# coding=utf-8
# Copyright 2019 The Google AI Language Team Authors and The HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" TF Electra model. """
import math
import warnings
from dataclasses import dataclass
from typing import Dict, Optional, Tuple, Union
import tensorflow as tf
from ...activations_tf import get_tf_activation
from ...file_utils import (
DUMMY_INPUTS,
MULTIPLE_CHOICE_DUMMY_INPUTS,
ModelOutput,
add_code_sample_docstrings,
add_start_docstrings,
add_start_docstrings_to_model_forward,
replace_return_docstrings,
)
from ...modeling_tf_outputs import (
TFBaseModelOutputWithPastAndCrossAttentions,
TFMaskedLMOutput,
TFMultipleChoiceModelOutput,
TFQuestionAnsweringModelOutput,
TFSequenceClassifierOutput,
TFTokenClassifierOutput,
)
from ...modeling_tf_utils import (
TFMaskedLanguageModelingLoss,
TFMultipleChoiceLoss,
TFPreTrainedModel,
TFQuestionAnsweringLoss,
TFSequenceClassificationLoss,
TFSequenceSummary,
TFTokenClassificationLoss,
get_initializer,
input_processing,
keras_serializable,
shape_list,
)
from ...utils import logging
from .configuration_electra import ElectraConfig
logger = logging.get_logger(__name__)
_CHECKPOINT_FOR_DOC = "google/electra-small-discriminator"
_CONFIG_FOR_DOC = "ElectraConfig"
_TOKENIZER_FOR_DOC = "ElectraTokenizer"
TF_ELECTRA_PRETRAINED_MODEL_ARCHIVE_LIST = [
"google/electra-small-generator",
"google/electra-base-generator",
"google/electra-large-generator",
"google/electra-small-discriminator",
"google/electra-base-discriminator",
"google/electra-large-discriminator",
# See all ELECTRA models at https://huggingface.co/models?filter=electra
]
# Copied from transformers.models.bert.modeling_tf_bert.TFBertSelfAttention with Bert->Electra
class TFElectraSelfAttention(tf.keras.layers.Layer):
def __init__(self, config: ElectraConfig, **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 TFElectraModel 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->Electra
class TFElectraSelfOutput(tf.keras.layers.Layer):
def __init__(self, config: ElectraConfig, **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 from transformers.models.bert.modeling_tf_bert.TFBertAttention with Bert->Electra
class TFElectraAttention(tf.keras.layers.Layer):
def __init__(self, config: ElectraConfig, **kwargs):
super().__init__(**kwargs)
self.self_attention = TFElectraSelfAttention(config, name="self")
self.dense_output = TFElectraSelfOutput(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->Electra
class TFElectraIntermediate(tf.keras.layers.Layer):
def __init__(self, config: ElectraConfig, **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->Electra
class TFElectraOutput(tf.keras.layers.Layer):
def __init__(self, config: ElectraConfig, **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->Electra
class TFElectraLayer(tf.keras.layers.Layer):
def __init__(self, config: ElectraConfig, **kwargs):
super().__init__(**kwargs)
self.attention = TFElectraAttention(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 = TFElectraAttention(config, name="crossattention")
self.intermediate = TFElectraIntermediate(config, name="intermediate")
self.bert_output = TFElectraOutput(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->Electra
class TFElectraEncoder(tf.keras.layers.Layer):
def __init__(self, config: ElectraConfig, **kwargs):
super().__init__(**kwargs)
self.config = config
self.layer = [TFElectraLayer(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,
)
# Copied from transformers.models.bert.modeling_tf_bert.TFBertPooler with Bert->Electra
class TFElectraPooler(tf.keras.layers.Layer):
def __init__(self, config: ElectraConfig, **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.albert.modeling_tf_albert.TFAlbertEmbeddings with Albert->Electra
class TFElectraEmbeddings(tf.keras.layers.Layer):
"""Construct the embeddings from word, position and token_type embeddings."""
