Restore TF embeddings and attention layers to their previous version (#9890)
* Refacto BERT * Restore all the concerned models * Remove print * Update template * Apply Sylvain's and Morgan's comments * Fix cast * Put the cast inside call * Remove cond in ebds * Fix funnel * Restore previous dot product (attention_scores) computation * Add ConvBERT and BART * Make all the S2S models ONNX compliant * Fix test * Fix check copies
This commit is contained in:
Julien Plu
@@ -15,8 +15,9 @@
|
||||
# limitations under the License.
|
||||
""" TF 2.0 RoBERTa model. """
|
||||
|
||||
import math
|
||||
import warnings
|
||||
from typing import Any, Dict, Optional, Tuple, Union
|
||||
from typing import Optional, Tuple, Union
|
||||
|
||||
import numpy as np
|
||||
import tensorflow as tf
|
||||
@@ -68,127 +69,6 @@ TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST = [
|
||||
]
|
||||
|
||||
|
||||
# Copied from transformers.models.bert.modeling_tf_bert.TFBertWordEmbeddings
|
||||
class TFRobertaWordEmbeddings(tf.keras.layers.Layer):
|
||||
def __init__(self, vocab_size: int, hidden_size: int, initializer_range: float, **kwargs):
|
||||
super().__init__(**kwargs)
|
||||
|
||||
self.vocab_size = vocab_size
|
||||
self.hidden_size = hidden_size
|
||||
self.initializer_range = initializer_range
|
||||
|
||||
def build(self, input_shape: tf.TensorShape):
|
||||
self.weight = self.add_weight(
|
||||
name="weight",
|
||||
shape=[self.vocab_size, self.hidden_size],
|
||||
initializer=get_initializer(self.initializer_range),
|
||||
)
|
||||
|
||||
super().build(input_shape)
|
||||
|
||||
def get_config(self) -> Dict[str, Any]:
|
||||
config = {
|
||||
"vocab_size": self.vocab_size,
|
||||
"hidden_size": self.hidden_size,
|
||||
"initializer_range": self.initializer_range,
|
||||
}
|
||||
base_config = super().get_config()
|
||||
|
||||
return dict(list(base_config.items()) + list(config.items()))
|
||||
|
||||
def call(self, input_ids: tf.Tensor) -> tf.Tensor:
|
||||
flat_input_ids = tf.reshape(tensor=input_ids, shape=[-1])
|
||||
embeddings = tf.gather(params=self.weight, indices=flat_input_ids)
|
||||
embeddings = tf.reshape(
|
||||
tensor=embeddings, shape=tf.concat(values=[shape_list(input_ids), [self.hidden_size]], axis=0)
|
||||
)
|
||||
|
||||
embeddings.set_shape(input_ids.shape.as_list() + [self.hidden_size])
|
||||
|
||||
return embeddings
|
||||
|
||||
|
||||
# Copied from transformers.models.bert.modeling_tf_bert.TFBertTokenTypeEmbeddings
|
||||
class TFRobertaTokenTypeEmbeddings(tf.keras.layers.Layer):
|
||||
def __init__(self, type_vocab_size: int, hidden_size: int, initializer_range: float, **kwargs):
|
||||
super().__init__(**kwargs)
|
||||
|
||||
self.type_vocab_size = type_vocab_size
|
||||
self.hidden_size = hidden_size
|
||||
self.initializer_range = initializer_range
|
||||
|
||||
def build(self, input_shape: tf.TensorShape):
|
||||
self.token_type_embeddings = self.add_weight(
|
||||
name="embeddings",
|
||||
shape=[self.type_vocab_size, self.hidden_size],
|
||||
initializer=get_initializer(self.initializer_range),
|
||||
)
|
||||
|
||||
super().build(input_shape)
|
||||
|
||||
def get_config(self) -> Dict[str, Any]:
|
||||
config = {
|
||||
"type_vocab_size": self.type_vocab_size,
|
||||
"hidden_size": self.hidden_size,
|
||||
"initializer_range": self.initializer_range,
|
||||
}
|
||||
base_config = super().get_config()
|
||||
|
||||
return dict(list(base_config.items()) + list(config.items()))
|
||||
|
||||
def call(self, token_type_ids: tf.Tensor) -> tf.Tensor:
|
||||
flat_token_type_ids = tf.reshape(tensor=token_type_ids, shape=[-1])
|
||||
one_hot_data = tf.one_hot(indices=flat_token_type_ids, depth=self.type_vocab_size, dtype=self._compute_dtype)
|
||||
embeddings = tf.matmul(a=one_hot_data, b=self.token_type_embeddings)
|
||||
embeddings = tf.reshape(
|
||||
tensor=embeddings, shape=tf.concat(values=[shape_list(token_type_ids), [self.hidden_size]], axis=0)
|
||||
)
