# coding=utf-8
# Copyright 2020 The SqueezeBert authors 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 SqueezeBert model. """
import math
import torch
from torch import nn
from torch.nn import CrossEntropyLoss, MSELoss
from ...activations import ACT2FN
from ...file_utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward
from ...modeling_outputs import (
BaseModelOutput,
BaseModelOutputWithPooling,
MaskedLMOutput,
MultipleChoiceModelOutput,
QuestionAnsweringModelOutput,
SequenceClassifierOutput,
TokenClassifierOutput,
)
from ...modeling_utils import PreTrainedModel
from ...utils import logging
from .configuration_squeezebert import SqueezeBertConfig
logger = logging.get_logger(__name__)
_CONFIG_FOR_DOC = "SqueezeBertConfig"
_TOKENIZER_FOR_DOC = "SqueezeBertTokenizer"
SQUEEZEBERT_PRETRAINED_MODEL_ARCHIVE_LIST = [
"squeezebert/squeezebert-uncased",
"squeezebert/squeezebert-mnli",
"squeezebert/squeezebert-mnli-headless",
]
class SqueezeBertEmbeddings(nn.Module):
"""Construct the embeddings from word, position and token_type embeddings."""
def __init__(self, config):
super().__init__()
self.word_embeddings = nn.Embedding(config.vocab_size, config.embedding_size, padding_idx=config.pad_token_id)
self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.embedding_size)
self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.embedding_size)
# self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
# any TensorFlow checkpoint file
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
# position_ids (1, len position emb) is contiguous in memory and exported when serialized
self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
def forward(self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None):
if input_ids is not None:
input_shape = input_ids.size()
else:
input_shape = inputs_embeds.size()[:-1]
seq_length = input_shape[1]
if position_ids is None:
position_ids = self.position_ids[:, :seq_length]
if token_type_ids is None:
token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)
if inputs_embeds is None:
inputs_embeds = self.word_embeddings(input_ids)
position_embeddings = self.position_embeddings(position_ids)
token_type_embeddings = self.token_type_embeddings(token_type_ids)
embeddings = inputs_embeds + position_embeddings + token_type_embeddings
embeddings = self.LayerNorm(embeddings)
embeddings = self.dropout(embeddings)
return embeddings
class MatMulWrapper(torch.nn.Module):
"""
Wrapper for torch.matmul(). This makes flop-counting easier to implement. Note that if you directly call
torch.matmul() in your code, the flop counter will typically ignore the flops of the matmul.
"""
def __init__(self):
super().__init__()
def forward(self, mat1, mat2):
"""
:param inputs: two torch tensors :return: matmul of these tensors
Here are the typical dimensions found in BERT (the B is optional) mat1.shape: [B, <optional extra dims>, M, K]
mat2.shape: [B, <optional extra dims>, K, N] output shape: [B, <optional extra dims>, M, N]
"""
return torch.matmul(mat1, mat2)
class SqueezeBertLayerNorm(nn.LayerNorm):
"""
This is a nn.LayerNorm subclass that accepts NCW data layout and performs normalization in the C dimension.