def __init__(self, config: ElectraConfig, **kwargs):
super().__init__(**kwargs)
self.vocab_size = config.vocab_size
self.type_vocab_size = config.type_vocab_size
self.embedding_size = config.embedding_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.embedding_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.embedding_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.embedding_size],
initializer=get_initializer(self.initializer_range),
)
super().build(input_shape)
# Copied from transformers.models.bert.modeling_tf_bert.TFBertEmbeddings.call
def call(
self,
input_ids: tf.Tensor = None,
position_ids: tf.Tensor = None,
token_type_ids: tf.Tensor = None,
inputs_embeds: tf.Tensor = None,
past_key_values_length=0,
training: bool = False,
) -> tf.Tensor:
"""
Applies embedding based on inputs tensor.
Returns:
final_embeddings (`tf.Tensor`): output embedding tensor.
"""
if input_ids is None and inputs_embeds is None:
raise ValueError("Need to provide either `input_ids` or `input_embeds`.")
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:
position_ids = tf.expand_dims(
tf.range(start=past_key_values_length, limit=input_shape[1] + past_key_values_length), 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
class TFElectraDiscriminatorPredictions(tf.keras.layers.Layer):
def __init__(self, config, **kwargs):
super().__init__(**kwargs)
self.dense = tf.keras.layers.Dense(config.hidden_size, name="dense")
self.dense_prediction = tf.keras.layers.Dense(1, name="dense_prediction")
self.config = config
def call(self, discriminator_hidden_states, training=False):
hidden_states = self.dense(discriminator_hidden_states)
hidden_states = get_tf_activation(self.config.hidden_act)(hidden_states)
logits = tf.squeeze(self.dense_prediction(hidden_states), -1)
return logits
class TFElectraGeneratorPredictions(tf.keras.layers.Layer):
def __init__(self, config, **kwargs):
super().__init__(**kwargs)
self.LayerNorm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
self.dense = tf.keras.layers.Dense(config.embedding_size, name="dense")
def call(self, generator_hidden_states, training=False):
hidden_states = self.dense(generator_hidden_states)
hidden_states = get_tf_activation("gelu")(hidden_states)
hidden_states = self.LayerNorm(hidden_states)
return hidden_states
class TFElectraPreTrainedModel(TFPreTrainedModel):
"""
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
models.
"""
config_class = ElectraConfig
base_model_prefix = "electra"
# When the model is loaded from a PT model
_keys_to_ignore_on_load_unexpected = [r"generator_lm_head.weight"]
_keys_to_ignore_on_load_missing = [r"dropout"]
@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
@keras_serializable
class TFElectraMainLayer(tf.keras.layers.Layer):
config_class = ElectraConfig
def __init__(self, config, **kwargs):
super().__init__(**kwargs)
self.config = config
self.is_decoder = config.is_decoder
self.embeddings = TFElectraEmbeddings(config, name="embeddings")
if config.embedding_size != config.hidden_size:
self.embeddings_project = tf.keras.layers.Dense(config.hidden_size, name="embeddings_project")
self.encoder = TFElectraEncoder(config, name="encoder")
def get_input_embeddings(self):
return self.embeddings
def set_input_embeddings(self, value):
self.embeddings.weight = value
self.embeddings.vocab_size = shape_list(value)[0]
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
def get_extended_attention_mask(self, attention_mask, input_shape, dtype, past_key_values_length=0):
batch_size, seq_length = input_shape
if attention_mask is None:
attention_mask = tf.fill(dims=(batch_size, seq_length + past_key_values_length), value=1)
# 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(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=attention_mask.dtype)
extended_attention_mask = causal_mask * 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(
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=dtype)
one_cst = tf.constant(1.0, dtype=dtype)
ten_thousand_cst = tf.constant(-10000.0, dtype=dtype)
extended_attention_mask = tf.multiply(tf.subtract(one_cst, extended_attention_mask), ten_thousand_cst)
return extended_attention_mask
def get_head_mask(self, head_mask):
if head_mask is not None:
raise NotImplementedError
else:
head_mask = [None] * self.config.num_hidden_layers
return head_mask
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,
):
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)
hidden_states = 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"],
)
extended_attention_mask = self.get_extended_attention_mask(
inputs["attention_mask"], input_shape, hidden_states.dtype, past_key_values_length
)
# 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
inputs["head_mask"] = self.get_head_mask(inputs["head_mask"])
if hasattr(self, "embeddings_project"):
hidden_states = self.embeddings_project(hidden_states, training=inputs["training"])
hidden_states = self.encoder(
hidden_states=hidden_states,
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"],
)
return hidden_states
@dataclass
class TFElectraForPreTrainingOutput(ModelOutput):
"""
Output type of [`TFElectraForPreTraining`].