|
||||
|
||||
embeddings.set_shape(token_type_ids.shape.as_list() + [self.hidden_size])
|
||||
|
||||
return embeddings
|
||||
|
||||
|
||||
# Copied from transformers.models.longformer.modeling_tf_longformer.TFLongformerPositionEmbeddings
|
||||
class TFRobertaPositionEmbeddings(tf.keras.layers.Layer):
|
||||
def __init__(self, max_position_embeddings: int, hidden_size: int, initializer_range: float, **kwargs):
|
||||
super().__init__(**kwargs)
|
||||
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
self.hidden_size = hidden_size
|
||||
self.initializer_range = initializer_range
|
||||
|
||||
def build(self, input_shape):
|
||||
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 get_config(self):
|
||||
config = {
|
||||
"max_position_embeddings": self.max_position_embeddings,
|
||||
"hidden_size": self.hidden_size,
|
||||
"initializer_range": self.initializer_range,
|
||||
}
|
||||
base_config = super().get_config()
|
||||
|
||||
return dict(list(base_config.items()) + list(config.items()))
|
||||
|
||||
def call(self, position_ids):
|
||||
flat_position_ids = tf.reshape(tensor=position_ids, shape=[-1])
|
||||
embeddings = tf.gather(params=self.position_embeddings, indices=flat_position_ids)
|
||||
embeddings = tf.reshape(
|
||||
tensor=embeddings, shape=tf.concat(values=[shape_list(position_ids), [self.hidden_size]], axis=0)
|
||||
)
|
||||
|
||||
embeddings.set_shape(position_ids.shape.as_list() + [self.hidden_size])
|
||||
|
||||
return embeddings
|
||||
|
||||
|
||||
class TFRobertaEmbeddings(tf.keras.layers.Layer):
|
||||
"""
|
||||
Same as BertEmbeddings with a tiny tweak for positional embeddings indexing.
|
||||
@@ -198,28 +78,39 @@ class TFRobertaEmbeddings(tf.keras.layers.Layer):
|
||||
super().__init__(**kwargs)
|
||||
|
||||
self.padding_idx = 1
|
||||
self.word_embeddings = TFRobertaWordEmbeddings(
|
||||
vocab_size=config.vocab_size,
|
||||
hidden_size=config.hidden_size,
|
||||
initializer_range=config.initializer_range,
|
||||
name="word_embeddings",
|
||||
)
|
||||
self.position_embeddings = TFRobertaPositionEmbeddings(
|
||||
max_position_embeddings=config.max_position_embeddings,
|
||||
hidden_size=config.hidden_size,
|
||||
initializer_range=config.initializer_range,
|
||||
name="position_embeddings",
|
||||
)
|
||||
self.token_type_embeddings = TFRobertaTokenTypeEmbeddings(
|
||||
type_vocab_size=config.type_vocab_size,
|
||||
hidden_size=config.hidden_size,
|
||||
initializer_range=config.initializer_range,
|
||||
name="token_type_embeddings",
|
||||
)
|
||||
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.embeddings_sum = tf.keras.layers.Add()
|
||||
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(initializer_range=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(initializer_range=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(initializer_range=self.initializer_range),
|
||||
)
|
||||
|
||||
super().build(input_shape)
|
||||
|
||||
def create_position_ids_from_input_ids(self, input_ids):
|
||||
"""
|
||||
Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding
|
||||
@@ -227,36 +118,13 @@ class TFRobertaEmbeddings(tf.keras.layers.Layer):
|
||||
|
||||
Args:
|
||||
input_ids: tf.Tensor
|
||||
|
||||
Returns: tf.Tensor
|
||||
"""
|
||||
input_ids_shape = shape_list(tensor=input_ids)
|
||||
|
||||
# multiple choice has 3 dimensions
|
||||
if len(input_ids_shape) == 3:
|
||||
input_ids = tf.reshape(
|
||||
tensor=input_ids, shape=(input_ids_shape[0] * input_ids_shape[1], input_ids_shape[2])
|
||||
)
|
||||
|
||||
mask = tf.cast(tf.math.not_equal(input_ids, self.padding_idx), dtype=input_ids.dtype)
|
||||
incremental_indices = tf.math.cumsum(mask, axis=1) * mask
|
||||
|
||||
return incremental_indices + self.padding_idx
|
||||
|
||||
def create_position_ids_from_inputs_embeds(self, inputs_embeds):
|
||||
"""
|
||||
We are provided embeddings directly. We cannot infer which are padded so just generate sequential position ids.
