# coding=utf-8
# Copyright 2021 The HuggingFace Inc. team. 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 RoFormer model."""
from __future__ import annotations
import math
from typing import Dict, Optional, Tuple, Union
import numpy as np
import tensorflow as tf
from ...activations_tf import get_tf_activation
from ...modeling_tf_outputs import (
TFBaseModelOutput,
TFBaseModelOutputWithPooling,
TFCausalLMOutput,
TFMaskedLMOutput,
TFMultipleChoiceModelOutput,
TFQuestionAnsweringModelOutput,
TFSequenceClassifierOutput,
TFTokenClassifierOutput,
)
from ...modeling_tf_utils import (
TFCausalLanguageModelingLoss,
TFMaskedLanguageModelingLoss,
TFModelInputType,
TFMultipleChoiceLoss,
TFPreTrainedModel,
TFQuestionAnsweringLoss,
TFSequenceClassificationLoss,
TFSequenceSummary,
TFTokenClassificationLoss,
get_initializer,
keras_serializable,
unpack_inputs,
)
from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax
from ...utils import (
add_code_sample_docstrings,
add_start_docstrings,
add_start_docstrings_to_model_forward,
logging,
)
from .configuration_roformer import RoFormerConfig
logger = logging.get_logger(__name__)
_CHECKPOINT_FOR_DOC = "junnyu/roformer_chinese_base"
_CONFIG_FOR_DOC = "RoFormerConfig"
TF_ROFORMER_PRETRAINED_MODEL_ARCHIVE_LIST = [
"junnyu/roformer_chinese_small",
"junnyu/roformer_chinese_base",
"junnyu/roformer_chinese_char_small",
"junnyu/roformer_chinese_char_base",
"junnyu/roformer_small_discriminator",
"junnyu/roformer_small_generator"
# See all RoFormer models at https://huggingface.co/models?filter=roformer
]
class TFRoFormerSinusoidalPositionalEmbedding(tf.keras.layers.Layer):
"""This module produces sinusoidal positional embeddings of any length."""
def __init__(self, num_positions: int, embedding_dim: int, **kwargs):
super().__init__(**kwargs)
if embedding_dim % 2 != 0:
raise NotImplementedError(f"odd embedding_dim {embedding_dim} not supported")
self.embedding_dim = embedding_dim
self.num_positions = num_positions
def build(self, input_shape: tf.TensorShape):
"""
Build shared token embedding layer Shared weights logic adapted from
https://github.com/tensorflow/models/blob/a009f4fb9d2fc4949e32192a944688925ef78659/official/transformer/v2/embedding_layer.py#L24
"""
weight = self._init_weight(self.num_positions, self.embedding_dim)
self.weight = self.add_weight(
name="embeddings",
shape=[self.num_positions, self.embedding_dim],
)
weight = tf.cast(weight, dtype=self.weight.dtype)
self.weight.assign(weight)
super().build(input_shape)
@staticmethod
def _init_weight(n_pos: int, dim: int):
"""
Identical to the XLM create_sinusoidal_embeddings except features are not interleaved. The cos features are in
the 2nd half of the vector. [dim // 2:]
"""
position_enc = np.array(
[[pos / np.power(10000, 2 * (j // 2) / dim) for j in range(dim)] for pos in range(n_pos)]
)
table = np.zeros_like(position_enc)
# index 0 is all zero
table[:, 0 : dim // 2] = np.sin(position_enc[:, 0::2])
table[:, dim // 2 :] = np.cos(position_enc[:, 1::2])
# convert to tensor
table = tf.convert_to_tensor(table)
tf.stop_gradient(table)
return table
def call(self, input_shape: tf.TensorShape, past_key_values_length: int = 0):
"""Input is expected to be of size [bsz x seqlen]."""
bsz, seq_len = input_shape[:2]
positions = tf.range(past_key_values_length, seq_len + past_key_values_length, delta=1, name="range")
return tf.gather(self.weight, positions)
class TFRoFormerEmbeddings(tf.keras.layers.Layer):
"""Construct the embeddings from word, position and token_type embeddings."""