Args:
loss (*optional*, returned when `labels` is provided, `tf.Tensor` of shape `(1,)`):
Total loss of the ELECTRA objective.
logits (`tf.Tensor` of shape `(batch_size, sequence_length)`):
Prediction scores of the head (scores for each token before SoftMax).
hidden_states (`tuple(tf.Tensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
Tuple of `tf.Tensor` (one for the output of the embeddings + one for the output of each layer) of
shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the initial embedding outputs.
attentions (`tuple(tf.Tensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
Tuple of `tf.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
"""
logits: tf.Tensor = None
hidden_states: Optional[Tuple[tf.Tensor]] = None
attentions: Optional[Tuple[tf.Tensor]] = None
ELECTRA_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 ([`ElectraConfig`]): 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.
"""
ELECTRA_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 [`ElectraTokenizer`]. 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)
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 Electra Model transformer outputting raw hidden-states without any specific head on top. Identical to "
"the BERT model except that it uses an additional linear layer between the embedding layer and the encoder if the "
"hidden size and embedding size are different. "
""
"Both the generator and discriminator checkpoints may be loaded into this model.",
ELECTRA_START_DOCSTRING,
)
class TFElectraModel(TFElectraPreTrainedModel):
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.electra = TFElectraMainLayer(config, name="electra")
@add_start_docstrings_to_model_forward(ELECTRA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
processor_class=_TOKENIZER_FOR_DOC,
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFBaseModelOutputWithPastAndCrossAttentions,
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.electra(
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"],
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"],
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"],
)
return outputs
def serving_output(self, output):
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 TFBaseModelOutputWithPastAndCrossAttentions(
last_hidden_state=output.last_hidden_state,
past_key_values=pkv,
hidden_states=hs,
attentions=attns,
cross_attentions=cross_attns,
)
@add_start_docstrings(
"""
Electra model with a binary classification head on top as used during pretraining for identifying generated tokens.
Even though both the discriminator and generator may be loaded into this model, the discriminator is the only model
of the two to have the correct classification head to be used for this model.
""",
ELECTRA_START_DOCSTRING,
)
class TFElectraForPreTraining(TFElectraPreTrainedModel):
def __init__(self, config, **kwargs):
super().__init__(config, **kwargs)
self.electra = TFElectraMainLayer(config, name="electra")
self.discriminator_predictions = TFElectraDiscriminatorPredictions(config, name="discriminator_predictions")
@add_start_docstrings_to_model_forward(ELECTRA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@replace_return_docstrings(output_type=TFElectraForPreTrainingOutput, 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,
training=False,
**kwargs,
):
r"""
Returns:
Examples:
```python
>>> import tensorflow as tf
>>> from transformers import ElectraTokenizer, TFElectraForPreTraining
>>> tokenizer = ElectraTokenizer.from_pretrained('google/electra-small-discriminator')
>>> model = TFElectraForPreTraining.from_pretrained('google/electra-small-discriminator')
>>> input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
>>> outputs = model(input_ids)
>>> scores = outputs[0]
```"""
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,
training=training,
kwargs_call=kwargs,
)
discriminator_hidden_states = self.electra(
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"],
output_attentions=inputs["output_attentions"],
output_hidden_states=inputs["output_hidden_states"],
return_dict=inputs["return_dict"],
training=inputs["training"],
)
discriminator_sequence_output = discriminator_hidden_states[0]
logits = self.discriminator_predictions(discriminator_sequence_output)
if not inputs["return_dict"]:
return (logits,) + discriminator_hidden_states[1:]
return TFElectraForPreTrainingOutput(
logits=logits,
hidden_states=discriminator_hidden_states.hidden_states,
attentions=discriminator_hidden_states.attentions,
)
def serving_output(self, output):
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 TFElectraForPreTrainingOutput(logits=output.logits, hidden_states=hs, attentions=attns)
class TFElectraMaskedLMHead(tf.keras.layers.Layer):
def __init__(self, config, input_embeddings, **kwargs):
super().__init__(**kwargs)
self.vocab_size = config.vocab_size
self.embedding_size = config.embedding_size
self.input_embeddings = 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.input_embeddings
def set_output_embeddings(self, value):
self.input_embeddings.weight = value
self.input_embeddings.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):
seq_length = shape_list(tensor=hidden_states)[1]
hidden_states = tf.reshape(tensor=hidden_states, shape=[-1, self.embedding_size])
hidden_states = tf.matmul(a=hidden_states, b=self.input_embeddings.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(
"""
Electra model with a language modeling head on top.
