# coding=utf-8
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" TF 2.0 RoBERTa model. """
import math
import warnings
from typing import Optional, Tuple, Union
import numpy as np
import tensorflow as tf
from ...activations_tf import get_tf_activation
from ...file_utils import (
DUMMY_INPUTS,
MULTIPLE_CHOICE_DUMMY_INPUTS,
add_code_sample_docstrings,
add_start_docstrings,
add_start_docstrings_to_model_forward,
)
from ...modeling_tf_outputs import (
TFBaseModelOutputWithPastAndCrossAttentions,
TFBaseModelOutputWithPoolingAndCrossAttentions,
TFCausalLMOutputWithCrossAttentions,
TFMaskedLMOutput,
TFMultipleChoiceModelOutput,
TFQuestionAnsweringModelOutput,
TFSequenceClassifierOutput,
TFTokenClassifierOutput,
)
from ...modeling_tf_utils import (
TFCausalLanguageModelingLoss,
TFMaskedLanguageModelingLoss,
TFModelInputType,
TFMultipleChoiceLoss,
TFPreTrainedModel,
TFQuestionAnsweringLoss,
TFSequenceClassificationLoss,
TFTokenClassificationLoss,
get_initializer,
input_processing,
keras_serializable,
shape_list,
)
from ...utils import logging
from .configuration_roberta import RobertaConfig
logger = logging.get_logger(__name__)
_CHECKPOINT_FOR_DOC = "roberta-base"
_CONFIG_FOR_DOC = "RobertaConfig"
_TOKENIZER_FOR_DOC = "RobertaTokenizer"
TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST = [
"roberta-base",
"roberta-large",
"roberta-large-mnli",
"distilroberta-base",
# See all RoBERTa models at https://huggingface.co/models?filter=roberta
]
class TFRobertaEmbeddings(tf.keras.layers.Layer):
"""
Same as BertEmbeddings with a tiny tweak for positional embeddings indexing.
"""
def __init__(self, config, **kwargs):
super().__init__(**kwargs)
self.padding_idx = 1
self.vocab_size = config.vocab_size
self.type_vocab_size = config.type_vocab_size
self.hidden_size = config.hidden_size
self.max_position_embeddings = config.max_position_embeddings
self.initializer_range = config.initializer_range
self.LayerNorm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
self.dropout = tf.keras.layers.Dropout(rate=config.hidden_dropout_prob)
def build(self, input_shape: tf.TensorShape):
with tf.name_scope("word_embeddings"):
self.weight = self.add_weight(
name="weight",
shape=[self.vocab_size, self.hidden_size],
initializer=get_initializer(self.initializer_range),
)
with tf.name_scope("token_type_embeddings"):
self.token_type_embeddings = self.add_weight(
name="embeddings",
shape=[self.type_vocab_size, self.hidden_size],
initializer=get_initializer(self.initializer_range),
)
with tf.name_scope("position_embeddings"):
self.position_embeddings = self.add_weight(
name="embeddings",
shape=[self.max_position_embeddings, self.hidden_size],
initializer=get_initializer(self.initializer_range),
)
super().build(input_shape)
def create_position_ids_from_input_ids(self, input_ids, 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: tf.Tensor
Returns: tf.Tensor
"""
mask = tf.cast(tf.math.not_equal(input_ids, self.padding_idx), dtype=input_ids.dtype)
incremental_indices = (tf.math.cumsum(mask, axis=1) + past_key_values_length) * mask
return incremental_indices + self.padding_idx
def call(
self,
input_ids=None,
position_ids=None,
token_type_ids=None,
inputs_embeds=None,
past_key_values_length=0,
training=False,
):
"""
Applies embedding based on inputs tensor.
Returns:
final_embeddings (`tf.Tensor`): output embedding tensor.
"""
assert not (input_ids is None and inputs_embeds is None)
if input_ids is not None:
inputs_embeds = tf.gather(params=self.weight, indices=input_ids)
input_shape = shape_list(inputs_embeds)[:-1]
if token_type_ids is None:
token_type_ids = tf.fill(dims=input_shape, value=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 = self.create_position_ids_from_input_ids(
input_ids=input_ids, past_key_values_length=past_key_values_length
)
else:
position_ids = tf.expand_dims(
tf.range(start=self.padding_idx + 1, limit=input_shape[-1] + self.padding_idx + 1), axis=0
)
position_embeds = tf.gather(params=self.position_embeddings, indices=position_ids)
token_type_embeds = tf.gather(params=self.token_type_embeddings, indices=token_type_ids)
final_embeddings = inputs_embeds + position_embeds + token_type_embeds
final_embeddings = self.LayerNorm(inputs=final_embeddings)
final_embeddings = self.dropout(inputs=final_embeddings, training=training)
return final_embeddings
# Copied from transformers.models.bert.modeling_tf_bert.TFBertPooler with Bert->Roberta
class TFRobertaPooler(tf.keras.layers.Layer):
def __init__(self, config: RobertaConfig, **kwargs):
super().__init__(**kwargs)
self.dense = tf.keras.layers.Dense(
units=config.hidden_size,
kernel_initializer=get_initializer(config.initializer_range),
activation="tanh",
name="dense",
)
def call(self, hidden_states: tf.Tensor) -> tf.Tensor:
# 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(inputs=first_token_tensor)
return pooled_output
# Copied from transformers.models.bert.modeling_tf_bert.TFBertSelfAttention with Bert->Roberta
class TFRobertaSelfAttention(tf.keras.layers.Layer):
def __init__(self, config: RobertaConfig, **kwargs):
super().__init__(**kwargs)
if config.hidden_size % config.num_attention_heads != 0:
raise ValueError(
f"The hidden size ({config.hidden_size}) is not a multiple of the number "
f"of attention heads ({config.num_attention_heads})"
)
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
self.sqrt_att_head_size = math.sqrt(self.attention_head_size)
self.query = tf.keras.layers.Dense(
units=self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="query"
)
self.key = tf.keras.layers.Dense(
units=self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="key"
)
self.value = tf.keras.layers.Dense(
units=self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="value"
)
self.dropout = tf.keras.layers.Dropout(rate=config.attention_probs_dropout_prob)
self.is_decoder = config.is_decoder
def transpose_for_scores(self, tensor: tf.Tensor, batch_size: int) -> tf.Tensor:
# Reshape from [batch_size, seq_length, all_head_size] to [batch_size, seq_length, num_attention_heads, attention_head_size]
tensor = tf.reshape(tensor=tensor, shape=(batch_size, -1, self.num_attention_heads, self.attention_head_size))
# Transpose the tensor from [batch_size, seq_length, num_attention_heads, attention_head_size] to [batch_size, num_attention_heads, seq_length, attention_head_size]
return tf.transpose(tensor, perm=[0, 2, 1, 3])
def call(
self,
hidden_states: tf.Tensor,
attention_mask: tf.Tensor,
head_mask: tf.Tensor,
encoder_hidden_states: tf.Tensor,
encoder_attention_mask: tf.Tensor,
past_key_value: Tuple[tf.Tensor],
output_attentions: bool,
training: bool = False,
) -> Tuple[tf.Tensor]:
batch_size = shape_list(hidden_states)[0]
mixed_query_layer = self.query(inputs=hidden_states)
