# coding=utf-8
# Copyright 2021 The I-BERT Authors (Sehoon Kim, Amir Gholami, Zhewei Yao,
# Michael Mahoney, Kurt Keutzer - UC Berkeley) and The HuggingFace Inc. team.
# Copyright (c) 20121, 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 I-BERT model. """
import math
import torch
import torch.nn as nn
import torch.utils.checkpoint
from torch.nn import CrossEntropyLoss, MSELoss
from ...activations import gelu
from ...file_utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward
from ...modeling_outputs import (
BaseModelOutputWithPastAndCrossAttentions,
BaseModelOutputWithPoolingAndCrossAttentions,
MaskedLMOutput,
MultipleChoiceModelOutput,
QuestionAnsweringModelOutput,
SequenceClassifierOutput,
TokenClassifierOutput,
)
from ...modeling_utils import PreTrainedModel, find_pruneable_heads_and_indices, prune_linear_layer
from ...utils import logging
from .configuration_ibert import IBertConfig
from .quant_modules import IntGELU, IntLayerNorm, IntSoftmax, QuantAct, QuantEmbedding, QuantLinear
logger = logging.get_logger(__name__)
_CHECKPOINT_FOR_DOC = "kssteven/ibert-roberta-base"
_CONFIG_FOR_DOC = "IBertConfig"
_TOKENIZER_FOR_DOC = "RobertaTokenizer"
IBERT_PRETRAINED_MODEL_ARCHIVE_LIST = [
"kssteven/ibert-roberta-base",
"kssteven/ibert-roberta-large",
"kssteven/ibert-roberta-large-mnli",
]
class IBertEmbeddings(nn.Module):
"""
Same as BertEmbeddings with a tiny tweak for positional embeddings indexing.
"""
def __init__(self, config):
super().__init__()
self.quant_mode = config.quant_mode
self.embedding_bit = 8
self.embedding_act_bit = 16
self.act_bit = 8
self.ln_input_bit = 22
self.ln_output_bit = 32
self.word_embeddings = QuantEmbedding(
config.vocab_size,
config.hidden_size,
padding_idx=config.pad_token_id,
weight_bit=self.embedding_bit,
quant_mode=self.quant_mode,
)
self.token_type_embeddings = QuantEmbedding(
config.type_vocab_size, config.hidden_size, weight_bit=self.embedding_bit, quant_mode=self.quant_mode
)
# 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)))
self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
# End copy
self.padding_idx = config.pad_token_id
self.position_embeddings = QuantEmbedding(
config.max_position_embeddings,
config.hidden_size,
padding_idx=self.padding_idx,
weight_bit=self.embedding_bit,
quant_mode=self.quant_mode,
)
# Integer-only addition between embeddings
self.embeddings_act1 = QuantAct(self.embedding_act_bit, quant_mode=self.quant_mode)
self.embeddings_act2 = QuantAct(self.embedding_act_bit, quant_mode=self.quant_mode)
# self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
# any TensorFlow checkpoint file
self.LayerNorm = IntLayerNorm(
config.hidden_size,
eps=config.layer_norm_eps,
output_bit=self.ln_output_bit,
quant_mode=self.quant_mode,
force_dequant=config.force_dequant,
)
self.output_activation = QuantAct(self.act_bit, quant_mode=self.quant_mode)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
def forward(
self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None, past_key_values_length=0
):
if position_ids is None:
if input_ids is not None:
# Create the position ids from the input token ids. Any padded tokens remain padded.
position_ids = create_position_ids_from_input_ids(
input_ids, self.padding_idx, past_key_values_length
).to(input_ids.device)
else:
position_ids = self.create_position_ids_from_inputs_embeds(inputs_embeds)
if input_ids is not None:
input_shape = input_ids.size()
else:
input_shape = inputs_embeds.size()[:-1]
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, inputs_embeds_scaling_factor = self.word_embeddings(input_ids)
else:
inputs_embeds_scaling_factor = None
token_type_embeddings, token_type_embeddings_scaling_factor = self.token_type_embeddings(token_type_ids)
embeddings, embeddings_scaling_factor = self.embeddings_act1(
inputs_embeds,
inputs_embeds_scaling_factor,
identity=token_type_embeddings,
identity_scaling_factor=token_type_embeddings_scaling_factor,
)
if self.position_embedding_type == "absolute":
position_embeddings, position_embeddings_scaling_factor = self.position_embeddings(position_ids)
embeddings, embeddings_scaling_factor = self.embeddings_act1(
embeddings,
embeddings_scaling_factor,
identity=position_embeddings,
identity_scaling_factor=position_embeddings_scaling_factor,
)
embeddings, embeddings_scaling_factor = self.LayerNorm(embeddings, embeddings_scaling_factor)
embeddings = self.dropout(embeddings)
embeddings, embeddings_scaling_factor = self.output_activation(embeddings, embeddings_scaling_factor)
return embeddings, embeddings_scaling_factor
def create_position_ids_from_inputs_embeds(self, inputs_embeds):
"""
We are provided embeddings directly. We cannot infer which are padded so just generate sequential position ids.