Even though both the discriminator and generator may be loaded into this model, the generator is the only model of
the two to have been trained for the masked language modeling task.
""",
ELECTRA_START_DOCSTRING,
)
class TFElectraForMaskedLM(TFElectraPreTrainedModel, TFMaskedLanguageModelingLoss):
def __init__(self, config, **kwargs):
super().__init__(config, **kwargs)
self.vocab_size = config.vocab_size
self.electra = TFElectraMainLayer(config, name="electra")
self.generator_predictions = TFElectraGeneratorPredictions(config, name="generator_predictions")
if isinstance(config.hidden_act, str):
self.activation = get_tf_activation(config.hidden_act)
else:
self.activation = config.hidden_act
self.generator_lm_head = TFElectraMaskedLMHead(config, self.electra.embeddings, name="generator_lm_head")
def get_lm_head(self):
return self.generator_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.generator_lm_head.name
@add_start_docstrings_to_model_forward(ELECTRA_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,
)
generator_hidden_states = self.electra(
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"],
output_attentions=inputs["output_attentions"],
output_hidden_states=inputs["output_hidden_states"],
return_dict=inputs["return_dict"],
training=inputs["training"],
)
generator_sequence_output = generator_hidden_states[0]
prediction_scores = self.generator_predictions(generator_sequence_output, training=inputs["training"])
prediction_scores = self.generator_lm_head(prediction_scores, training=inputs["training"])
loss = None if inputs["labels"] is None else self.compute_loss(inputs["labels"], prediction_scores)
if not inputs["return_dict"]:
output = (prediction_scores,) + generator_hidden_states[1:]
return ((loss,) + output) if loss is not None else output
return TFMaskedLMOutput(
loss=loss,
logits=prediction_scores,
hidden_states=generator_hidden_states.hidden_states,
attentions=generator_hidden_states.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 TFElectraClassificationHead(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), name="dense"
)
classifier_dropout = (
config.classifhidden_dropout_probier_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, inputs, **kwargs):
x = inputs[:, 0, :] # take token (equiv. to [CLS])
x = self.dropout(x)
x = self.dense(x)
x = get_tf_activation("gelu")(x) # although BERT uses tanh here, it seems Electra authors used gelu here
x = self.dropout(x)
x = self.out_proj(x)
return x
@add_start_docstrings(
"""
ELECTRA Model transformer with a sequence classification/regression head on top (a linear layer on top of the
pooled output) e.g. for GLUE tasks.
""",
ELECTRA_START_DOCSTRING,
)
class TFElectraForSequenceClassification(TFElectraPreTrainedModel, TFSequenceClassificationLoss):
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.num_labels = config.num_labels
self.electra = TFElectraMainLayer(config, name="electra")
self.classifier = TFElectraClassificationHead(config, name="classifier")
@add_start_docstrings_to_model_forward(ELECTRA_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.electra(
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"],
output_attentions=inputs["output_attentions"],
output_hidden_states=inputs["output_hidden_states"],
return_dict=inputs["return_dict"],
training=inputs["training"],
)
logits = self.classifier(outputs[0])
loss = None if inputs["labels"] is None else self.compute_loss(inputs["labels"], logits)
if not inputs["return_dict"]:
output = (logits,) + outputs[1:]
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(
"""
ELECTRA 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.
