,)
        if self.output_hidden_states:
            outputs = outputs + (hidden_states,)
        if self.output_attentions:
            outputs = outputs + (attentions,)
        return outputs  # outputs, (hidden_states), (attentions)


class XLMPredLayer(nn.Module):
    """
    Prediction layer (cross_entropy or adaptive_softmax).
    """
    def __init__(self, config):
        super(XLMPredLayer, self).__init__()
        self.asm = config.asm
        self.n_words = config.n_words
        self.pad_index = config.pad_index
        dim = config.emb_dim

        if config.asm is False:
            self.proj = nn.Linear(dim, config.n_words, bias=True)
        else:
            self.proj = nn.AdaptiveLogSoftmaxWithLoss(
                in_features=dim,
                n_classes=config.n_words,
                cutoffs=config.asm_cutoffs,
                div_value=config.asm_div_value,
                head_bias=True,  # default is False
            )

    def forward(self, x, y=None):
        """ Compute the loss, and optionally the scores.
        """
        outputs = ()
        if self.asm is False:
            scores = self.proj(x).view(-1, self.n_words)
            outputs = (scores,) + outputs
            if y is not None:
                loss = F.cross_entropy(scores, y, reduction='elementwise_mean')
                outputs = (loss,) + outputs
        else:
            scores = self.proj.log_prob(x)
            outputs = (scores,) + outputs
            if y is not None:
                _, loss = self.proj(x, y)
                outputs = (loss,) + outputs

        return outputs


@add_start_docstrings("""The XLM Model transformer with a language modeling head on top
    (linear layer with weights tied to the input embeddings). """,
    XLM_START_DOCSTRING, XLM_INPUTS_DOCSTRING)
class XLMWithLMHeadModel(XLMPreTrainedModel):
    r"""
        **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
            Labels for language modeling.
            Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
            Indices are selected in ``[-1, 0, ..., config.vocab_size]``
            All labels set to ``-1`` are ignored (masked), the loss is only
            computed for labels in ``[0, ..., config.vocab_size]``

    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
            Language modeling loss.
        **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).
        **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 = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
        model = XLMWithLMHeadModel.from_pretrained('xlm-mlm-en-2048')
        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):
        super(XLMWithLMHeadModel, self).__init__(config)
        self.transformer = XLMModel(config)
        self.pred_layer = XLMPredLayer(config)

        self.init_weights()
        self.tie_weights()

    def tie_weights(self):
        """ Make sure we are sharing the embeddings
        """
        self._tie_or_clone_weights(self.pred_layer.proj, self.transformer.embeddings)

    def forward(self, input_ids, lengths=None, position_ids=None, langs=None, token_type_ids=None,
                attention_mask=None, cache=None, labels=None, head_mask=None):
        transformer_outputs = self.transformer(input_ids, lengths=lengths, position_ids=position_ids,
                                               token_type_ids=token_type_ids, langs=langs,
                                               attention_mask=attention_mask, cache=cache, head_mask=head_mask)

        output = transformer_outputs[0]
        outputs = self.pred_layer(output, labels)
        outputs = outputs + transformer_outputs[1:]  # Keep new_mems and attention/hidden states if they are here

        return outputs


@add_start_docstrings("""XLM Model with a sequence classification/regression head on top (a linear layer on top of
    the pooled output) e.g. for GLUE tasks. """,
    XLM_START_DOCSTRING, XLM_INPUTS_DOCSTRING)
class XLMForSequenceClassification(XLMPreTrainedModel):
    r"""
        **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
            Labels for computing the sequence classification/regression loss.
            Indices should be in ``[0, ..., config.num_labels - 1]``.
            If ``config.num_labels == 1`` a regression loss is computed (Mean-Square loss),
            If ``config.num_labels > 1`` a classification loss is computed (Cross-Entropy).