N = batch C = channels W = sequence length
"""
def __init__(self, hidden_size, eps=1e-12):
nn.LayerNorm.__init__(self, normalized_shape=hidden_size, eps=eps) # instantiates self.{weight, bias, eps}
def forward(self, x):
x = x.permute(0, 2, 1)
x = nn.LayerNorm.forward(self, x)
return x.permute(0, 2, 1)
class ConvDropoutLayerNorm(nn.Module):
"""
ConvDropoutLayerNorm: Conv, Dropout, LayerNorm
"""
def __init__(self, cin, cout, groups, dropout_prob):
super().__init__()
self.conv1d = nn.Conv1d(in_channels=cin, out_channels=cout, kernel_size=1, groups=groups)
self.layernorm = SqueezeBertLayerNorm(cout)
self.dropout = nn.Dropout(dropout_prob)
def forward(self, hidden_states, input_tensor):
x = self.conv1d(hidden_states)
x = self.dropout(x)
x = x + input_tensor
x = self.layernorm(x)
return x
class ConvActivation(nn.Module):
"""
ConvActivation: Conv, Activation
"""
def __init__(self, cin, cout, groups, act):
super().__init__()
self.conv1d = nn.Conv1d(in_channels=cin, out_channels=cout, kernel_size=1, groups=groups)
self.act = ACT2FN[act]
def forward(self, x):
output = self.conv1d(x)
return self.act(output)
class SqueezeBertSelfAttention(nn.Module):
def __init__(self, config, cin, q_groups=1, k_groups=1, v_groups=1):
"""
config = used for some things; ignored for others (work in progress...) cin = input channels = output channels
groups = number of groups to use in conv1d layers
"""
super().__init__()
if cin % config.num_attention_heads != 0:
raise ValueError(
"cin (%d) is not a multiple of the number of attention "
"heads (%d)" % (cin, config.num_attention_heads)
)
self.num_attention_heads = config.num_attention_heads
self.attention_head_size = int(cin / config.num_attention_heads)
self.all_head_size = self.num_attention_heads * self.attention_head_size
self.query = nn.Conv1d(in_channels=cin, out_channels=cin, kernel_size=1, groups=q_groups)
self.key = nn.Conv1d(in_channels=cin, out_channels=cin, kernel_size=1, groups=k_groups)
self.value = nn.Conv1d(in_channels=cin, out_channels=cin, kernel_size=1, groups=v_groups)
self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
self.softmax = nn.Softmax(dim=-1)
self.matmul_qk = MatMulWrapper()
self.matmul_qkv = MatMulWrapper()
def transpose_for_scores(self, x):
"""
- input: [N, C, W]
- output: [N, C1, W, C2] where C1 is the head index, and C2 is one head's contents
"""
new_x_shape = (x.size()[0], self.num_attention_heads, self.attention_head_size, x.size()[-1]) # [N, C1, C2, W]
x = x.view(*new_x_shape)
return x.permute(0, 1, 3, 2) # [N, C1, C2, W] --> [N, C1, W, C2]
def transpose_key_for_scores(self, x):
"""
- input: [N, C, W]
- output: [N, C1, C2, W] where C1 is the head index, and C2 is one head's contents
"""
new_x_shape = (x.size()[0], self.num_attention_heads, self.attention_head_size, x.size()[-1]) # [N, C1, C2, W]
x = x.view(*new_x_shape)
# no `permute` needed
return x
def transpose_output(self, x):
"""
- input: [N, C1, W, C2]
- output: [N, C, W]
"""
x = x.permute(0, 1, 3, 2).contiguous() # [N, C1, C2, W]
new_x_shape = (x.size()[0], self.all_head_size, x.size()[3]) # [N, C, W]
x = x.view(*new_x_shape)
return x
def forward(self, hidden_states, attention_mask, output_attentions):
"""
expects hidden_states in [N, C, W] data layout.
The attention_mask data layout is [N, W], and it does not need to be transposed.
"""
mixed_query_layer = self.query(hidden_states)
mixed_key_layer = self.key(hidden_states)
mixed_value_layer = self.value(hidden_states)
query_layer = self.transpose_for_scores(mixed_query_layer)
key_layer = self.transpose_key_for_scores(mixed_key_layer)
value_layer = self.transpose_for_scores(mixed_value_layer)
# Take the dot product between "query" and "key" to get the raw attention scores.
attention_score = self.matmul_qk(query_layer, key_layer)
attention_score = attention_score / math.sqrt(self.attention_head_size)
# Apply the attention mask is (precomputed for all layers in BertModel forward() function)
attention_score = attention_score + attention_mask
# Normalize the attention scores to probabilities.
attention_probs = self.softmax(attention_score)