def __init__(self, config: RoFormerConfig, **kwargs):
super().__init__(**kwargs)
self.config = config
self.embedding_size = config.embedding_size
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.config.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.config.type_vocab_size, self.embedding_size],
initializer=get_initializer(self.initializer_range),
)
super().build(input_shape)
def call(
self,
input_ids: tf.Tensor = None,
token_type_ids: tf.Tensor = None,
inputs_embeds: tf.Tensor = None,
training: bool = False,
) -> tf.Tensor:
"""
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:
check_embeddings_within_bounds(input_ids, self.config.vocab_size)
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)
token_type_embeds = tf.gather(params=self.token_type_embeddings, indices=token_type_ids)
final_embeddings = inputs_embeds + token_type_embeds
final_embeddings = self.LayerNorm(inputs=final_embeddings)
final_embeddings = self.dropout(inputs=final_embeddings, training=training)
return final_embeddings
class TFRoFormerSelfAttention(tf.keras.layers.Layer):
def __init__(self, config: RoFormerConfig, **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.rotary_value = config.rotary_value
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,
sinusoidal_pos: tf.Tensor,
head_mask: 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)
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)
if sinusoidal_pos is not None:
if self.rotary_value:
query_layer, key_layer, value_layer = self.apply_rotary_position_embeddings(
sinusoidal_pos, query_layer, key_layer, value_layer
)
else:
query_layer, key_layer = self.apply_rotary_position_embeddings(sinusoidal_pos, query_layer, key_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 TFRoFormerModel call() function)
attention_scores = tf.add(attention_scores, attention_mask)
# Normalize the attention scores to probabilities.
attention_probs = stable_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,)
return outputs
@staticmethod
def apply_rotary_position_embeddings(sinusoidal_pos, query_layer, key_layer, value_layer=None):
# https://kexue.fm/archives/8265
# sin [batch_size, num_heads, sequence_length, embed_size_per_head//2]
# cos [batch_size, num_heads, sequence_length, embed_size_per_head//2]
sin, cos = tf.split(sinusoidal_pos, num_or_size_splits=2, axis=-1)
# sin [θ0,θ1,θ2......θd/2-1]-> sin_pos [θ0,θ0,θ1,θ1,θ2,θ2......θd/2-1,θd/2-1]
# cos [θ0,θ1,θ2......θd/2-1]-> cos_pos [θ0,θ0,θ1,θ1,θ2,θ2......θd/2-1,θd/2-1]
sin_pos = tf.repeat(sin, 2, axis=-1)
cos_pos = tf.repeat(cos, 2, axis=-1)
# rotate_half_query_layer [-q1,q0,-q3,q2......,-qd-1,qd-2]
rotate_half_query_layer = tf.stack([-query_layer[..., 1::2], query_layer[..., ::2]], axis=-1)
rotate_half_query_layer = tf.reshape(rotate_half_query_layer, shape_list(query_layer))
query_layer = query_layer * cos_pos + rotate_half_query_layer * sin_pos
# rotate_half_key_layer [-k1,k0,-k3,k2......,-kd-1,kd-2]
rotate_half_key_layer = tf.stack([-key_layer[..., 1::2], key_layer[..., ::2]], axis=-1)
rotate_half_key_layer = tf.reshape(rotate_half_key_layer, shape_list(key_layer))
key_layer = key_layer * cos_pos + rotate_half_key_layer * sin_pos
if value_layer is not None:
# rotate_half_value_layer [-v1,v0,-v3,v2......,-vd-1,vd-2]
rotate_half_value_layer = tf.stack([-value_layer[..., 1::2], value_layer[..., ::2]], axis=-1)
rotate_half_value_layer = tf.reshape(rotate_half_value_layer, shape_list(value_layer))
value_layer = value_layer * cos_pos + rotate_half_value_layer * sin_pos
return query_layer, key_layer, value_layer
return query_layer, key_layer
# Copied from transformers.models.bert.modeling_tf_bert.TFBertSelfOutput with Bert->RoFormer
class TFRoFormerSelfOutput(tf.keras.layers.Layer):
def __init__(self, config: RoFormerConfig, **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
class TFRoFormerAttention(tf.keras.layers.Layer):
def __init__(self, config: RoFormerConfig, **kwargs):
super().__init__(**kwargs)
self.self_attention = TFRoFormerSelfAttention(config, name="self")