""",
ELECTRA_START_DOCSTRING,
)
class TFElectraForMultipleChoice(TFElectraPreTrainedModel, TFMultipleChoiceLoss):
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.electra = TFElectraMainLayer(config, name="electra")
self.sequence_summary = TFSequenceSummary(
config, initializer_range=config.initializer_range, name="sequence_summary"
)
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(ELECTRA_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["inputs_embeds"])[1]
seq_length = shape_list(inputs["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
)
flat_inputs_embeds = (
tf.reshape(inputs["inputs_embeds"], (-1, seq_length, shape_list(inputs["inputs_embeds"])[3]))
if inputs["inputs_embeds"] is not None
else None
)
outputs = self.electra(
flat_input_ids,
flat_attention_mask,
flat_token_type_ids,
flat_position_ids,
inputs["head_mask"],
flat_inputs_embeds,
inputs["output_attentions"],
inputs["output_hidden_states"],
return_dict=inputs["return_dict"],
training=inputs["training"],
)
logits = self.sequence_summary(outputs[0])
logits = self.classifier(logits)
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[1:]
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"),
"token_type_ids": tf.TensorSpec((None, None, None), tf.int32, name="token_type_ids"),
}
]
)
# Copied from transformers.models.bert.modeling_tf_bert.TFBertForMultipleChoice.serving
def serving(self, inputs: Dict[str, tf.Tensor]):
output = self.call(input_ids=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(
"""
Electra model with a token classification head on top.
Both the discriminator and generator may be loaded into this model.
""",
ELECTRA_START_DOCSTRING,
)
class TFElectraForTokenClassification(TFElectraPreTrainedModel, TFTokenClassificationLoss):
def __init__(self, config, **kwargs):
super().__init__(config, **kwargs)
self.electra = TFElectraMainLayer(config, name="electra")
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(ELECTRA_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,
)
discriminator_hidden_states = self.electra(
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"],
output_attentions=inputs["output_attentions"],
output_hidden_states=inputs["output_hidden_states"],
return_dict=inputs["return_dict"],
training=inputs["training"],
)
discriminator_sequence_output = discriminator_hidden_states[0]
discriminator_sequence_output = self.dropout(discriminator_sequence_output)
logits = self.classifier(discriminator_sequence_output)
loss = None if inputs["labels"] is None else self.compute_loss(inputs["labels"], logits)
if not inputs["return_dict"]:
output = (logits,) + discriminator_hidden_states[1:]
return ((loss,) + output) if loss is not None else output
return TFTokenClassifierOutput(
loss=loss,
logits=logits,
hidden_states=discriminator_hidden_states.hidden_states,
attentions=discriminator_hidden_states.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(
"""
Electra 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`).
""",
ELECTRA_START_DOCSTRING,
)
class TFElectraForQuestionAnswering(TFElectraPreTrainedModel, TFQuestionAnsweringLoss):
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.num_labels = config.num_labels
self.electra = TFElectraMainLayer(config, name="electra")
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(ELECTRA_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,
)
discriminator_hidden_states = self.electra(
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"],
output_attentions=inputs["output_attentions"],
output_hidden_states=inputs["output_hidden_states"],
return_dict=inputs["return_dict"],
training=inputs["training"],
)
discriminator_sequence_output = discriminator_hidden_states[0]
logits = self.qa_outputs(discriminator_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,
) + discriminator_hidden_states[1:]
return ((loss,) + output) if loss is not None else output
return TFQuestionAnsweringModelOutput(
loss=loss,
start_logits=start_logits,
end_logits=end_logits,
hidden_states=discriminator_hidden_states.hidden_states,
attentions=discriminator_hidden_states.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
)