# If this is instantiated as a cross-attention module, the keys
# and values come from an encoder; the attention mask needs to be
# such that the encoder's padding tokens are not attended to.
is_cross_attention = encoder_hidden_states is not None
if is_cross_attention and past_key_value is not None:
# reuse k,v, cross_attentions
key_layer = past_key_value[0]
value_layer = past_key_value[1]
attention_mask = encoder_attention_mask
elif is_cross_attention:
key_layer = self.transpose_for_scores(self.key(inputs=encoder_hidden_states), batch_size)
value_layer = self.transpose_for_scores(self.value(inputs=encoder_hidden_states), batch_size)
attention_mask = encoder_attention_mask
elif past_key_value is not None:
key_layer = self.transpose_for_scores(self.key(inputs=hidden_states), batch_size)
value_layer = self.transpose_for_scores(self.value(inputs=hidden_states), batch_size)
key_layer = tf.concatenate([past_key_value[0], key_layer], dim=2)
value_layer = tf.concatenate([past_key_value[1], value_layer], dim=2)
else:
key_layer = self.transpose_for_scores(self.key(inputs=hidden_states), batch_size)
value_layer = self.transpose_for_scores(self.value(inputs=hidden_states), batch_size)
query_layer = self.transpose_for_scores(mixed_query_layer, batch_size)
if self.is_decoder:
# if cross_attention save Tuple(tf.Tensor, tf.Tensor) of all cross attention key/value_states.
# Further calls to cross_attention layer can then reuse all cross-attention
# key/value_states (first "if" case)
# if uni-directional self-attention (decoder) save Tuple(tf.Tensor, tf.Tensor) of
# all previous decoder key/value_states. Further calls to uni-directional self-attention
# can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
# if encoder bi-directional self-attention `past_key_value` is always `None`
past_key_value = (key_layer, value_layer)
# Take the dot product between "query" and "key" to get the raw attention scores.
# (batch size, num_heads, seq_len_q, seq_len_k)
attention_scores = tf.matmul(query_layer, key_layer, transpose_b=True)
dk = tf.cast(self.sqrt_att_head_size, dtype=attention_scores.dtype)
attention_scores = tf.divide(attention_scores, dk)
if attention_mask is not None:
# Apply the attention mask is (precomputed for all layers in TFRobertaModel call() function)
attention_scores = tf.add(attention_scores, attention_mask)
# Normalize the attention scores to probabilities.
attention_probs = tf.nn.softmax(logits=attention_scores, axis=-1)
# 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(inputs=attention_probs, training=training)
# Mask heads if we want to
if head_mask is not None:
attention_probs = tf.multiply(attention_probs, head_mask)
attention_output = tf.matmul(attention_probs, value_layer)
attention_output = tf.transpose(attention_output, perm=[0, 2, 1, 3])
# (batch_size, seq_len_q, all_head_size)
attention_output = tf.reshape(tensor=attention_output, shape=(batch_size, -1, self.all_head_size))
outputs = (attention_output, attention_probs) if output_attentions else (attention_output,)
if self.is_decoder:
outputs = outputs + (past_key_value,)
return outputs
# Copied from transformers.models.bert.modeling_tf_bert.TFBertSelfOutput with Bert->Roberta
class TFRobertaSelfOutput(tf.keras.layers.Layer):
def __init__(self, config: RobertaConfig, **kwargs):
super().__init__(**kwargs)
self.dense = tf.keras.layers.Dense(
units=config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
)
self.LayerNorm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
self.dropout = tf.keras.layers.Dropout(rate=config.hidden_dropout_prob)
def call(self, hidden_states: tf.Tensor, input_tensor: tf.Tensor, training: bool = False) -> tf.Tensor:
hidden_states = self.dense(inputs=hidden_states)
hidden_states = self.dropout(inputs=hidden_states, training=training)
hidden_states = self.LayerNorm(inputs=hidden_states + input_tensor)
return hidden_states
# Copied from transformers.models.bert.modeling_tf_bert.TFBertAttention with Bert->Roberta
class TFRobertaAttention(tf.keras.layers.Layer):
def __init__(self, config: RobertaConfig, **kwargs):
super().__init__(**kwargs)
self.self_attention = TFRobertaSelfAttention(config, name="self")
self.dense_output = TFRobertaSelfOutput(config, name="output")
def prune_heads(self, heads):
raise NotImplementedError
def call(
self,
input_tensor: tf.Tensor,
attention_mask: tf.Tensor,
head_mask: tf.Tensor,
encoder_hidden_states: tf.Tensor,
encoder_attention_mask: tf.Tensor,
past_key_value: Tuple[tf.Tensor],
output_attentions: bool,
training: bool = False,
) -> Tuple[tf.Tensor]:
self_outputs = self.self_attention(
hidden_states=input_tensor,
attention_mask=attention_mask,
head_mask=head_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_value=past_key_value,
output_attentions=output_attentions,
training=training,
)
attention_output = self.dense_output(
hidden_states=self_outputs[0], input_tensor=input_tensor, training=training
)
# add attentions (possibly with past_key_value) if we output them
outputs = (attention_output,) + self_outputs[1:]
return outputs
# Copied from transformers.models.bert.modeling_tf_bert.TFBertIntermediate with Bert->Roberta
class TFRobertaIntermediate(tf.keras.layers.Layer):
def __init__(self, config: RobertaConfig, **kwargs):
super().__init__(**kwargs)
self.dense = tf.keras.layers.Dense(
units=config.intermediate_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
)
if isinstance(config.hidden_act, str):
self.intermediate_act_fn = get_tf_activation(config.hidden_act)
else:
self.intermediate_act_fn = config.hidden_act
def call(self, hidden_states: tf.Tensor) -> tf.Tensor:
hidden_states = self.dense(inputs=hidden_states)
hidden_states = self.intermediate_act_fn(hidden_states)
return hidden_states
# Copied from transformers.models.bert.modeling_tf_bert.TFBertOutput with Bert->Roberta
class TFRobertaOutput(tf.keras.layers.Layer):
def __init__(self, config: RobertaConfig, **kwargs):
super().__init__(**kwargs)
self.dense = tf.keras.layers.Dense(
units=config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
)
self.LayerNorm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
self.dropout = tf.keras.layers.Dropout(rate=config.hidden_dropout_prob)
def call(self, hidden_states: tf.Tensor, input_tensor: tf.Tensor, training: bool = False) -> tf.Tensor:
hidden_states = self.dense(inputs=hidden_states)
hidden_states = self.dropout(inputs=hidden_states, training=training)
hidden_states = self.LayerNorm(inputs=hidden_states + input_tensor)
return hidden_states
# Copied from transformers.models.bert.modeling_tf_bert.TFBertLayer with Bert->Roberta
class TFRobertaLayer(tf.keras.layers.Layer):