# This is actually dropping out entire tokens to attend to, which might
# seem a bit unusual, but is taken from the original Transformer paper.
attention_probs = self.dropout(attention_probs)
context_layer = self.matmul_qkv(attention_probs, value_layer)
context_layer = self.transpose_output(context_layer)
result = {"context_layer": context_layer}
if output_attentions:
result["attention_score"] = attention_score
return result
class SqueezeBertModule(nn.Module):
def __init__(self, config):
"""
- hidden_size = input chans = output chans for Q, K, V (they are all the same ... for now) = output chans for
the module
- intermediate_size = output chans for intermediate layer
- groups = number of groups for all layers in the BertModule. (eventually we could change the interface to
allow different groups for different layers)
"""
super().__init__()
c0 = config.hidden_size
c1 = config.hidden_size
c2 = config.intermediate_size
c3 = config.hidden_size
self.attention = SqueezeBertSelfAttention(
config=config, cin=c0, q_groups=config.q_groups, k_groups=config.k_groups, v_groups=config.v_groups
)
self.post_attention = ConvDropoutLayerNorm(
cin=c0, cout=c1, groups=config.post_attention_groups, dropout_prob=config.hidden_dropout_prob
)
self.intermediate = ConvActivation(cin=c1, cout=c2, groups=config.intermediate_groups, act=config.hidden_act)
self.output = ConvDropoutLayerNorm(
cin=c2, cout=c3, groups=config.output_groups, dropout_prob=config.hidden_dropout_prob
)
def forward(self, hidden_states, attention_mask, output_attentions):
att = self.attention(hidden_states, attention_mask, output_attentions)
attention_output = att["context_layer"]
post_attention_output = self.post_attention(attention_output, hidden_states)
intermediate_output = self.intermediate(post_attention_output)
layer_output = self.output(intermediate_output, post_attention_output)
output_dict = {"feature_map": layer_output}
if output_attentions:
output_dict["attention_score"] = att["attention_score"]
return output_dict
class SqueezeBertEncoder(nn.Module):
def __init__(self, config):
super().__init__()
assert config.embedding_size == config.hidden_size, (
"If you want embedding_size != intermediate hidden_size,"
"please insert a Conv1d layer to adjust the number of channels "
"before the first SqueezeBertModule."
)
self.layers = nn.ModuleList(SqueezeBertModule(config) for _ in range(config.num_hidden_layers))
def forward(
self,
hidden_states,
attention_mask=None,
head_mask=None,
output_attentions=False,
output_hidden_states=False,
return_dict=True,
):
if head_mask is None:
head_mask_is_all_none = True
elif head_mask.count(None) == len(head_mask):
head_mask_is_all_none = True
else:
head_mask_is_all_none = False
assert head_mask_is_all_none is True, "head_mask is not yet supported in the SqueezeBert implementation."
# [batch_size, sequence_length, hidden_size] --> [batch_size, hidden_size, sequence_length]
hidden_states = hidden_states.permute(0, 2, 1)
all_hidden_states = () if output_hidden_states else None
all_attentions = () if output_attentions else None
for layer in self.layers:
if output_hidden_states:
hidden_states = hidden_states.permute(0, 2, 1)
all_hidden_states += (hidden_states,)
hidden_states = hidden_states.permute(0, 2, 1)
layer_output = layer.forward(hidden_states, attention_mask, output_attentions)
hidden_states = layer_output["feature_map"]
if output_attentions:
all_attentions += (layer_output["attention_score"],)
# [batch_size, hidden_size, sequence_length] --> [batch_size, sequence_length, hidden_size]
hidden_states = hidden_states.permute(0, 2, 1)
if output_hidden_states:
all_hidden_states += (hidden_states,)
if not return_dict:
return tuple(v for v in [hidden_states, all_hidden_states, all_attentions] if v is not None)
return BaseModelOutput(
last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
)
class SqueezeBertPooler(nn.Module):
def __init__(self, config):
super().__init__()
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
self.activation = nn.Tanh()
def forward(self, hidden_states):