Args:
inputs_embeds: torch.Tensor
Returns: torch.Tensor
"""
input_shape = inputs_embeds.size()[:-1]
sequence_length = input_shape[1]
position_ids = torch.arange(
self.padding_idx + 1, sequence_length + self.padding_idx + 1, dtype=torch.long, device=inputs_embeds.device
)
return position_ids.unsqueeze(0).expand(input_shape)
class IBertSelfAttention(nn.Module):
def __init__(self, config):
super().__init__()
if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
raise ValueError(
f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
f"heads ({config.num_attention_heads})"
)
self.quant_mode = config.quant_mode
self.weight_bit = 8
self.bias_bit = 32
self.act_bit = 8
self.num_attention_heads = config.num_attention_heads
self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
self.all_head_size = self.num_attention_heads * self.attention_head_size
# Q, K, V Linear layers
self.query = QuantLinear(
config.hidden_size,
self.all_head_size,
bias=True,
weight_bit=self.weight_bit,
bias_bit=self.bias_bit,
quant_mode=self.quant_mode,
per_channel=True,
)
self.key = QuantLinear(
config.hidden_size,
self.all_head_size,
bias=True,
weight_bit=self.weight_bit,
bias_bit=self.bias_bit,
quant_mode=self.quant_mode,
per_channel=True,
)
self.value = QuantLinear(
config.hidden_size,
self.all_head_size,
bias=True,
weight_bit=self.weight_bit,
bias_bit=self.bias_bit,
quant_mode=self.quant_mode,
per_channel=True,
)
# Requantization (32bit -> 8bit) for Q, K, V activations
self.query_activation = QuantAct(self.act_bit, quant_mode=self.quant_mode)
self.key_activation = QuantAct(self.act_bit, quant_mode=self.quant_mode)
self.value_activation = QuantAct(self.act_bit, quant_mode=self.quant_mode)
self.output_activation = QuantAct(self.act_bit, quant_mode=self.quant_mode)
self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
assert (
self.position_embedding_type == "absolute"
), "I-BERT only supports 'absolute' for `config.position_embedding_type`"
self.softmax = IntSoftmax(self.act_bit, quant_mode=self.quant_mode, force_dequant=config.force_dequant)
def transpose_for_scores(self, x):
new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
x = x.view(*new_x_shape)
return x.permute(0, 2, 1, 3)
def forward(
self,
hidden_states,
hidden_states_scaling_factor,
attention_mask=None,
head_mask=None,
output_attentions=False,
):
# Projection
mixed_query_layer, mixed_query_layer_scaling_factor = self.query(hidden_states, hidden_states_scaling_factor)
mixed_key_layer, mixed_key_layer_scaling_factor = self.key(hidden_states, hidden_states_scaling_factor)
mixed_value_layer, mixed_value_layer_scaling_factor = self.value(hidden_states, hidden_states_scaling_factor)
# Requantization
query_layer, query_layer_scaling_factor = self.query_activation(
mixed_query_layer, mixed_query_layer_scaling_factor
)
key_layer, key_layer_scaling_factor = self.key_activation(mixed_key_layer, mixed_key_layer_scaling_factor)
value_layer, value_layer_scaling_factor = self.value_activation(
mixed_value_layer, mixed_value_layer_scaling_factor
)
# Transpose
query_layer = self.transpose_for_scores(query_layer)
key_layer = self.transpose_for_scores(key_layer)
value_layer = self.transpose_for_scores(value_layer)