def __init__(self, config: RobertaConfig, **kwargs):
super().__init__(**kwargs)
self.attention = TFRobertaAttention(config, name="attention")
self.is_decoder = config.is_decoder
self.add_cross_attention = config.add_cross_attention
if self.add_cross_attention:
if not self.is_decoder:
raise ValueError(f"{self} should be used as a decoder model if cross attention is added")
self.crossattention = TFRobertaAttention(config, name="crossattention")
self.intermediate = TFRobertaIntermediate(config, name="intermediate")
self.bert_output = TFRobertaOutput(config, name="output")
def call(
self,
hidden_states: tf.Tensor,
attention_mask: tf.Tensor,
head_mask: tf.Tensor,
encoder_hidden_states: Optional[tf.Tensor],
encoder_attention_mask: Optional[tf.Tensor],
past_key_value: Optional[Tuple[tf.Tensor]],
output_attentions: bool,
training: bool = False,
) -> Tuple[tf.Tensor]:
# decoder uni-directional self-attention cached key/values tuple is at positions 1,2
self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
self_attention_outputs = self.attention(
input_tensor=hidden_states,
attention_mask=attention_mask,
head_mask=head_mask,
encoder_hidden_states=None,
encoder_attention_mask=None,
past_key_value=self_attn_past_key_value,
output_attentions=output_attentions,
training=training,
)
attention_output = self_attention_outputs[0]
# if decoder, the last output is tuple of self-attn cache
if self.is_decoder:
outputs = self_attention_outputs[1:-1]
present_key_value = self_attention_outputs[-1]
else:
outputs = self_attention_outputs[1:] # add self attentions if we output attention weights
cross_attn_present_key_value = None
if self.is_decoder and encoder_hidden_states is not None:
if not hasattr(self, "crossattention"):
raise ValueError(
f"If `encoder_hidden_states` are passed, {self} has to be instantiated with cross-attention layers "
"by setting `config.add_cross_attention=True`"
)
# cross_attn cached key/values tuple is at positions 3,4 of past_key_value tuple
cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
cross_attention_outputs = self.crossattention(
input_tensor=attention_output,
attention_mask=attention_mask,
head_mask=head_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_value=cross_attn_past_key_value,
output_attentions=output_attentions,
training=training,
)
attention_output = cross_attention_outputs[0]
outputs = outputs + cross_attention_outputs[1:-1] # add cross attentions if we output attention weights
# add cross-attn cache to positions 3,4 of present_key_value tuple
cross_attn_present_key_value = cross_attention_outputs[-1]
present_key_value = present_key_value + cross_attn_present_key_value
intermediate_output = self.intermediate(hidden_states=attention_output)
layer_output = self.bert_output(
hidden_states=intermediate_output, input_tensor=attention_output, training=training
)
outputs = (layer_output,) + outputs # add attentions if we output them
# if decoder, return the attn key/values as the last output
if self.is_decoder:
outputs = outputs + (present_key_value,)
return outputs
# Copied from transformers.models.bert.modeling_tf_bert.TFBertEncoder with Bert->Roberta
class TFRobertaEncoder(tf.keras.layers.Layer):
def __init__(self, config: RobertaConfig, **kwargs):
super().__init__(**kwargs)
self.config = config
self.layer = [TFRobertaLayer(config, name=f"layer_._{i}") for i in range(config.num_hidden_layers)]
def call(
self,
hidden_states: tf.Tensor,
attention_mask: tf.Tensor,
head_mask: tf.Tensor,
encoder_hidden_states: Optional[tf.Tensor],
encoder_attention_mask: Optional[tf.Tensor],
past_key_values: Optional[Tuple[Tuple[tf.Tensor]]],
use_cache: Optional[bool],
output_attentions: bool,
output_hidden_states: bool,
return_dict: bool,
training: bool = False,
) -> Union[TFBaseModelOutputWithPastAndCrossAttentions, Tuple[tf.Tensor]]:
all_hidden_states = () if output_hidden_states else None
all_attentions = () if output_attentions else None
all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
next_decoder_cache = () if use_cache else None
for i, layer_module in enumerate(self.layer):
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
past_key_value = past_key_values[i] if past_key_values is not None else None
layer_outputs = layer_module(
hidden_states=hidden_states,
attention_mask=attention_mask,
head_mask=head_mask[i],
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_value=past_key_value,
output_attentions=output_attentions,
training=training,
)
hidden_states = layer_outputs[0]
if use_cache:
next_decoder_cache += (layer_outputs[-1],)
if output_attentions:
all_attentions = all_attentions + (layer_outputs[1],)
if self.config.add_cross_attention and encoder_hidden_states is not None:
all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
# Add last layer
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
if not return_dict:
return tuple(
v for v in [hidden_states, all_hidden_states, all_attentions, all_cross_attentions] if v is not None
)
return TFBaseModelOutputWithPastAndCrossAttentions(
last_hidden_state=hidden_states,
past_key_values=next_decoder_cache,
hidden_states=all_hidden_states,
attentions=all_attentions,
cross_attentions=all_cross_attentions,
)
@keras_serializable
class TFRobertaMainLayer(tf.keras.layers.Layer):
config_class = RobertaConfig
def __init__(self, config, add_pooling_layer=True, **kwargs):
super().__init__(**kwargs)
self.config = config
self.is_decoder = config.is_decoder
self.num_hidden_layers = config.num_hidden_layers
self.initializer_range = config.initializer_range
self.output_attentions = config.output_attentions
self.output_hidden_states = config.output_hidden_states
self.return_dict = config.use_return_dict
self.encoder = TFRobertaEncoder(config, name="encoder")
self.pooler = TFRobertaPooler(config, name="pooler") if add_pooling_layer else None
# The embeddings must be the last declaration in order to follow the weights order
self.embeddings = TFRobertaEmbeddings(config, name="embeddings")
# Copied from transformers.models.bert.modeling_tf_bert.TFBertMainLayer.get_input_embeddings
def get_input_embeddings(self) -> tf.keras.layers.Layer:
return self.embeddings
# Copied from transformers.models.bert.modeling_tf_bert.TFBertMainLayer.set_input_embeddings
def set_input_embeddings(self, value: tf.Variable):
self.embeddings.weight = value
self.embeddings.vocab_size = shape_list(value)[0]
# Copied from transformers.models.bert.modeling_tf_bert.TFBertMainLayer._prune_heads
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
"""
raise NotImplementedError
# Copied from transformers.models.bert.modeling_tf_bert.TFBertMainLayer.call
def call(
self,
input_ids: Optional[TFModelInputType] = None,