|
||||
|
||||
Args:
|
||||
inputs_embeds: tf.Tensor
|
||||
|
||||
Returns: tf.Tensor
|
||||
"""
|
||||
batch_size, seq_length = shape_list(tensor=inputs_embeds)[:2]
|
||||
position_ids = tf.range(start=self.padding_idx + 1, limit=seq_length + self.padding_idx + 1)[tf.newaxis, :]
|
||||
|
||||
return tf.tile(input=position_ids, multiples=(batch_size, 1))
|
||||
|
||||
def call(self, input_ids=None, position_ids=None, token_type_ids=None, inputs_embeds=None, training=False):
|
||||
"""
|
||||
Applies embedding based on inputs tensor.
|
||||
@@ -267,10 +135,11 @@ class TFRobertaEmbeddings(tf.keras.layers.Layer):
|
||||
assert not (input_ids is None and inputs_embeds is None)
|
||||
|
||||
if input_ids is not None:
|
||||
inputs_embeds = self.word_embeddings(input_ids=input_ids)
|
||||
inputs_embeds = tf.gather(params=self.weight, indices=input_ids)
|
||||
|
||||
input_shape = shape_list(inputs_embeds)[:-1]
|
||||
|
||||
if token_type_ids is None:
|
||||
input_shape = shape_list(tensor=inputs_embeds)[:-1]
|
||||
token_type_ids = tf.fill(dims=input_shape, value=0)
|
||||
|
||||
if position_ids is None:
|
||||
@@ -278,10 +147,13 @@ class TFRobertaEmbeddings(tf.keras.layers.Layer):
|
||||
# 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)
|
||||
else:
|
||||
position_ids = self.create_position_ids_from_inputs_embeds(inputs_embeds=inputs_embeds)
|
||||
position_ids = tf.range(start=self.padding_idx + 1, limit=input_shape[-1] + self.padding_idx + 1)[
|
||||
tf.newaxis, :
|
||||
]
|
||||
position_ids = tf.tile(input=position_ids, multiples=(input_shape[0], 1))
|
||||
|
||||
position_embeds = self.position_embeddings(position_ids=position_ids)
|
||||
token_type_embeds = self.token_type_embeddings(token_type_ids=token_type_ids)
|
||||
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 = self.embeddings_sum(inputs=[inputs_embeds, position_embeds, token_type_embeds])
|
||||
final_embeddings = self.LayerNorm(inputs=final_embeddings)
|
||||
final_embeddings = self.dropout(inputs=final_embeddings, training=training)
|
||||
@@ -321,31 +193,29 @@ class TFRobertaSelfAttention(tf.keras.layers.Layer):
|
||||
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.experimental.EinsumDense(
|
||||
equation="abc,cde->abde",
|
||||
output_shape=(None, config.num_attention_heads, self.attention_head_size),
|
||||
bias_axes="de",
|
||||
kernel_initializer=get_initializer(config.initializer_range),
|
||||
name="query",
|
||||
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.experimental.EinsumDense(
|
||||
equation="abc,cde->abde",
|
||||
output_shape=(None, config.num_attention_heads, self.attention_head_size),
|
||||
bias_axes="de",
|
||||
kernel_initializer=get_initializer(config.initializer_range),
|
||||
name="key",
|
||||
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.experimental.EinsumDense(
|
||||
equation="abc,cde->abde",
|
||||
output_shape=(None, config.num_attention_heads, self.attention_head_size),
|
||||
bias_axes="de",
|
||||
kernel_initializer=get_initializer(config.initializer_range),
|
||||
name="value",
|
||||
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)
|
||||
|
||||
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,
|
||||
@@ -354,15 +224,20 @@ class TFRobertaSelfAttention(tf.keras.layers.Layer):
|
||||
output_attentions: bool,
|
||||
training: bool = False,
|
||||
) -> Tuple[tf.Tensor]:
|
||||
query_layer = self.query(inputs=hidden_states)
|
||||
key_layer = self.key(inputs=hidden_states)
|
||||
value_layer = self.value(inputs=hidden_states)
|
||||
batch_size = shape_list(hidden_states)[0]
|
||||
mixed_query_layer = self.query(inputs=hidden_states)
|
||||
mixed_key_layer = self.key(inputs=hidden_states)
|
||||
mixed_value_layer = self.value(inputs=hidden_states)
|
||||
query_layer = self.transpose_for_scores(mixed_query_layer, batch_size)
|
||||
key_layer = self.transpose_for_scores(mixed_key_layer, batch_size)
|
||||
value_layer = self.transpose_for_scores(mixed_value_layer, batch_size)