# Take the dot product between "query" and "key" to get the raw attention scores.
attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
scale = math.sqrt(self.attention_head_size)
attention_scores = attention_scores / scale
if self.quant_mode:
attention_scores_scaling_factor = query_layer_scaling_factor * key_layer_scaling_factor / scale
else:
attention_scores_scaling_factor = None
if attention_mask is not None:
# Apply the attention mask is (precomputed for all layers in IBertModel forward() function)
attention_scores = attention_scores + attention_mask
# Normalize the attention scores to probabilities.
attention_probs, attention_probs_scaling_factor = self.softmax(
attention_scores, attention_scores_scaling_factor
)
# 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)
# Mask heads if we want to
if head_mask is not None:
attention_probs = attention_probs * head_mask
context_layer = torch.matmul(attention_probs, value_layer)
if attention_probs_scaling_factor is not None:
context_layer_scaling_factor = attention_probs_scaling_factor * value_layer_scaling_factor
else:
context_layer_scaling_factor = None
context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
context_layer = context_layer.view(*new_context_layer_shape)
# requantization: 32-bit -> 8-bit
context_layer, context_layer_scaling_factor = self.output_activation(
context_layer, context_layer_scaling_factor
)
outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
output_scaling_factor = (
(context_layer_scaling_factor, attention_probs_scaling_factor)
if output_attentions
else (context_layer_scaling_factor,)
)
return outputs, output_scaling_factor
class IBertSelfOutput(nn.Module):
def __init__(self, config):
super().__init__()
self.quant_mode = config.quant_mode
self.act_bit = 8
self.weight_bit = 8
self.bias_bit = 32
self.ln_input_bit = 22
self.ln_output_bit = 32
self.dense = QuantLinear(
config.hidden_size,
config.hidden_size,
bias=True,
weight_bit=self.weight_bit,
bias_bit=self.bias_bit,
quant_mode=self.quant_mode,
per_channel=True,
)
self.ln_input_act = QuantAct(self.ln_input_bit, quant_mode=self.quant_mode)
self.LayerNorm = IntLayerNorm(
config.hidden_size,
eps=config.layer_norm_eps,
output_bit=self.ln_output_bit,
quant_mode=self.quant_mode,
force_dequant=config.force_dequant,
)
self.output_activation = QuantAct(self.act_bit, quant_mode=self.quant_mode)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
def forward(self, hidden_states, hidden_states_scaling_factor, input_tensor, input_tensor_scaling_factor):
hidden_states, hidden_states_scaling_factor = self.dense(hidden_states, hidden_states_scaling_factor)
hidden_states = self.dropout(hidden_states)
hidden_states, hidden_states_scaling_factor = self.ln_input_act(
hidden_states,
hidden_states_scaling_factor,
identity=input_tensor,
identity_scaling_factor=input_tensor_scaling_factor,
)
hidden_states, hidden_states_scaling_factor = self.LayerNorm(hidden_states, hidden_states_scaling_factor)
hidden_states, hidden_states_scaling_factor = self.output_activation(
hidden_states, hidden_states_scaling_factor
)
return hidden_states, hidden_states_scaling_factor
class IBertAttention(nn.Module):
def __init__(self, config):
super().__init__()
self.quant_mode = config.quant_mode
self.self = IBertSelfAttention(config)
self.output = IBertSelfOutput(config)
self.pruned_heads = set()
def prune_heads(self, heads):
if len(heads) == 0:
return
heads, index = find_pruneable_heads_and_indices(
heads, self.self.num_attention_heads, self.self.attention_head_size, self.pruned_heads
)
# Prune linear layers
self.self.query = prune_linear_layer(self.self.query, index)
self.self.key = prune_linear_layer(self.self.key, index)
self.self.value = prune_linear_layer(self.self.value, index)
self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
# Update hyper params and store pruned heads
self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads
self.pruned_heads = self.pruned_heads.union(heads)
def forward(
self,
hidden_states,
hidden_states_scaling_factor,
attention_mask=None,
head_mask=None,
output_attentions=False,
):
self_outputs, self_outputs_scaling_factor = self.self(
hidden_states,
hidden_states_scaling_factor,
attention_mask,
head_mask,
output_attentions,
)
attention_output, attention_output_scaling_factor = self.output(
self_outputs[0], self_outputs_scaling_factor[0], hidden_states, hidden_states_scaling_factor
)
outputs = (attention_output,) + self_outputs[1:] # add attentions if we output them