    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
            Classification (or regression if config.num_labels==1) loss.
        **logits**: ``torch.FloatTensor`` of shape ``(batch_size, config.num_labels)``
            Classification (or regression if config.num_labels==1) scores (before SoftMax).
        **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 = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
        model = XLMForSequenceClassification.from_pretrained('xlm-mlm-en-2048')
        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
        labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
        outputs = model(input_ids, labels=labels)
        loss, logits = outputs[:2]

    """
    def __init__(self, config):
        super(XLMForSequenceClassification, self).__init__(config)
        self.num_labels = config.num_labels

        self.transformer = XLMModel(config)
        self.sequence_summary = SequenceSummary(config)

        self.init_weights()

    def forward(self, input_ids, lengths=None, position_ids=None, langs=None, token_type_ids=None,
                attention_mask=None, cache=None, labels=None, head_mask=None):
        transformer_outputs = self.transformer(input_ids, lengths=lengths, position_ids=position_ids,
                                               token_type_ids=token_type_ids, langs=langs,
                                               attention_mask=attention_mask, cache=cache, head_mask=head_mask)

        output = transformer_outputs[0]
        logits = self.sequence_summary(output)

        outputs = (logits,) + transformer_outputs[1:]  # Keep new_mems and attention/hidden states if they are here

        if labels is not None:
            if self.num_labels == 1:
                #  We are doing regression
                loss_fct = MSELoss()
                loss = loss_fct(logits.view(-1), labels.view(-1))
            else:
                loss_fct = CrossEntropyLoss()
                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
            outputs = (loss,) + outputs

        return outputs


@add_start_docstrings("""XLM Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
    the hidden-states output to compute `span start logits` and `span end logits`). """,
    XLM_START_DOCSTRING, XLM_INPUTS_DOCSTRING)
class XLMForQuestionAnswering(XLMPreTrainedModel):
    r"""
        **start_positions**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
            Labels for position (index) of the start of the labelled span for computing the token classification loss.
            Positions are clamped to the length of the sequence (`sequence_length`).
            Position outside of the sequence are not taken into account for computing the loss.
        **end_positions**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
            Labels for position (index) of the end of the labelled span for computing the token classification loss.
            Positions are clamped to the length of the sequence (`sequence_length`).
            Position outside of the sequence are not taken into account for computing the loss.
        **is_impossible**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
            Labels whether a question has an answer or no answer (SQuAD 2.0)
        **cls_index**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
            Labels for position (index) of the classification token to use as input for computing plausibility of the answer.
        **p_mask**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
            Optional mask of tokens which can't be in answers (e.g. [CLS], [PAD], ...) 

    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
            Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.
        **start_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length,)``
            Span-start scores (before SoftMax).
        **end_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length,)``
            Span-end scores (before SoftMax).
        **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 = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
        model = XLMForQuestionAnswering.from_pretrained('xlm-mlm-en-2048')
        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
        start_positions = torch.tensor([1])
        end_positions = torch.tensor([3])
        outputs = model(input_ids, start_positions=start_positions, end_positions=end_positions)
        loss, start_scores, end_scores = outputs[:2]

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

        self.transformer = XLMModel(config)
        self.qa_outputs = SQuADHead(config)

        self.init_weights()

    def forward(self, input_ids, lengths=None, position_ids=None, langs=None, token_type_ids=None,
                attention_mask=None, cache=None, start_positions=None, end_positions=None,
                cls_index=None, is_impossible=None, p_mask=None, head_mask=None):
        transformer_outputs = self.transformer(input_ids, lengths=lengths, position_ids=position_ids,
                                               token_type_ids=token_type_ids, langs=langs,
                                               attention_mask=attention_mask, cache=cache, head_mask=head_mask)

        output = transformer_outputs[0]

        outputs = self.qa_outputs(output, start_positions=start_positions, end_positions=end_positions,
                                  cls_index=cls_index, is_impossible=is_impossible, p_mask=p_mask)

        outputs = outputs + transformer_outputs[1:]  # Keep new_mems and attention/hidden states if they are here

        return outputs