|
||||
|
||||
# Take the dot product between "query" and "key" to get the raw
|
||||
# attention scores.
|
||||
dk = tf.cast(self.attention_head_size, dtype=query_layer.dtype)
|
||||
query_layer = tf.multiply(query_layer, tf.math.rsqrt(dk))
|
||||
attention_scores = tf.einsum("aecd,abcd->acbe", key_layer, query_layer)
|
||||
# (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)
|
||||
@@ -379,7 +254,11 @@ class TFRobertaSelfAttention(tf.keras.layers.Layer):
|
||||
if head_mask is not None:
|
||||
attention_probs = tf.multiply(attention_probs, head_mask)
|
||||
|
||||
attention_output = tf.einsum("acbe,aecd->abcd", attention_probs, value_layer)
|
||||
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,)
|
||||
|
||||
return outputs
|
||||
@@ -390,21 +269,8 @@ class TFRobertaSelfOutput(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.attention_head_size = int(config.hidden_size / config.num_attention_heads)
|
||||
self.all_head_size = config.num_attention_heads * self.attention_head_size
|
||||
|
||||
self.dense = tf.keras.layers.experimental.EinsumDense(
|
||||
equation="abcd,cde->abe",
|
||||
output_shape=(None, self.all_head_size),
|
||||
bias_axes="e",
|
||||
kernel_initializer=get_initializer(config.initializer_range),
|
||||
name="dense",
|
||||
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)
|
||||
@@ -456,12 +322,8 @@ class TFRobertaIntermediate(tf.keras.layers.Layer):
|
||||
def __init__(self, config: RobertaConfig, **kwargs):
|
||||
super().__init__(**kwargs)
|
||||
|
||||
self.dense = tf.keras.layers.experimental.EinsumDense(
|
||||
equation="abc,cd->abd",
|
||||
output_shape=(None, config.intermediate_size),
|
||||
bias_axes="d",
|
||||
kernel_initializer=get_initializer(config.initializer_range),
|
||||
name="dense",
|
||||
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):
|
||||
@@ -481,12 +343,8 @@ class TFRobertaOutput(tf.keras.layers.Layer):
|
||||
def __init__(self, config: RobertaConfig, **kwargs):
|
||||
super().__init__(**kwargs)
|
||||
|
||||
self.dense = tf.keras.layers.experimental.EinsumDense(
|
||||
equation="abc,cd->abd",
|
||||
bias_axes="d",
|
||||
output_shape=(None, config.hidden_size),
|
||||
kernel_initializer=get_initializer(config.initializer_range),
|
||||
name="dense",
|
||||
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)
|
||||
@@ -601,12 +459,12 @@ class TFRobertaMainLayer(tf.keras.layers.Layer):
|
||||
|
||||
# Copied from transformers.models.bert.modeling_tf_bert.TFBertMainLayer.get_input_embeddings
|
||||
def get_input_embeddings(self) -> tf.keras.layers.Layer:
|
||||
return self.embeddings.word_embeddings
|
||||
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.word_embeddings.weight = value
|
||||
self.embeddings.word_embeddings.vocab_size = shape_list(value)[0]
|
||||
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):
|
||||
@@ -972,7 +830,7 @@ class TFRobertaForMaskedLM(TFRobertaPreTrainedModel, TFMaskedLanguageModelingLos
|
||||
super().__init__(config, *inputs, **kwargs)
|
||||
|
||||
self.roberta = TFRobertaMainLayer(config, add_pooling_layer=False, name="roberta")
|
||||
self.lm_head = TFRobertaLMHead(config, self.roberta.embeddings.word_embeddings, name="lm_head")
|
||||
self.lm_head = TFRobertaLMHead(config, self.roberta.embeddings, name="lm_head")
|
||||
|
||||
def get_lm_head(self):
|
||||
return self.lm_head
|
||||
|
||||
Reference in New Issue
Block a user