# We "pool" the model by simply taking the hidden state corresponding
# to the first token.
first_token_tensor = hidden_states[:, 0]
pooled_output = self.dense(first_token_tensor)
pooled_output = self.activation(pooled_output)
return pooled_output
class SqueezeBertPredictionHeadTransform(nn.Module):
def __init__(self, config):
super().__init__()
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
if isinstance(config.hidden_act, str):
self.transform_act_fn = ACT2FN[config.hidden_act]
else:
self.transform_act_fn = config.hidden_act
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
def forward(self, hidden_states):
hidden_states = self.dense(hidden_states)
hidden_states = self.transform_act_fn(hidden_states)
hidden_states = self.LayerNorm(hidden_states)
return hidden_states
class SqueezeBertLMPredictionHead(nn.Module):
def __init__(self, config):
super().__init__()
self.transform = SqueezeBertPredictionHeadTransform(config)
# The output weights are the same as the input embeddings, but there is
# an output-only bias for each token.
self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
self.bias = nn.Parameter(torch.zeros(config.vocab_size))
# Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
self.decoder.bias = self.bias
def forward(self, hidden_states):
hidden_states = self.transform(hidden_states)
hidden_states = self.decoder(hidden_states)
return hidden_states
class SqueezeBertOnlyMLMHead(nn.Module):
def __init__(self, config):
super().__init__()
self.predictions = SqueezeBertLMPredictionHead(config)
def forward(self, sequence_output):
prediction_scores = self.predictions(sequence_output)
return prediction_scores
class SqueezeBertPreTrainedModel(PreTrainedModel):
"""
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
models.
"""
config_class = SqueezeBertConfig
base_model_prefix = "transformer"
_keys_to_ignore_on_load_missing = [r"position_ids"]
def _init_weights(self, module):
""" Initialize the weights """
if isinstance(module, (nn.Linear, nn.Conv1d, 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, SqueezeBertLayerNorm):
module.bias.data.zero_()
module.weight.data.fill_(1.0)
if isinstance(module, (nn.Linear, nn.Conv1d)) and module.bias is not None:
module.bias.data.zero_()
SQUEEZEBERT_START_DOCSTRING = r"""
The SqueezeBERT model was proposed in `SqueezeBERT: What can computer vision teach NLP about efficient neural
networks? <https://arxiv.org/abs/2006.11316>`__ by Forrest N. Iandola, Albert E. Shaw, Ravi Krishna, and Kurt W.
Keutzer
This model inherits from :class:`~transformers.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.
For best results finetuning SqueezeBERT on text classification tasks, it is recommended to use the
`squeezebert/squeezebert-mnli-headless` checkpoint as a starting point.
Parameters:
config (:class:`~transformers.SqueezeBertConfig`): 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:`~transformers.PreTrainedModel.from_pretrained` method to load the model
weights.
Hierarchy::
Internal class hierarchy:
SqueezeBertModel
SqueezeBertEncoder
SqueezeBertModule
SqueezeBertSelfAttention
ConvActivation
ConvDropoutLayerNorm
Data layouts::
Input data is in [batch, sequence_length, hidden_size] format.
Data inside the encoder is in [batch, hidden_size, sequence_length] format. But, if :obj:`output_hidden_states
== True`, the data from inside the encoder is returned in [batch, sequence_length, hidden_size] format.
The final output of the encoder is in [batch, sequence_length, hidden_size] format.
"""
SQUEEZEBERT_INPUTS_DOCSTRING = r"""
Args:
input_ids (:obj:`torch.LongTensor` of shape :obj:`({0})`):
Indices of input sequence tokens in the vocabulary.
Indices can be obtained using :class:`~transformers.SqueezeBertTokenizer`. See
:meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
details.
`What are input IDs? <../glossary.html#input-ids>`__
attention_mask (:obj:`torch.FloatTensor` of shape :obj:`({0})`, `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.html#attention-mask>`__
token_type_ids (:obj:`torch.LongTensor` of shape :obj:`({0})`, `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.html#token-type-ids>`_
position_ids (:obj:`torch.LongTensor` of shape :obj:`({0})`, `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.html#position-ids>`_
head_mask (:obj:`torch.FloatTensor` of shape :obj:`(num_heads,)` or :obj:`(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 (:obj:`torch.FloatTensor` of shape :obj:`({0}, hidden_size)`, `optional`):
Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
This is useful if you want more control over how to convert :obj:`input_ids` indices into associated
vectors than the model's internal embedding lookup matrix.
output_attentions (:obj:`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 (:obj:`bool`, `optional`):
Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
more detail.
return_dict (:obj:`bool`, `optional`):
Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.
