FM / flaubert

/usr/local/lib/python3.6/site-packages/transformers/models/flaubert/

/usr/local/lib/python3.6/site-packages/transformers/models/flaubert

Name	Size	Mode	Actions
__pycache__/	-	0755	rm
configuration_flaubert.py	8940	0644	edit dl rm
modeling_flaubert.py	18489	0644	edit dl rm
modeling_tf_flaubert.py	38347	0644	edit dl rm
tokenization_flaubert.py	5633	0644	edit dl rm
__init__.py	3059	0644	edit dl rm

Edit: /usr/local/lib/python3.6/site-packages/transformers/models/flaubert/modeling_flaubert.py (18489B)

# coding=utf-8
# Copyright 2019-present CNRS, 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 Flaubert model, based on XLM."""

import random
from typing import Dict, Optional, Tuple, Union

import torch
from packaging import version
from torch import nn

from ...modeling_outputs import BaseModelOutput
from ...utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward, logging
from ..xlm.modeling_xlm import (
    XLMForMultipleChoice,
    XLMForQuestionAnswering,
    XLMForQuestionAnsweringSimple,
    XLMForSequenceClassification,
    XLMForTokenClassification,
    XLMModel,
    XLMWithLMHeadModel,
    get_masks,
)
from .configuration_flaubert import FlaubertConfig

logger = logging.get_logger(__name__)

_CHECKPOINT_FOR_DOC = "flaubert/flaubert_base_cased"
_CONFIG_FOR_DOC = "FlaubertConfig"
_TOKENIZER_FOR_DOC = "FlaubertTokenizer"

FLAUBERT_PRETRAINED_MODEL_ARCHIVE_LIST = [
    "flaubert/flaubert_small_cased",
    "flaubert/flaubert_base_uncased",
    "flaubert/flaubert_base_cased",
    "flaubert/flaubert_large_cased",
    # See all Flaubert models at https://huggingface.co/models?filter=flaubert
]

FLAUBERT_START_DOCSTRING = r"""

This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
    etc.)

This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
    and behavior.

Parameters:
        config ([`FlaubertConfig`]): 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 [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""

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

Indices can be obtained using [`FlaubertTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.

[What are input IDs?](../glossary#input-ids)
        attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

- 1 for tokens that are **not masked**,
            - 0 for tokens that are **masked**.

[What are attention masks?](../glossary#attention-mask)
        token_type_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
            1]`:

- 0 corresponds to a *sentence A* token,
            - 1 corresponds to a *sentence B* token.

[What are token type IDs?](../glossary#token-type-ids)
        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
            config.max_position_embeddings - 1]`.

[What are position IDs?](../glossary#position-ids)
        lengths (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            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 compatibility. Indices selected in
            `[0, ..., input_ids.size(-1)]`:
        cache (`Dict[str, torch.FloatTensor]`, *optional*):
            Dictionary strings to `torch.FloatTensor` that contains precomputed 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 (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
            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**.

inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
            model's internal embedding lookup matrix.
        output_attentions (`bool`, *optional*):
            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
            tensors for more detail.
        output_hidden_states (`bool`, *optional*):
            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
            more detail.
        return_dict (`bool`, *optional*):
            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""

@add_start_docstrings(
    "The bare Flaubert Model transformer outputting raw hidden-states without any specific head on top.",
    FLAUBERT_START_DOCSTRING,
)
class FlaubertModel(XLMModel):

config_class = FlaubertConfig

def __init__(self, config):  # , dico, is_encoder, with_output):
        super().__init__(config)
        self.layerdrop = getattr(config, "layerdrop", 0.0)
        self.pre_norm = getattr(config, "pre_norm", False)
        if version.parse(torch.__version__) > version.parse("1.6.0"):
            self.register_buffer(
                "position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)), persistent=False
            )

@add_start_docstrings_to_model_forward(FLAUBERT_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(
        processor_class=_TOKENIZER_FOR_DOC,
        checkpoint=_CHECKPOINT_FOR_DOC,
        output_type=BaseModelOutput,
        config_class=_CONFIG_FOR_DOC,
    )
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        langs: Optional[torch.Tensor] = None,
        token_type_ids: Optional[torch.LongTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        lengths: Optional[torch.LongTensor] = None,
        cache: Optional[Dict[str, torch.FloatTensor]] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, BaseModelOutput]:
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