self.dense_output = TFRoFormerSelfOutput(config, name="output")
def prune_heads(self, heads):
raise NotImplementedError
def call(
self,
input_tensor: tf.Tensor,
attention_mask: tf.Tensor,
sinusoidal_pos: tf.Tensor,
head_mask: tf.Tensor,
output_attentions: bool,
training: bool = False,
) -> Tuple[tf.Tensor]:
self_outputs = self.self_attention(
hidden_states=input_tensor,
attention_mask=attention_mask,
sinusoidal_pos=sinusoidal_pos,
head_mask=head_mask,
output_attentions=output_attentions,
training=training,
)
attention_output = self.dense_output(
hidden_states=self_outputs[0], input_tensor=input_tensor, training=training
)
outputs = (attention_output,) + self_outputs[1:] # add attentions if we output them
return outputs
# Copied from transformers.models.bert.modeling_tf_bert.TFBertIntermediate with Bert->RoFormer
class TFRoFormerIntermediate(tf.keras.layers.Layer):
def __init__(self, config: RoFormerConfig, **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->RoFormer
class TFRoFormerOutput(tf.keras.layers.Layer):
def __init__(self, config: RoFormerConfig, **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
class TFRoFormerLayer(tf.keras.layers.Layer):
def __init__(self, config: RoFormerConfig, **kwargs):
super().__init__(**kwargs)
self.attention = TFRoFormerAttention(config, name="attention")
self.intermediate = TFRoFormerIntermediate(config, name="intermediate")
self.roformer_output = TFRoFormerOutput(config, name="output")
def call(
self,
hidden_states: tf.Tensor,
attention_mask: tf.Tensor,
sinusoidal_pos: tf.Tensor,
head_mask: tf.Tensor,
output_attentions: bool,
training: bool = False,
) -> Tuple[tf.Tensor]:
attention_outputs = self.attention(
input_tensor=hidden_states,
attention_mask=attention_mask,
sinusoidal_pos=sinusoidal_pos,
head_mask=head_mask,
output_attentions=output_attentions,
training=training,
)
attention_output = attention_outputs[0]
intermediate_output = self.intermediate(hidden_states=attention_output)
layer_output = self.roformer_output(
hidden_states=intermediate_output, input_tensor=attention_output, training=training
)
outputs = (layer_output,) + attention_outputs[1:] # add attentions if we output them
return outputs
class TFRoFormerEncoder(tf.keras.layers.Layer):
def __init__(self, config: RoFormerConfig, **kwargs):
super().__init__(**kwargs)
self.embed_positions = TFRoFormerSinusoidalPositionalEmbedding(
config.max_position_embeddings,
config.hidden_size // config.num_attention_heads,
name="embed_positions",
)
self.layer = [TFRoFormerLayer(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,
output_attentions: bool,
output_hidden_states: bool,
return_dict: bool,
training: bool = False,
) -> Union[TFBaseModelOutput, Tuple[tf.Tensor]]:
all_hidden_states = () if output_hidden_states else None
all_attentions = () if output_attentions else None
# [sequence_length, embed_size_per_head] -> [batch_size, num_heads, sequence_length, embed_size_per_head]
sinusoidal_pos = self.embed_positions(shape_list(hidden_states)[:-1])[None, None, :, :]
for i, layer_module in enumerate(self.layer):
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
layer_outputs = layer_module(
hidden_states=hidden_states,
attention_mask=attention_mask,
sinusoidal_pos=sinusoidal_pos,
head_mask=head_mask[i],
output_attentions=output_attentions,
training=training,
)
hidden_states = layer_outputs[0]
if output_attentions:
all_attentions = all_attentions + (layer_outputs[1],)
# 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] if v is not None)
return TFBaseModelOutput(
last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
)
class TFRoFormerPredictionHeadTransform(tf.keras.layers.Layer):
def __init__(self, config: RoFormerConfig, **kwargs):
super().__init__(**kwargs)
self.dense = tf.keras.layers.Dense(
units=config.embedding_size,
kernel_initializer=get_initializer(config.initializer_range),
name="dense",
)
if isinstance(config.hidden_act, str):
self.transform_act_fn = get_tf_activation(config.hidden_act)
else:
self.transform_act_fn = config.hidden_act
self.LayerNorm = tf.keras.layers.LayerNormalization(epsilon=config.layer_norm_eps, name="LayerNorm")
def call(self, hidden_states: tf.Tensor) -> tf.Tensor:
hidden_states = self.dense(inputs=hidden_states)
hidden_states = self.transform_act_fn(hidden_states)
hidden_states = self.LayerNorm(inputs=hidden_states)
return hidden_states
class TFRoFormerLMPredictionHead(tf.keras.layers.Layer):
def __init__(self, config: RoFormerConfig, input_embeddings: tf.keras.layers.Layer, **kwargs):
super().__init__(**kwargs)
self.config = config
self.embedding_size = config.embedding_size
self.transform = TFRoFormerPredictionHeadTransform(config, name="transform")
# The output weights are the same as the input embeddings, but there is
# an output-only bias for each token.
self.input_embeddings = input_embeddings
def build(self, input_shape: tf.TensorShape):
self.bias = self.add_weight(shape=(self.config.vocab_size,), initializer="zeros", trainable=True, name="bias")
super().build(input_shape)
def get_output_embeddings(self) -> tf.keras.layers.Layer:
return self.input_embeddings
def set_output_embeddings(self, value: tf.Variable):
self.input_embeddings.weight = value
self.input_embeddings.vocab_size = shape_list(value)[0]
def get_bias(self) -> Dict[str, tf.Variable]:
return {"bias": self.bias}
def set_bias(self, value: tf.Variable):
self.bias = value["bias"]
self.config.vocab_size = shape_list(value["bias"])[0]
def call(self, hidden_states: tf.Tensor) -> tf.Tensor:
hidden_states = self.transform(hidden_states=hidden_states)
seq_length = shape_list(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.config.vocab_size])
hidden_states = tf.nn.bias_add(value=hidden_states, bias=self.bias)
return hidden_states
# Copied from transformers.models.bert.modeling_tf_bert.TFBertMLMHead with Bert->RoFormer
class TFRoFormerMLMHead(tf.keras.layers.Layer):
def __init__(self, config: RoFormerConfig, input_embeddings: tf.keras.layers.Layer, **kwargs):
super().__init__(**kwargs)
self.predictions = TFRoFormerLMPredictionHead(config, input_embeddings, name="predictions")
def call(self, sequence_output: tf.Tensor) -> tf.Tensor:
prediction_scores = self.predictions(hidden_states=sequence_output)
return prediction_scores
@keras_serializable
class TFRoFormerMainLayer(tf.keras.layers.Layer):
config_class = RoFormerConfig
def __init__(self, config: RoFormerConfig, add_pooling_layer: bool = True, **kwargs):
super().__init__(**kwargs)
self.config = config
self.embeddings = TFRoFormerEmbeddings(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 = TFRoFormerEncoder(config, name="encoder")
def get_input_embeddings(self) -> tf.keras.layers.Layer:
return self.embeddings
def set_input_embeddings(self, value: tf.Variable):
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
@unpack_inputs
def call(
self,
input_ids: TFModelInputType | None = None,
attention_mask: np.ndarray | tf.Tensor | None = None,
token_type_ids: np.ndarray | tf.Tensor | None = None,
head_mask: np.ndarray | tf.Tensor | None = None,
inputs_embeds: np.ndarray | tf.Tensor | None = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
training: bool = False,
) -> Union[TFBaseModelOutput, Tuple[tf.Tensor]]:
if input_ids is not None and inputs_embeds is not None:
raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
elif input_ids is not None:
input_shape = shape_list(input_ids)
elif inputs_embeds is not None:
input_shape = shape_list(inputs_embeds)[:-1]
else:
raise ValueError("You have to specify either input_ids or inputs_embeds")
if attention_mask is None:
attention_mask = tf.fill(dims=input_shape, value=1)
if token_type_ids is None:
token_type_ids = tf.fill(dims=input_shape, value=0)
embedding_output = self.embeddings(
input_ids=input_ids,
token_type_ids=token_type_ids,
inputs_embeds=inputs_embeds,
training=training,
)
if hasattr(self, "embeddings_project"):
embedding_output = self.embeddings_project(embedding_output, training=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.