"""
[docs]@add_start_docstrings(
"The bare SqueezeBERT Model transformer outputting raw hidden-states without any specific head on top.",
SQUEEZEBERT_START_DOCSTRING,
)
class SqueezeBertModel(SqueezeBertPreTrainedModel):
def __init__(self, config):
super().__init__(config)
self.embeddings = SqueezeBertEmbeddings(config)
self.encoder = SqueezeBertEncoder(config)
self.pooler = SqueezeBertPooler(config)
self.init_weights()
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.encoder.layer[layer].attention.prune_heads(heads)
[docs] @add_start_docstrings_to_model_forward(SQUEEZEBERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint="squeezebert/squeezebert-mnli-headless",
output_type=BaseModelOutputWithPooling,
config_class=_CONFIG_FOR_DOC,
)
def forward(
self,
input_ids=None,
attention_mask=None,
token_type_ids=None,
position_ids=None,
head_mask=None,
inputs_embeds=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
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
if input_ids is not None and inputs_embeds is not None:
raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
elif input_ids is not None:
input_shape = input_ids.size()
elif inputs_embeds is not None:
input_shape = inputs_embeds.size()[:-1]
else:
raise ValueError("You have to specify either input_ids or inputs_embeds")
device = input_ids.device if input_ids is not None else inputs_embeds.device
if attention_mask is None:
attention_mask = torch.ones(input_shape, device=device)
if token_type_ids is None:
token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
extended_attention_mask = self.get_extended_attention_mask(attention_mask, input_shape, device)
# Prepare head mask if needed
# 1.0 in head_mask indicate we keep the head
# attention_probs has shape bsz x n_heads x N x N
# input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
# and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
embedding_output = self.embeddings(
input_ids=input_ids, position_ids=position_ids, token_type_ids=token_type_ids, inputs_embeds=inputs_embeds
)
encoder_outputs = self.encoder(
hidden_states=embedding_output,
attention_mask=extended_attention_mask,
head_mask=head_mask,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
sequence_output = encoder_outputs[0]
pooled_output = self.pooler(sequence_output)
if not return_dict:
return (sequence_output, pooled_output) + encoder_outputs[1:]
return BaseModelOutputWithPooling(
last_hidden_state=sequence_output,
pooler_output=pooled_output,
hidden_states=encoder_outputs.hidden_states,
attentions=encoder_outputs.attentions,
)
[docs]@add_start_docstrings(
"""
SqueezeBERT Model transformer with a sequence classification/regression head on top (a linear layer on top of the
pooled output) e.g. for GLUE tasks.
""",
SQUEEZEBERT_START_DOCSTRING,
)
class SqueezeBertForSequenceClassification(SqueezeBertPreTrainedModel):
def __init__(self, config):
super().__init__(config)
self.num_labels = config.num_labels
self.transformer = SqueezeBertModel(config)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)
self.init_weights()
[docs] @add_start_docstrings_to_model_forward(SQUEEZEBERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint="squeezebert/squeezebert-mnli-headless",
output_type=SequenceClassifierOutput,
config_class=_CONFIG_FOR_DOC,
)
def forward(
self,
input_ids=None,
attention_mask=None,
token_type_ids=None,
position_ids=None,
head_mask=None,
inputs_embeds=None,
labels=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
r"""
labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
Labels for computing the sequence classification/regression loss. Indices should be in :obj:`[0, ...,
config.num_labels - 1]`. If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
outputs = self.transformer(
input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
position_ids=position_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
pooled_output = outputs[1]
pooled_output = self.dropout(pooled_output)
logits = self.classifier(pooled_output)
loss = None
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))
if not return_dict:
output = (logits,) + outputs[2:]
return ((loss,) + output) if loss is not None else output
return SequenceClassifierOutput(
loss=loss,
logits=logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
[docs]@add_start_docstrings(
"""
SqueezeBERT Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g.
for Named-Entity-Recognition (NER) tasks.