# removed: src_enc=None, src_len=None
        if input_ids is not None:
            bs, slen = input_ids.size()
        else:
            bs, slen = inputs_embeds.size()[:-1]

device = input_ids.device if input_ids is not None else inputs_embeds.device

if lengths is None:
            if input_ids is not None:
                lengths = (input_ids != self.pad_index).sum(dim=1).long()
            else:
                lengths = torch.tensor([slen] * bs, device=device)
        # mask = input_ids != self.pad_index

# check inputs
        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]

# Setting the position-ids to the registered buffer in constructor, it helps
        # when tracing the model without passing position-ids, solves
        # isues similar to issue #5664
        if position_ids is None:
            if hasattr(self, "position_ids"):
                position_ids = self.position_ids[:, :slen]
                position_ids = position_ids.expand((bs, slen))
            else:
                position_ids = torch.arange(slen, dtype=torch.long, device=device)
                position_ids = position_ids.unsqueeze(0).expand((bs, slen))
        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
        head_mask = self.get_head_mask(head_mask, self.config.n_layers)

# do not recompute cached elements
        if cache is not None and input_ids 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
        if inputs_embeds is None:
            inputs_embeds = self.embeddings(input_ids)

tensor = inputs_embeds + self.position_embeddings(position_ids).expand_as(inputs_embeds)
        if langs is not None and self.use_lang_emb and self.config.n_langs > 1:
            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 = nn.functional.dropout(tensor, p=self.dropout, training=self.training)
        tensor *= mask.unsqueeze(-1).to(tensor.dtype)

# transformer layers
        hidden_states = () if output_hidden_states else None
        attentions = () if output_attentions else None
        for i in range(self.n_layers):
            # LayerDrop
            dropout_probability = random.uniform(0, 1)
            if self.training and (dropout_probability < self.layerdrop):
                continue

if output_hidden_states:
                hidden_states = hidden_states + (tensor,)

# self attention
            if not self.pre_norm:
                attn_outputs = self.attentions[i](
                    tensor,
                    attn_mask,
                    cache=cache,
                    head_mask=head_mask[i],
                    output_attentions=output_attentions,
                )
                attn = attn_outputs[0]
                if output_attentions:
                    attentions = attentions + (attn_outputs[1],)
                attn = nn.functional.dropout(attn, p=self.dropout, training=self.training)
                tensor = tensor + attn
                tensor = self.layer_norm1[i](tensor)
            else:
                tensor_normalized = self.layer_norm1[i](tensor)
                attn_outputs = self.attentions[i](tensor_normalized, attn_mask, cache=cache, head_mask=head_mask[i])
                attn = attn_outputs[0]
                if output_attentions:
                    attentions = attentions + (attn_outputs[1],)
                attn = nn.functional.dropout(attn, p=self.dropout, training=self.training)
                tensor = tensor + attn

# 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 = nn.functional.dropout(attn, p=self.dropout, training=self.training)
            #     tensor = tensor + attn
            #     tensor = self.layer_norm15[i](tensor)

# FFN
            if not self.pre_norm:
                tensor = tensor + self.ffns[i](tensor)
                tensor = self.layer_norm2[i](tensor)
            else:
                tensor_normalized = self.layer_norm2[i](tensor)
                tensor = tensor + self.ffns[i](tensor_normalized)

tensor *= mask.unsqueeze(-1).to(tensor.dtype)

# Add last hidden state
        if 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)

if not return_dict:
            return tuple(v for v in [tensor, hidden_states, attentions] if v is not None)

return BaseModelOutput(last_hidden_state=tensor, hidden_states=hidden_states, attentions=attentions)

@add_start_docstrings(
    """
    The Flaubert Model transformer with a language modeling head on top (linear layer with weights tied to the input
    embeddings).
    """,
    FLAUBERT_START_DOCSTRING,
)
class FlaubertWithLMHeadModel(XLMWithLMHeadModel):
    """
    This class overrides [`XLMWithLMHeadModel`]. Please check the superclass for the appropriate documentation
    alongside usage examples.
    """

config_class = FlaubertConfig