extended_attention_mask = tf.reshape(attention_mask, (input_shape[0], 1, 1, input_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)
# 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 head_mask is not None:
raise NotImplementedError
else:
head_mask = [None] * self.config.num_hidden_layers
encoder_outputs = self.encoder(
hidden_states=embedding_output,
attention_mask=extended_attention_mask,
head_mask=head_mask,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
training=training,
)
sequence_output = encoder_outputs[0]
if not return_dict:
return (sequence_output,) + encoder_outputs[1:]
return TFBaseModelOutput(
last_hidden_state=sequence_output,
hidden_states=encoder_outputs.hidden_states,
attentions=encoder_outputs.attentions,
)
class TFRoFormerPreTrainedModel(TFPreTrainedModel):
"""
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
models.
"""
config_class = RoFormerConfig
base_model_prefix = "roformer"
ROFORMER_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.
TensorFlow models and layers in `transformers` accept two formats as input:
- having all inputs as keyword arguments (like PyTorch models), or
- having all inputs as a list, tuple or dict in the first positional argument.
The reason the second format is supported is that Keras methods prefer this format when passing inputs to models
and layers. Because of this support, when using methods like `model.fit()` things should "just work" for you - just
pass your inputs and labels in any format that `model.fit()` supports! If, however, you want to use the second
format outside of Keras methods like `fit()` and `predict()`, such as when creating your own layers or models with
the Keras `Functional` API, 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(input_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})`
Note that when creating models and layers with
[subclassing](https://keras.io/guides/making_new_layers_and_models_via_subclassing/) then you don't need to worry
about any of this, as you can just pass inputs like you would to any other Python function!
Args:
config ([`RoFormerConfig`]): 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.
"""
ROFORMER_INPUTS_DOCSTRING = r"""
Args:
input_ids (`np.ndarray`, `tf.Tensor`, `List[tf.Tensor]` ``Dict[str, tf.Tensor]` or `Dict[str, np.ndarray]` and each example must have the shape `({0})`):
Indices of input sequence tokens in the vocabulary.
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.__call__`] and
[`PreTrainedTokenizer.encode`] for details.