outputs_scaling_factor = (attention_output_scaling_factor,) + self_outputs_scaling_factor[1:]
return outputs, outputs_scaling_factor
class IBertIntermediate(nn.Module):
def __init__(self, config):
super().__init__()
self.quant_mode = config.quant_mode
self.act_bit = 8
self.weight_bit = 8
self.bias_bit = 32
self.dense = QuantLinear(
config.hidden_size,
config.intermediate_size,
bias=True,
weight_bit=self.weight_bit,
bias_bit=self.bias_bit,
quant_mode=self.quant_mode,
per_channel=True,
)
assert config.hidden_act == "gelu", "I-BERT only supports 'gelu' for `config.hidden_act`"
self.intermediate_act_fn = IntGELU(quant_mode=self.quant_mode, force_dequant=config.force_dequant)
self.output_activation = QuantAct(self.act_bit, quant_mode=self.quant_mode)
def forward(self, hidden_states, hidden_states_scaling_factor):
hidden_states, hidden_states_scaling_factor = self.dense(hidden_states, hidden_states_scaling_factor)
hidden_states, hidden_states_scaling_factor = self.intermediate_act_fn(
hidden_states, hidden_states_scaling_factor
)
# Requantization: 32bit -> 8-bit
hidden_states, hidden_states_scaling_factor = self.output_activation(
hidden_states, hidden_states_scaling_factor
)
return hidden_states, hidden_states_scaling_factor
class IBertOutput(nn.Module):
def __init__(self, config):
super().__init__()
self.quant_mode = config.quant_mode
self.act_bit = 8
self.weight_bit = 8
self.bias_bit = 32
self.ln_input_bit = 22
self.ln_output_bit = 32
self.dense = QuantLinear(
config.intermediate_size,
config.hidden_size,
bias=True,
weight_bit=self.weight_bit,
bias_bit=self.bias_bit,
quant_mode=self.quant_mode,
per_channel=True,
)
self.ln_input_act = QuantAct(self.ln_input_bit, quant_mode=self.quant_mode)
self.LayerNorm = IntLayerNorm(
config.hidden_size,
eps=config.layer_norm_eps,
output_bit=self.ln_output_bit,
quant_mode=self.quant_mode,
force_dequant=config.force_dequant,
)
self.output_activation = QuantAct(self.act_bit, quant_mode=self.quant_mode)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
def forward(self, hidden_states, hidden_states_scaling_factor, input_tensor, input_tensor_scaling_factor):
hidden_states, hidden_states_scaling_factor = self.dense(hidden_states, hidden_states_scaling_factor)
hidden_states = self.dropout(hidden_states)
hidden_states, hidden_states_scaling_factor = self.ln_input_act(
hidden_states,
hidden_states_scaling_factor,
identity=input_tensor,
identity_scaling_factor=input_tensor_scaling_factor,
)
hidden_states, hidden_states_scaling_factor = self.LayerNorm(hidden_states, hidden_states_scaling_factor)
hidden_states, hidden_states_scaling_factor = self.output_activation(
hidden_states, hidden_states_scaling_factor
)
return hidden_states, hidden_states_scaling_factor
class IBertLayer(nn.Module):
def __init__(self, config):
super().__init__()
self.quant_mode = config.quant_mode
self.act_bit = 8
self.seq_len_dim = 1
self.attention = IBertAttention(config)
self.intermediate = IBertIntermediate(config)
self.output = IBertOutput(config)
self.pre_intermediate_act = QuantAct(self.act_bit, quant_mode=self.quant_mode)
self.pre_output_act = QuantAct(self.act_bit, quant_mode=self.quant_mode)
def forward(
self,
hidden_states,
hidden_states_scaling_factor,
attention_mask=None,
head_mask=None,
output_attentions=False,
):
self_attention_outputs, self_attention_outputs_scaling_factor = self.attention(
hidden_states,
hidden_states_scaling_factor,
attention_mask,
head_mask,
output_attentions=output_attentions,
)
attention_output = self_attention_outputs[0]
attention_output_scaling_factor = self_attention_outputs_scaling_factor[0]
outputs = self_attention_outputs[1:] # add self attentions if we output attention weights
layer_output, layer_output_scaling_factor = self.feed_forward_chunk(
attention_output, attention_output_scaling_factor
)
outputs = (layer_output,) + outputs
return outputs
def feed_forward_chunk(self, attention_output, attention_output_scaling_factor):
attention_output, attention_output_scaling_factor = self.pre_intermediate_act(
attention_output, attention_output_scaling_factor
)
intermediate_output, intermediate_output_scaling_factor = self.intermediate(
attention_output, attention_output_scaling_factor
)
intermediate_output, intermediate_output_scaling_factor = self.pre_output_act(
intermediate_output, intermediate_output_scaling_factor
)
layer_output, layer_output_scaling_factor = self.output(
intermediate_output, intermediate_output_scaling_factor, attention_output, attention_output_scaling_factor
)
return layer_output, layer_output_scaling_factor