attention_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
token_type_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
position_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
head_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
inputs_embeds: Optional[Union[np.ndarray, tf.Tensor]] = None,
encoder_hidden_states: Optional[Union[np.ndarray, tf.Tensor]] = None,
encoder_attention_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
past_key_values: Optional[Tuple[Tuple[Union[np.ndarray, tf.Tensor]]]] = None,
use_cache: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
training: bool = False,
**kwargs,
) -> Union[TFBaseModelOutputWithPoolingAndCrossAttentions, Tuple[tf.Tensor]]:
inputs = input_processing(
func=self.call,
config=self.config,
input_ids=input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
position_ids=position_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
use_cache=use_cache,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
training=training,
kwargs_call=kwargs,
)
if not self.config.is_decoder:
inputs["use_cache"] = False
if inputs["input_ids"] is not None and inputs["inputs_embeds"] is not None:
raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
elif inputs["input_ids"] is not None:
input_shape = shape_list(inputs["input_ids"])
elif inputs["inputs_embeds"] is not None:
input_shape = shape_list(inputs["inputs_embeds"])[:-1]
else:
raise ValueError("You have to specify either input_ids or inputs_embeds")
batch_size, seq_length = input_shape
if inputs["past_key_values"] is None:
past_key_values_length = 0
inputs["past_key_values"] = [None] * len(self.encoder.layer)
else:
past_key_values_length = shape_list(inputs["past_key_values"][0][0])[-2]
if inputs["attention_mask"] is None:
inputs["attention_mask"] = tf.fill(dims=(batch_size, seq_length + past_key_values_length), value=1)
if inputs["token_type_ids"] is None:
inputs["token_type_ids"] = tf.fill(dims=input_shape, value=0)
embedding_output = self.embeddings(
input_ids=inputs["input_ids"],
position_ids=inputs["position_ids"],
token_type_ids=inputs["token_type_ids"],
inputs_embeds=inputs["inputs_embeds"],
past_key_values_length=past_key_values_length,
training=inputs["training"],
)
# We create a 3D attention mask from a 2D tensor mask.
# Sizes are [batch_size, 1, 1, to_seq_length]
# So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
# this attention mask is more simple than the triangular masking of causal attention
# used in OpenAI GPT, we just need to prepare the broadcast dimension here.
attention_mask_shape = shape_list(inputs["attention_mask"])
mask_seq_length = seq_length + past_key_values_length
# Copied from `modeling_tf_t5.py`
# Provided a padding mask of dimensions [batch_size, mask_seq_length]
# - if the model is a decoder, apply a causal mask in addition to the padding mask
# - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, mask_seq_length, mask_seq_length]
if self.is_decoder:
seq_ids = tf.range(mask_seq_length)
causal_mask = tf.less_equal(
tf.tile(seq_ids[None, None, :], (batch_size, mask_seq_length, 1)),
seq_ids[None, :, None],
)
causal_mask = tf.cast(causal_mask, dtype=inputs["attention_mask"].dtype)
extended_attention_mask = causal_mask * inputs["attention_mask"][:, None, :]
attention_mask_shape = shape_list(extended_attention_mask)
extended_attention_mask = tf.reshape(
extended_attention_mask, (attention_mask_shape[0], 1, attention_mask_shape[1], attention_mask_shape[2])
)
else:
extended_attention_mask = tf.reshape(
inputs["attention_mask"], (attention_mask_shape[0], 1, 1, attention_mask_shape[1])
)
# Since attention_mask is 1.0 for positions we want to attend and 0.0 for
# masked positions, this operation will create a tensor which is 0.0 for
# positions we want to attend and -10000.0 for masked positions.
# Since we are adding it to the raw scores before the softmax, this is
# effectively the same as removing these entirely.
extended_attention_mask = tf.cast(extended_attention_mask, dtype=embedding_output.dtype)
one_cst = tf.constant(1.0, dtype=embedding_output.dtype)
ten_thousand_cst = tf.constant(-10000.0, dtype=embedding_output.dtype)
extended_attention_mask = tf.multiply(tf.subtract(one_cst, extended_attention_mask), ten_thousand_cst)
# Copied from `modeling_tf_t5.py` with -1e9 -> -10000
if self.is_decoder and inputs["encoder_attention_mask"] is not None:
# If a 2D ou 3D attention mask is provided for the cross-attention
# we need to make broadcastable to [batch_size, num_heads, mask_seq_length, mask_seq_length]
# we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
inputs["encoder_attention_mask"] = tf.cast(
inputs["encoder_attention_mask"], dtype=extended_attention_mask.dtype
)
num_dims_encoder_attention_mask = len(shape_list(inputs["encoder_attention_mask"]))
if num_dims_encoder_attention_mask == 3:
encoder_extended_attention_mask = inputs["encoder_attention_mask"][:, None, :, :]
if num_dims_encoder_attention_mask == 2:
encoder_extended_attention_mask = inputs["encoder_attention_mask"][:, None, None, :]
# T5 has a mask that can compare sequence ids, we can simulate this here with this transposition
# Cf. https://github.com/tensorflow/mesh/blob/8d2465e9bc93129b913b5ccc6a59aa97abd96ec6/mesh_tensorflow/transformer/transformer_layers.py#L270
# encoder_extended_attention_mask = tf.math.equal(encoder_extended_attention_mask,
# tf.transpose(encoder_extended_attention_mask, perm=(-1, -2)))
encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -10000.0
else:
encoder_extended_attention_mask = None
# 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]
if inputs["head_mask"] is not None:
raise NotImplementedError
else:
inputs["head_mask"] = [None] * self.config.num_hidden_layers
encoder_outputs = self.encoder(
hidden_states=embedding_output,
attention_mask=extended_attention_mask,
head_mask=inputs["head_mask"],
encoder_hidden_states=inputs["encoder_hidden_states"],
encoder_attention_mask=encoder_extended_attention_mask,
past_key_values=inputs["past_key_values"],
use_cache=inputs["use_cache"],
output_attentions=inputs["output_attentions"],
output_hidden_states=inputs["output_hidden_states"],
return_dict=inputs["return_dict"],
training=inputs["training"],
)
sequence_output = encoder_outputs[0]
pooled_output = self.pooler(hidden_states=sequence_output) if self.pooler is not None else None
if not inputs["return_dict"]:
return (
sequence_output,
pooled_output,
) + encoder_outputs[1:]
return TFBaseModelOutputWithPoolingAndCrossAttentions(
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 TFRobertaPreTrainedModel(TFPreTrainedModel):
"""
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
models.