[What are input IDs?](../glossary#input-ids)
attention_mask (`np.ndarray` 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 (`np.ndarray` 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)
head_mask (`np.ndarray` 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 (`np.ndarray` or `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 [`~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 RoFormer Model transformer outputing raw hidden-states without any specific head on top.",
ROFORMER_START_DOCSTRING,
)
class TFRoFormerModel(TFRoFormerPreTrainedModel):
def __init__(self, config: RoFormerConfig, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.roformer = TFRoFormerMainLayer(config, name="roformer")
@unpack_inputs
@add_start_docstrings_to_model_forward(ROFORMER_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFBaseModelOutputWithPooling,
config_class=_CONFIG_FOR_DOC,
)
def call(
self,
input_ids: TFModelInputType | None = None,
attention_mask: np.ndarray | tf.Tensor | None = None,
token_type_ids: np.ndarray | tf.Tensor | None = None,
head_mask: np.ndarray | tf.Tensor | None = None,
inputs_embeds: np.ndarray | tf.Tensor | None = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
training: Optional[bool] = False,
) -> Union[TFBaseModelOutputWithPooling, Tuple[tf.Tensor]]:
outputs = self.roformer(
input_ids=input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_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,
)
return outputs
@add_start_docstrings("""RoFormer Model with a `language modeling` head on top.""", ROFORMER_START_DOCSTRING)
class TFRoFormerForMaskedLM(TFRoFormerPreTrainedModel, TFMaskedLanguageModelingLoss):
def __init__(self, config: RoFormerConfig, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
if config.is_decoder:
logger.warning(
"If you want to use `TFRoFormerForMaskedLM` make sure `config.is_decoder=False` for "
"bi-directional self-attention."
)
self.roformer = TFRoFormerMainLayer(config, name="roformer")
self.mlm = TFRoFormerMLMHead(config, input_embeddings=self.roformer.embeddings, name="mlm___cls")
def get_lm_head(self) -> tf.keras.layers.Layer:
return self.mlm.predictions
@unpack_inputs
@add_start_docstrings_to_model_forward(ROFORMER_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFMaskedLMOutput,
config_class=_CONFIG_FOR_DOC,
)
def call(
self,
input_ids: TFModelInputType | None = None,
attention_mask: np.ndarray | tf.Tensor | None = None,
token_type_ids: np.ndarray | tf.Tensor | None = None,
head_mask: np.ndarray | tf.Tensor | None = None,
inputs_embeds: np.ndarray | tf.Tensor | None = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
labels: np.ndarray | tf.Tensor | None = None,
training: Optional[bool] = False,
) -> Union[TFMaskedLMOutput, Tuple[tf.Tensor]]:
r"""
labels (`tf.Tensor` or `np.ndarray` 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]`
"""
outputs = self.roformer(
input_ids=input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_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,
)
sequence_output = outputs[0]
prediction_scores = self.mlm(sequence_output=sequence_output, training=training)
loss = None if labels is None else self.hf_compute_loss(labels=labels, logits=prediction_scores)
if not 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,
)
@add_start_docstrings(
"""RoFormer Model with a `language modeling` head on top for CLM fine-tuning.""", ROFORMER_START_DOCSTRING
)
class TFRoFormerForCausalLM(TFRoFormerPreTrainedModel, TFCausalLanguageModelingLoss):
def __init__(self, config: RoFormerConfig, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
if not config.is_decoder:
logger.warning("If you want to use `TFRoFormerForCausalLM` as a standalone, add `is_decoder=True.`")
self.roformer = TFRoFormerMainLayer(config, name="roformer")
self.mlm = TFRoFormerMLMHead(config, input_embeddings=self.roformer.embeddings, name="mlm___cls")
def get_lm_head(self) -> tf.keras.layers.Layer:
return self.mlm.predictions
@unpack_inputs
@add_code_sample_docstrings(
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFCausalLMOutput,
config_class=_CONFIG_FOR_DOC,
)
def call(
self,
input_ids: TFModelInputType | None = None,
attention_mask: np.ndarray | tf.Tensor | None = None,
token_type_ids: np.ndarray | tf.Tensor | None = None,
head_mask: np.ndarray | tf.Tensor | None = None,
inputs_embeds: np.ndarray | tf.Tensor | None = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
labels: np.ndarray | tf.Tensor | None = None,
training: Optional[bool] = False,
) -> Union[TFCausalLMOutput, Tuple[tf.Tensor]]:
r"""
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]`.