class IBertEncoder(nn.Module):
def __init__(self, config):
super().__init__()
self.config = config
self.quant_mode = config.quant_mode
self.layer = nn.ModuleList([IBertLayer(config) for _ in range(config.num_hidden_layers)])
def forward(
self,
hidden_states,
hidden_states_scaling_factor,
attention_mask=None,
head_mask=None,
output_attentions=False,
output_hidden_states=False,
return_dict=True,
):
all_hidden_states = () if output_hidden_states else None
all_self_attentions = () if output_attentions else None
all_cross_attentions = None # `config.add_cross_attention` is not supported
next_decoder_cache = None # `config.use_cache` is not supported
for i, layer_module in enumerate(self.layer):
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
layer_head_mask = head_mask[i] if head_mask is not None else None
if getattr(self.config, "gradient_checkpointing", False) and self.training:
raise NotImplementedError("gradient checkpointing is not currently supported")
else:
layer_outputs = layer_module(
hidden_states,
hidden_states_scaling_factor,
attention_mask,
layer_head_mask,
output_attentions,
)
hidden_states = layer_outputs[0]
if output_attentions:
all_self_attentions = all_self_attentions + (layer_outputs[1],)
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
if not return_dict:
return tuple(
v
for v in [
hidden_states,
next_decoder_cache,
all_hidden_states,
all_self_attentions,
all_cross_attentions,
]
if v is not None
)
return BaseModelOutputWithPastAndCrossAttentions(
last_hidden_state=hidden_states,
past_key_values=next_decoder_cache,
hidden_states=all_hidden_states,
attentions=all_self_attentions,
cross_attentions=all_cross_attentions,
)
class IBertPooler(nn.Module):
def __init__(self, config):
super().__init__()
self.quant_mode = config.quant_mode
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 IBertPreTrainedModel(PreTrainedModel):
"""
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
models.
"""
config_class = IBertConfig
base_model_prefix = "ibert"
def _init_weights(self, module):
""" Initialize the weights """
if isinstance(module, (QuantLinear, nn.Linear)):
# 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)
if module.bias is not None:
module.bias.data.zero_()
elif isinstance(module, (QuantEmbedding, nn.Embedding)):
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
if module.padding_idx is not None:
module.weight.data[module.padding_idx].zero_()
elif isinstance(module, (IntLayerNorm, nn.LayerNorm)):
module.bias.data.zero_()
module.weight.data.fill_(1.0)
def resize_token_embeddings(self, new_num_tokens=None):
raise NotImplementedError("`resize_token_embeddings` is not supported for I-BERT.")
IBERT_START_DOCSTRING = r"""
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.
Parameters:
config (:class:`~transformers.IBertConfig`): 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.
"""
IBERT_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.RobertaTokenizer`. 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 I-BERT Model transformer outputting raw hidden-states without any specific head on top.",
IBERT_START_DOCSTRING,
)
class IBertModel(IBertPreTrainedModel):
"""
The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of
cross-attention is added between the self-attention layers, following the architecture described in `Attention is
all you need <https://arxiv.org/abs/1706.03762>`__ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,
Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.
"""
_keys_to_ignore_on_load_missing = [r"position_ids"]
def __init__(self, config, add_pooling_layer=True):
super().__init__(config)
self.config = config
self.quant_mode = config.quant_mode
self.embeddings = IBertEmbeddings(config)
self.encoder = IBertEncoder(config)
self.pooler = IBertPooler(config) if add_pooling_layer else None
self.init_weights()
def get_input_embeddings(self):
return self.embeddings.word_embeddings
def set_input_embeddings(self, value):
self.embeddings.word_embeddings = value
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(IBERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=BaseModelOutputWithPoolingAndCrossAttentions,
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()
batch_size, seq_length = input_shape
elif inputs_embeds is not None:
input_shape = inputs_embeds.size()[:-1]
batch_size, seq_length = input_shape
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(((batch_size, seq_length)), device=device)
if token_type_ids is None:
token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
# We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
# ourselves in which case we just need to make it broadcastable to all heads.
extended_attention_mask: torch.Tensor = 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, embedding_output_scaling_factor = self.embeddings(
input_ids=input_ids,
position_ids=position_ids,
token_type_ids=token_type_ids,
inputs_embeds=inputs_embeds,
)
encoder_outputs = self.encoder(
embedding_output,
embedding_output_scaling_factor,
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 self.pooler is not None else None
if not return_dict:
return (sequence_output, pooled_output) + encoder_outputs[1:]
return BaseModelOutputWithPoolingAndCrossAttentions(
last_hidden_state=sequence_output,
pooler_output=pooled_output,
past_key_values=encoder_outputs.past_key_values,
hidden_states=encoder_outputs.hidden_states,
attentions=encoder_outputs.attentions,
cross_attentions=encoder_outputs.cross_attentions,
)
class IBertLMHead(nn.Module):
"""I-BERT Head for masked language modeling."""
def __init__(self, config):
super().__init__()
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
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, features, **kwargs):
x = self.dense(features)
x = gelu(x)
x = self.layer_norm(x)
# project back to size of vocabulary with bias
x = self.decoder(x)
return x
[docs]@add_start_docstrings(
"""
I-BERT Model transformer with a sequence classification/regression head on top (a linear layer on top of the pooled
output) e.g. for GLUE tasks.
""",
IBERT_START_DOCSTRING,
)
class IBertForSequenceClassification(IBertPreTrainedModel):
_keys_to_ignore_on_load_missing = [r"position_ids"]
def __init__(self, config):
super().__init__(config)
self.num_labels = config.num_labels
self.ibert = IBertModel(config, add_pooling_layer=False)
self.classifier = IBertClassificationHead(config)
self.init_weights()
[docs] @add_start_docstrings_to_model_forward(IBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
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.ibert(
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.classifier(sequence_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(
"""
I-BERT 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.
""",
IBERT_START_DOCSTRING,
)
class IBertForTokenClassification(IBertPreTrainedModel):
_keys_to_ignore_on_load_unexpected = [r"pooler"]
_keys_to_ignore_on_load_missing = [r"position_ids"]
def __init__(self, config):
super().__init__(config)
self.num_labels = config.num_labels
self.ibert = IBertModel(config, add_pooling_layer=False)
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(IBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
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.ibert(
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,
)
class IBertClassificationHead(nn.Module):
"""Head for sentence-level classification tasks."""
def __init__(self, config):
super().__init__()
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
self.dropout = nn.Dropout(config.hidden_dropout_prob)
self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
def forward(self, features, **kwargs):
hidden_states = features[:, 0, :] # take <s> token (equiv. to [CLS])
hidden_states = self.dropout(hidden_states)
hidden_states = self.dense(hidden_states)
hidden_states = torch.tanh(hidden_states)
hidden_states = self.dropout(hidden_states)
hidden_states = self.out_proj(hidden_states)
return hidden_states
[docs]@add_start_docstrings(
"""
I-BERT 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`).
""",
IBERT_START_DOCSTRING,
)
class IBertForQuestionAnswering(IBertPreTrainedModel):
_keys_to_ignore_on_load_unexpected = [r"pooler"]
_keys_to_ignore_on_load_missing = [r"position_ids"]
def __init__(self, config):
super().__init__(config)
self.num_labels = config.num_labels
self.ibert = IBertModel(config, add_pooling_layer=False)
self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
self.init_weights()
[docs] @add_start_docstrings_to_model_forward(IBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
tokenizer_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
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 (:obj:`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 (:obj:`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.ibert(
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,
)
def create_position_ids_from_input_ids(input_ids, padding_idx, past_key_values_length=0):
"""
Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding symbols
are ignored. This is modified from fairseq's `utils.make_positions`.
Args:
input_ids (:obj:`torch.LongTensor`):
Indices of input sequence tokens in the vocabulary.
Returns: torch.Tensor
"""
# The series of casts and type-conversions here are carefully balanced to both work with ONNX export and XLA.
mask = input_ids.ne(padding_idx).int()
incremental_indices = (torch.cumsum(mask, dim=1).type_as(mask) + past_key_values_length) * mask
return incremental_indices.long() + padding_idx