"""
config_class = RobertaConfig
base_model_prefix = "roberta"
@property
# Copied from transformers.models.bert.modeling_tf_bert.TFBertPreTrainedModel.dummy_inputs
def dummy_inputs(self):
"""
Dummy inputs to build the network.
Returns:
`Dict[str, tf.Tensor]`: The dummy inputs.
"""
dummy = {"input_ids": tf.constant(DUMMY_INPUTS)}
# Add `encoder_hidden_states` to make the cross-attention layers' weights initialized
if self.config.add_cross_attention:
batch_size, seq_len = tf.constant(DUMMY_INPUTS).shape
shape = (batch_size, seq_len) + (self.config.hidden_size,)
h = tf.random.uniform(shape=shape)
dummy["encoder_hidden_states"] = h
return dummy
@tf.function(
input_signature=[
{
"input_ids": tf.TensorSpec((None, None), tf.int32, name="input_ids"),
"attention_mask": tf.TensorSpec((None, None), tf.int32, name="attention_mask"),
}
]
)
def serving(self, inputs):
output = self.call(inputs)
return self.serving_output(output)
ROBERTA_START_DOCSTRING = r"""
This model inherits from [`TFPreTrainedModel`]. 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 [tf.keras.Model](https://www.tensorflow.org/api_docs/python/tf/keras/Model) subclass. Use
it as a regular TF 2.0 Keras Model and refer to the TF 2.0 documentation for all matter related to general usage
and behavior.
TF 2.0 models accepts two formats as inputs:
- having all inputs as keyword arguments (like PyTorch models), or
- having all inputs as a list, tuple or dict in the first positional arguments.
This second option is useful when using [`tf.keras.Model.fit`] method which currently requires having all
the tensors in the first argument of the model call function: `model(inputs)`.
If you choose this second option, there are three possibilities you can use to gather all the input Tensors in
the first positional argument :
- a single Tensor with `input_ids` only and nothing else: `model(inputs_ids)`
- a list of varying length with one or several input Tensors IN THE ORDER given in the docstring:
`model([input_ids, attention_mask])` or `model([input_ids, attention_mask, token_type_ids])`
- a dictionary with one or several input Tensors associated to the input names given in the docstring:
`model({"input_ids": input_ids, "token_type_ids": token_type_ids})`
Parameters:
config ([`RobertaConfig`]): 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.
"""
ROBERTA_INPUTS_DOCSTRING = r"""
Args:
input_ids (`Numpy array` or `tf.Tensor` of shape `({0})`):
Indices of input sequence tokens in the vocabulary.
Indices can be obtained using [`RobertaTokenizer`]. See
[`PreTrainedTokenizer.__call__`] and [`PreTrainedTokenizer.encode`] for
details.
[What are input IDs?](../glossary#input-ids)
attention_mask (`Numpy array` or `tf.Tensor` of shape `({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#attention-mask)
token_type_ids (`Numpy array` or `tf.Tensor` of shape `({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#token-type-ids)
position_ids (`Numpy array` or `tf.Tensor` of shape `({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#position-ids)
head_mask (`Numpy array` or `tf.Tensor` 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 (`tf.Tensor` of shape `({0}, 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. This argument can be used only in eager mode, in graph mode the value in the
config will be used instead.
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. This argument can be used only in eager mode, in graph mode the value in the config will be
used instead.
return_dict (`bool`, *optional*):
Whether or not to return a [`~file_utils.ModelOutput`] instead of a plain tuple. This
argument can be used in eager mode, in graph mode the value will always be set to True.
training (`bool`, *optional*, defaults to `False`):
Whether or not to use the model in training mode (some modules like dropout modules have different
behaviors between training and evaluation).
"""
@add_start_docstrings(
"The bare RoBERTa Model transformer outputting raw hidden-states without any specific head on top.",
ROBERTA_START_DOCSTRING,
)
class TFRobertaModel(TFRobertaPreTrainedModel):
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.roberta = TFRobertaMainLayer(config, name="roberta")
@add_start_docstrings_to_model_forward(ROBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
processor_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFBaseModelOutputWithPoolingAndCrossAttentions,
config_class=_CONFIG_FOR_DOC,
)
def call(
self,
input_ids=None,
attention_mask=None,
token_type_ids=None,
position_ids=None,
head_mask=None,
inputs_embeds=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
past_key_values=None,
use_cache=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
training=False,
**kwargs,
):
r"""
encoder_hidden_states (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
the model is configured as a decoder.
encoder_attention_mask (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
past_key_values (`Tuple[Tuple[tf.Tensor]]` of length `config.n_layers`)
contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids`
(those that don't have their past key value states given to this model) of shape `(batch_size, 1)`
instead of all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
use_cache (`bool`, *optional*, defaults to `True`):
If set to `True`, `past_key_values` key value states are returned and can be used to speed up
decoding (see `past_key_values`). Set to `False` during training, `True` during generation
"""
inputs = input_processing(
func=self.call,
config=self.config,
input_ids=input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
position_ids=position_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
use_cache=use_cache,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
training=training,
kwargs_call=kwargs,
)
outputs = self.roberta(
input_ids=inputs["input_ids"],
attention_mask=inputs["attention_mask"],
token_type_ids=inputs["token_type_ids"],
position_ids=inputs["position_ids"],
head_mask=inputs["head_mask"],
inputs_embeds=inputs["inputs_embeds"],
encoder_hidden_states=inputs["encoder_hidden_states"],
encoder_attention_mask=inputs["encoder_attention_mask"],
past_key_values=inputs["past_key_values"],
use_cache=inputs["use_cache"],
output_attentions=inputs["output_attentions"],
output_hidden_states=inputs["output_hidden_states"],
return_dict=inputs["return_dict"],
training=inputs["training"],
)
return outputs
# Copied from transformers.models.bert.modeling_tf_bert.TFBertModel.serving_output
def serving_output(
self, output: TFBaseModelOutputWithPoolingAndCrossAttentions
) -> TFBaseModelOutputWithPoolingAndCrossAttentions:
output_cache = self.config.use_cache and self.config.is_decoder
pkv = tf.convert_to_tensor(output.past_key_values) if output_cache else None
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
cross_attns = tf.convert_to_tensor(output.cross_attentions) if output.cross_attentions is not None else None
if not (self.config.output_attentions and self.config.add_cross_attention):
cross_attns = None
return TFBaseModelOutputWithPoolingAndCrossAttentions(
last_hidden_state=output.last_hidden_state,
pooler_output=output.pooler_output,
past_key_values=pkv,
hidden_states=hs,
attentions=attns,
cross_attentions=cross_attns,
)
class TFRobertaLMHead(tf.keras.layers.Layer):
"""Roberta Head for masked language modeling."""