""",
SQUEEZEBERT_START_DOCSTRING,
)
class SqueezeBertForTokenClassification(SqueezeBertPreTrainedModel):
def __init__(self, config):
super().__init__(config)
self.num_labels = config.num_labels
self.transformer = SqueezeBertModel(config)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
self.classifier = nn.Linear(config.hidden_size, config.num_labels)
self.init_weights()
[docs] @add_start_docstrings_to_model_forward(SQUEEZEBERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint="squeezebert/squeezebert-mnli-headless",
output_type=TokenClassifierOutput,
config_class=_CONFIG_FOR_DOC,
)
def forward(
self,
input_ids=None,
attention_mask=None,
token_type_ids=None,
position_ids=None,
head_mask=None,
inputs_embeds=None,
labels=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
r"""
labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
Labels for computing the token classification loss. Indices should be in ``[0, ..., config.num_labels -
1]``.
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
outputs = self.transformer(
input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
position_ids=position_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
sequence_output = outputs[0]
sequence_output = self.dropout(sequence_output)
logits = self.classifier(sequence_output)
loss = None
if labels is not None:
loss_fct = CrossEntropyLoss()
# Only keep active parts of the loss
if attention_mask is not None:
active_loss = attention_mask.view(-1) == 1
active_logits = logits.view(-1, self.num_labels)
active_labels = torch.where(
active_loss, labels.view(-1), torch.tensor(loss_fct.ignore_index).type_as(labels)
)
loss = loss_fct(active_logits, active_labels)
else:
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
if not return_dict:
output = (logits,) + outputs[2:]
return ((loss,) + output) if loss is not None else output
return TokenClassifierOutput(
loss=loss,
logits=logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
[docs]@add_start_docstrings(
"""
SqueezeBERT 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`).
""",
SQUEEZEBERT_START_DOCSTRING,
)
class SqueezeBertForQuestionAnswering(SqueezeBertPreTrainedModel):
def __init__(self, config):
super().__init__(config)
self.num_labels = config.num_labels
self.transformer = SqueezeBertModel(config)
self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
self.init_weights()
[docs] @add_start_docstrings_to_model_forward(SQUEEZEBERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint="squeezebert/squeezebert-mnli-headless",
output_type=QuestionAnsweringModelOutput,
config_class=_CONFIG_FOR_DOC,
)
def forward(
self,
input_ids=None,
attention_mask=None,
token_type_ids=None,
position_ids=None,
head_mask=None,
inputs_embeds=None,
start_positions=None,
end_positions=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
r"""
start_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
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 (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
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.
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
outputs = self.transformer(
input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
position_ids=position_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
sequence_output = outputs[0]
logits = self.qa_outputs(sequence_output)
start_logits, end_logits = logits.split(1, dim=-1)
start_logits = start_logits.squeeze(-1)
end_logits = end_logits.squeeze(-1)
total_loss = None
if start_positions is not None and end_positions is not None:
# If we are on multi-GPU, split add a dimension
if len(start_positions.size()) > 1:
start_positions = start_positions.squeeze(-1)
if len(end_positions.size()) > 1:
end_positions = end_positions.squeeze(-1)
# sometimes the start/end positions are outside our model inputs, we ignore these terms
ignored_index = start_logits.size(1)
start_positions.clamp_(0, ignored_index)
end_positions.clamp_(0, ignored_index)
loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
start_loss = loss_fct(start_logits, start_positions)
end_loss = loss_fct(end_logits, end_positions)
total_loss = (start_loss + end_loss) / 2
if not return_dict:
output = (start_logits, end_logits) + outputs[2:]
return ((total_loss,) + output) if total_loss is not None else output
return QuestionAnsweringModelOutput(
loss=total_loss,
start_logits=start_logits,
end_logits=end_logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)