def __init__(self, config, input_embeddings, **kwargs):
super().__init__(**kwargs)
self.vocab_size = config.vocab_size
self.hidden_size = config.hidden_size
self.dense = tf.keras.layers.Dense(
config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
)
self.layer_norm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="layer_norm")
self.act = get_tf_activation("gelu")
# The output weights are the same as the input embeddings, but there is
# an output-only bias for each token.
self.decoder = input_embeddings
def build(self, input_shape):
self.bias = self.add_weight(shape=(self.vocab_size,), initializer="zeros", trainable=True, name="bias")
super().build(input_shape)
def get_output_embeddings(self):
return self.decoder
def set_output_embeddings(self, value):
self.decoder.weight = value
self.decoder.vocab_size = shape_list(value)[0]
def get_bias(self):
return {"bias": self.bias}
def set_bias(self, value):
self.bias = value["bias"]
self.vocab_size = shape_list(value["bias"])[0]
def call(self, hidden_states):
hidden_states = self.dense(hidden_states)
hidden_states = self.act(hidden_states)
hidden_states = self.layer_norm(hidden_states)
# project back to size of vocabulary with bias
seq_length = shape_list(tensor=hidden_states)[1]
hidden_states = tf.reshape(tensor=hidden_states, shape=[-1, self.hidden_size])
hidden_states = tf.matmul(a=hidden_states, b=self.decoder.weight, transpose_b=True)
hidden_states = tf.reshape(tensor=hidden_states, shape=[-1, seq_length, self.vocab_size])
hidden_states = tf.nn.bias_add(value=hidden_states, bias=self.bias)
return hidden_states
@add_start_docstrings("""RoBERTa Model with a `language modeling` head on top. """, ROBERTA_START_DOCSTRING)
class TFRobertaForMaskedLM(TFRobertaPreTrainedModel, TFMaskedLanguageModelingLoss):
# names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model
_keys_to_ignore_on_load_unexpected = [r"pooler", r"lm_head.decoder.weight"]
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.roberta = TFRobertaMainLayer(config, add_pooling_layer=False, name="roberta")
self.lm_head = TFRobertaLMHead(config, self.roberta.embeddings, name="lm_head")
def get_lm_head(self):
return self.lm_head
def get_prefix_bias_name(self):
warnings.warn("The method get_prefix_bias_name is deprecated. Please use `get_bias` instead.", FutureWarning)
return self.name + "/" + self.lm_head.name
@add_start_docstrings_to_model_forward(ROBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
processor_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFMaskedLMOutput,
config_class=_CONFIG_FOR_DOC,
)
def call(
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,
labels=None,
training=False,
**kwargs,
):
r"""
labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ..., config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`
"""
inputs = input_processing(
func=self.call,
config=self.config,
input_ids=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,
labels=labels,
training=training,
kwargs_call=kwargs,
)
outputs = self.roberta(
inputs["input_ids"],
attention_mask=inputs["attention_mask"],
token_type_ids=inputs["token_type_ids"],
position_ids=inputs["position_ids"],
head_mask=inputs["head_mask"],
inputs_embeds=inputs["inputs_embeds"],
output_attentions=inputs["output_attentions"],
output_hidden_states=inputs["output_hidden_states"],
return_dict=inputs["return_dict"],
training=inputs["training"],
)
sequence_output = outputs[0]
prediction_scores = self.lm_head(sequence_output)
loss = None if inputs["labels"] is None else self.compute_loss(inputs["labels"], prediction_scores)
if not inputs["return_dict"]:
output = (prediction_scores,) + outputs[2:]
return ((loss,) + output) if loss is not None else output
return TFMaskedLMOutput(
loss=loss,
logits=prediction_scores,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForMaskedLM.serving_output
def serving_output(self, output: TFMaskedLMOutput) -> TFMaskedLMOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFMaskedLMOutput(logits=output.logits, hidden_states=hs, attentions=attns)
class TFRobertaForCausalLM(TFRobertaPreTrainedModel, TFCausalLanguageModelingLoss):
# names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model
_keys_to_ignore_on_load_unexpected = [r"pooler", r"lm_head.decoder.weight"]
def __init__(self, config: RobertaConfig, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
if not config.is_decoder:
logger.warning("If you want to use `TFRobertaLMHeadModel` as a standalone, add `is_decoder=True.`")
self.roberta = TFRobertaMainLayer(config, add_pooling_layer=False, name="roberta")
self.lm_head = TFRobertaLMHead(config, input_embeddings=self.roberta.embeddings, name="lm_head")
def get_lm_head(self):
return self.lm_head
def get_prefix_bias_name(self):
warnings.warn("The method get_prefix_bias_name is deprecated. Please use `get_bias` instead.", FutureWarning)
return self.name + "/" + self.lm_head.name
# Copied from transformers.models.bert.modeling_tf_bert.TFBertLMHeadModel.prepare_inputs_for_generation
def prepare_inputs_for_generation(self, inputs, past=None, attention_mask=None, **model_kwargs):
# cut decoder_input_ids if past is used
if past:
inputs = tf.expand_dims(inputs[:, -1], -1)
return {
"input_ids": inputs,
"attention_mask": attention_mask,
"past_key_values": past,
"use_cache": model_kwargs["use_cache"],
}
@add_start_docstrings_to_model_forward(ROBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
processor_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFCausalLMOutputWithCrossAttentions,
config_class=_CONFIG_FOR_DOC,
)
def call(
self,
input_ids=None,
attention_mask=None,
token_type_ids=None,
position_ids=None,
head_mask=None,
inputs_embeds=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
past_key_values=None,
use_cache=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
labels=None,
training=False,
**kwargs,
) -> Union[TFCausalLMOutputWithCrossAttentions, Tuple[tf.Tensor]]:
r"""
encoder_hidden_states (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
the model is configured as a decoder.
encoder_attention_mask (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
past_key_values (`Tuple[Tuple[tf.Tensor]]` of length `config.n_layers`)
contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids`
(those that don't have their past key value states given to this model) of shape `(batch_size, 1)`
instead of all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
use_cache (`bool`, *optional*, defaults to `True`):
If set to `True`, `past_key_values` key value states are returned and can be used to speed up
decoding (see `past_key_values`). Set to `False` during training, `True` during generation
labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
Labels for computing the cross entropy classification loss. Indices should be in `[0, ..., config.vocab_size - 1]`.
"""
inputs = input_processing(
func=self.call,
config=self.config,
input_ids=input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
position_ids=position_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
use_cache=use_cache,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
labels=labels,
training=training,
kwargs_call=kwargs,
)
outputs = self.roberta(
input_ids=inputs["input_ids"],
attention_mask=inputs["attention_mask"],
token_type_ids=inputs["token_type_ids"],
position_ids=inputs["position_ids"],
head_mask=inputs["head_mask"],
inputs_embeds=inputs["inputs_embeds"],
encoder_hidden_states=inputs["encoder_hidden_states"],
encoder_attention_mask=inputs["encoder_attention_mask"],
past_key_values=inputs["past_key_values"],
use_cache=inputs["use_cache"],
output_attentions=inputs["output_attentions"],
output_hidden_states=inputs["output_hidden_states"],
return_dict=inputs["return_dict"],
training=inputs["training"],
)
sequence_output = outputs[0]
logits = self.lm_head(hidden_states=sequence_output)
loss = None
if inputs["labels"] is not None:
# shift labels to the left and cut last logit token
logits = logits[:, :-1]
labels = inputs["labels"][:, 1:]
loss = self.compute_loss(labels=labels, logits=logits)
if not inputs["return_dict"]:
output = (logits,) + outputs[2:]
return ((loss,) + output) if loss is not None else output
return TFCausalLMOutputWithCrossAttentions(
loss=loss,
logits=logits,
past_key_values=outputs.past_key_values,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
cross_attentions=outputs.cross_attentions,
)
# Copied from transformers.models.bert.modeling_tf_bert.TFBertLMHeadModel.serving_output
def serving_output(self, output: TFCausalLMOutputWithCrossAttentions) -> TFCausalLMOutputWithCrossAttentions:
output_cache = self.config.use_cache and self.config.is_decoder
pkv = tf.convert_to_tensor(output.past_key_values) if output_cache else None
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
cross_attns = tf.convert_to_tensor(output.cross_attentions) if output.cross_attentions is not None else None
if not (self.config.output_attentions and self.config.add_cross_attention):
cross_attns = None
return TFCausalLMOutputWithCrossAttentions(
logits=output.logits, past_key_values=pkv, hidden_states=hs, attentions=attns, cross_attentions=cross_attns
)
class TFRobertaClassificationHead(tf.keras.layers.Layer):
"""Head for sentence-level classification tasks."""
def __init__(self, config, **kwargs):
super().__init__(**kwargs)
self.dense = tf.keras.layers.Dense(
config.hidden_size,
kernel_initializer=get_initializer(config.initializer_range),
activation="tanh",
name="dense",
)
classifier_dropout = (
config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
)
self.dropout = tf.keras.layers.Dropout(classifier_dropout)
self.out_proj = tf.keras.layers.Dense(
config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="out_proj"
)
def call(self, features, training=False):
x = features[:, 0, :] # take token (equiv. to [CLS])
x = self.dropout(x, training=training)
x = self.dense(x)
x = self.dropout(x, training=training)
x = self.out_proj(x)
return x
@add_start_docstrings(
"""
RoBERTa Model transformer with a sequence classification/regression head on top (a linear layer on top of the
pooled output) e.g. for GLUE tasks.
""",
ROBERTA_START_DOCSTRING,
)
class TFRobertaForSequenceClassification(TFRobertaPreTrainedModel, TFSequenceClassificationLoss):
# names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model
_keys_to_ignore_on_load_unexpected = [r"pooler", r"lm_head"]
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.num_labels = config.num_labels
self.roberta = TFRobertaMainLayer(config, add_pooling_layer=False, name="roberta")
self.classifier = TFRobertaClassificationHead(config, name="classifier")
@add_start_docstrings_to_model_forward(ROBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
processor_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFSequenceClassifierOutput,
config_class=_CONFIG_FOR_DOC,
)
def call(
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,
labels=None,
training=False,
**kwargs,
):
r"""
labels (`tf.Tensor` of shape `(batch_size,)`, *optional*):
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).
"""
inputs = input_processing(
func=self.call,
config=self.config,
input_ids=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,
labels=labels,
training=training,
kwargs_call=kwargs,
)
outputs = self.roberta(
inputs["input_ids"],
attention_mask=inputs["attention_mask"],
token_type_ids=inputs["token_type_ids"],
position_ids=inputs["position_ids"],
head_mask=inputs["head_mask"],
inputs_embeds=inputs["inputs_embeds"],
output_attentions=inputs["output_attentions"],
output_hidden_states=inputs["output_hidden_states"],
return_dict=inputs["return_dict"],
training=inputs["training"],
)
sequence_output = outputs[0]
logits = self.classifier(sequence_output, training=inputs["training"])
loss = None if inputs["labels"] is None else self.compute_loss(inputs["labels"], logits)
if not inputs["return_dict"]:
output = (logits,) + outputs[2:]
return ((loss,) + output) if loss is not None else output
return TFSequenceClassifierOutput(
loss=loss,
logits=logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForSequenceClassification.serving_output
def serving_output(self, output: TFSequenceClassifierOutput) -> TFSequenceClassifierOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFSequenceClassifierOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@add_start_docstrings(
"""
Roberta Model with a multiple choice classification head on top (a linear layer on top of the pooled output and a
softmax) e.g. for RocStories/SWAG tasks.
""",
ROBERTA_START_DOCSTRING,
)
class TFRobertaForMultipleChoice(TFRobertaPreTrainedModel, TFMultipleChoiceLoss):
# names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model
_keys_to_ignore_on_load_unexpected = [r"lm_head"]
_keys_to_ignore_on_load_missing = [r"dropout"]
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.roberta = TFRobertaMainLayer(config, name="roberta")
self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
self.classifier = tf.keras.layers.Dense(
1, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
)
@property
def dummy_inputs(self):
"""
Dummy inputs to build the network.
Returns:
tf.Tensor with dummy inputs
"""
return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS)}
@add_start_docstrings_to_model_forward(ROBERTA_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
@add_code_sample_docstrings(
processor_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFMultipleChoiceModelOutput,
config_class=_CONFIG_FOR_DOC,
)
def call(
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,
labels=None,
training=False,
**kwargs,
):
r"""
labels (`tf.Tensor` of shape `(batch_size,)`, *optional*):
Labels for computing the multiple choice classification loss. Indices should be in `[0, ..., num_choices]` where `num_choices` is the size of the second dimension of the input tensors. (See
`input_ids` above)
"""
inputs = input_processing(
func=self.call,
config=self.config,
input_ids=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,
labels=labels,
training=training,
kwargs_call=kwargs,
)
if inputs["input_ids"] is not None:
num_choices = shape_list(inputs["input_ids"])[1]
seq_length = shape_list(inputs["input_ids"])[2]
else:
num_choices = shape_list(inputs_embeds)[1]
seq_length = shape_list(inputs_embeds)[2]
flat_input_ids = tf.reshape(inputs["input_ids"], (-1, seq_length)) if inputs["input_ids"] is not None else None
flat_attention_mask = (
tf.reshape(inputs["attention_mask"], (-1, seq_length)) if inputs["attention_mask"] is not None else None
)
flat_token_type_ids = (
tf.reshape(inputs["token_type_ids"], (-1, seq_length)) if inputs["token_type_ids"] is not None else None
)
flat_position_ids = (
tf.reshape(inputs["position_ids"], (-1, seq_length)) if inputs["position_ids"] is not None else None
)
outputs = self.roberta(
flat_input_ids,
flat_attention_mask,
flat_token_type_ids,
flat_position_ids,
inputs["head_mask"],
inputs["inputs_embeds"],
inputs["output_attentions"],
inputs["output_hidden_states"],
return_dict=inputs["return_dict"],
training=inputs["training"],
)
pooled_output = outputs[1]
pooled_output = self.dropout(pooled_output, training=inputs["training"])
logits = self.classifier(pooled_output)
reshaped_logits = tf.reshape(logits, (-1, num_choices))
loss = None if inputs["labels"] is None else self.compute_loss(inputs["labels"], reshaped_logits)
if not inputs["return_dict"]:
output = (reshaped_logits,) + outputs[2:]
return ((loss,) + output) if loss is not None else output
return TFMultipleChoiceModelOutput(
loss=loss,
logits=reshaped_logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
@tf.function(
input_signature=[
{
"input_ids": tf.TensorSpec((None, None, None), tf.int32, name="input_ids"),
"attention_mask": tf.TensorSpec((None, None, None), tf.int32, name="attention_mask"),
}
]
)
def serving(self, inputs):
output = self.call(inputs)
return self.serving_output(output)
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForMultipleChoice.serving_output
def serving_output(self, output: TFMultipleChoiceModelOutput) -> TFMultipleChoiceModelOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFMultipleChoiceModelOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@add_start_docstrings(
"""
RoBERTa 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.
""",
ROBERTA_START_DOCSTRING,
)
class TFRobertaForTokenClassification(TFRobertaPreTrainedModel, TFTokenClassificationLoss):
# names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model
_keys_to_ignore_on_load_unexpected = [r"pooler", r"lm_head"]
_keys_to_ignore_on_load_missing = [r"dropout"]
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.num_labels = config.num_labels
self.roberta = TFRobertaMainLayer(config, add_pooling_layer=False, name="roberta")
classifier_dropout = (
config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
)
self.dropout = tf.keras.layers.Dropout(classifier_dropout)
self.classifier = tf.keras.layers.Dense(
config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
)
@add_start_docstrings_to_model_forward(ROBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
processor_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFTokenClassifierOutput,
config_class=_CONFIG_FOR_DOC,
)
def call(
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,
labels=None,
training=False,
**kwargs,
):
r"""
labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
"""
inputs = input_processing(
func=self.call,
config=self.config,
input_ids=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,
labels=labels,
training=training,
kwargs_call=kwargs,
)
outputs = self.roberta(
inputs["input_ids"],
attention_mask=inputs["attention_mask"],
token_type_ids=inputs["token_type_ids"],
position_ids=inputs["position_ids"],
head_mask=inputs["head_mask"],
inputs_embeds=inputs["inputs_embeds"],
output_attentions=inputs["output_attentions"],
output_hidden_states=inputs["output_hidden_states"],
return_dict=inputs["return_dict"],
training=inputs["training"],
)
sequence_output = outputs[0]
sequence_output = self.dropout(sequence_output, training=inputs["training"])
logits = self.classifier(sequence_output)
loss = None if inputs["labels"] is None else self.compute_loss(inputs["labels"], logits)
if not inputs["return_dict"]:
output = (logits,) + outputs[2:]
return ((loss,) + output) if loss is not None else output
return TFTokenClassifierOutput(
loss=loss,
logits=logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForTokenClassification.serving_output
def serving_output(self, output: TFTokenClassifierOutput) -> TFTokenClassifierOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFTokenClassifierOutput(logits=output.logits, hidden_states=hs, attentions=attns)
@add_start_docstrings(
"""
RoBERTa 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`).
""",
ROBERTA_START_DOCSTRING,
)
class TFRobertaForQuestionAnswering(TFRobertaPreTrainedModel, TFQuestionAnsweringLoss):
# names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model
_keys_to_ignore_on_load_unexpected = [r"pooler", r"lm_head"]
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.num_labels = config.num_labels
self.roberta = TFRobertaMainLayer(config, add_pooling_layer=False, name="roberta")
self.qa_outputs = tf.keras.layers.Dense(
config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="qa_outputs"
)
@add_start_docstrings_to_model_forward(ROBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
processor_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFQuestionAnsweringModelOutput,
config_class=_CONFIG_FOR_DOC,
)
def call(
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,
start_positions=None,
end_positions=None,
training=False,
**kwargs,
):
r"""
start_positions (`tf.Tensor` of shape `(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 (`tf.Tensor` of shape `(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.
"""
inputs = input_processing(
func=self.call,
config=self.config,
input_ids=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,
start_positions=start_positions,
end_positions=end_positions,
training=training,
kwargs_call=kwargs,
)
outputs = self.roberta(
inputs["input_ids"],
attention_mask=inputs["attention_mask"],
token_type_ids=inputs["token_type_ids"],
position_ids=inputs["position_ids"],
head_mask=inputs["head_mask"],
inputs_embeds=inputs["inputs_embeds"],
output_attentions=inputs["output_attentions"],
output_hidden_states=inputs["output_hidden_states"],
return_dict=inputs["return_dict"],
training=inputs["training"],
)
sequence_output = outputs[0]
logits = self.qa_outputs(sequence_output)
start_logits, end_logits = tf.split(logits, 2, axis=-1)
start_logits = tf.squeeze(start_logits, axis=-1)
end_logits = tf.squeeze(end_logits, axis=-1)
loss = None
if inputs["start_positions"] is not None and inputs["end_positions"] is not None:
labels = {"start_position": inputs["start_positions"]}
labels["end_position"] = inputs["end_positions"]
loss = self.compute_loss(labels, (start_logits, end_logits))
if not inputs["return_dict"]:
output = (start_logits, end_logits) + outputs[2:]
return ((loss,) + output) if loss is not None else output
return TFQuestionAnsweringModelOutput(
loss=loss,
start_logits=start_logits,
end_logits=end_logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForQuestionAnswering.serving_output
def serving_output(self, output: TFQuestionAnsweringModelOutput) -> TFQuestionAnsweringModelOutput:
hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
return TFQuestionAnsweringModelOutput(
start_logits=output.start_logits, end_logits=output.end_logits, hidden_states=hs, attentions=attns
)