"""
outputs = self.roformer(
input_ids=input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_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,
)
sequence_output = outputs[0]
logits = self.mlm(sequence_output=sequence_output, training=training)
loss = None
if labels is not None:
# shift labels to the left and cut last logit token
shifted_logits = logits[:, :-1]
labels = labels[:, 1:]
loss = self.hf_compute_loss(labels=labels, logits=shifted_logits)
if not return_dict:
output = (logits,) + outputs[2:]
return ((loss,) + output) if loss is not None else output
return TFCausalLMOutput(
loss=loss,
logits=logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
class TFRoFormerClassificationHead(tf.keras.layers.Layer):
"""Head for sentence-level classification tasks."""
def __init__(self, config: RoFormerConfig, *inputs, **kwargs):
super().__init__(*inputs, **kwargs)
self.dense = tf.keras.layers.Dense(
units=config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="dense"
)
self.dropout = tf.keras.layers.Dropout(rate=config.hidden_dropout_prob)
self.out_proj = tf.keras.layers.Dense(
units=config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="out_proj"
)
if isinstance(config.hidden_act, str):
self.classifier_act_fn = get_tf_activation(config.hidden_act)
else:
self.classifier_act_fn = config.hidden_act
def call(self, hidden_states: tf.Tensor, training: bool = False) -> tf.Tensor:
hidden_states = hidden_states[:, 0, :] # take token (equiv. to [CLS])
hidden_states = self.dropout(inputs=hidden_states, training=training)
hidden_states = self.dense(inputs=hidden_states)
hidden_states = self.classifier_act_fn(hidden_states)
hidden_states = self.dropout(inputs=hidden_states, training=training)
hidden_states = self.out_proj(hidden_states)
return hidden_states
@add_start_docstrings(
"""
RoFormer Model transformer with a sequence classification/regression head on top e.g., for GLUE tasks.
""",
ROFORMER_START_DOCSTRING,
)
class TFRoFormerForSequenceClassification(TFRoFormerPreTrainedModel, TFSequenceClassificationLoss):
def __init__(self, config: RoFormerConfig, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.num_labels = config.num_labels
self.roformer = TFRoFormerMainLayer(config, name="roformer")
self.classifier = TFRoFormerClassificationHead(config, name="classifier")
@unpack_inputs
@add_start_docstrings_to_model_forward(ROFORMER_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFSequenceClassifierOutput,
config_class=_CONFIG_FOR_DOC,
)
def call(
self,
input_ids: TFModelInputType | None = None,
attention_mask: np.ndarray | tf.Tensor | None = None,
token_type_ids: np.ndarray | tf.Tensor | None = None,
head_mask: np.ndarray | tf.Tensor | None = None,
inputs_embeds: np.ndarray | tf.Tensor | None = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
labels: np.ndarray | tf.Tensor | None = None,
training: Optional[bool] = False,
) -> Union[TFSequenceClassifierOutput, Tuple[tf.Tensor]]:
r"""
labels (`tf.Tensor` or `np.ndarray` 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).
"""
outputs = self.roformer(
input_ids=input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_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,
)
logits = self.classifier(hidden_states=outputs[0], training=training)
loss = None if labels is None else self.hf_compute_loss(labels=labels, logits=logits)
if not 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,
)
@add_start_docstrings(
"""
RoFormer 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.
""",
ROFORMER_START_DOCSTRING,
)
class TFRoFormerForMultipleChoice(TFRoFormerPreTrainedModel, TFMultipleChoiceLoss):
def __init__(self, config: RoFormerConfig, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.roformer = TFRoFormerMainLayer(config, name="roformer")
self.sequence_summary = TFSequenceSummary(config, config.initializer_range, name="sequence_summary")
self.classifier = tf.keras.layers.Dense(
units=1, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
)
@unpack_inputs
@add_start_docstrings_to_model_forward(
ROFORMER_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length")
)
@add_code_sample_docstrings(
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFMultipleChoiceModelOutput,
config_class=_CONFIG_FOR_DOC,
)
def call(
self,
input_ids: TFModelInputType | None = None,
attention_mask: np.ndarray | tf.Tensor | None = None,
token_type_ids: np.ndarray | tf.Tensor | None = None,
head_mask: np.ndarray | tf.Tensor | None = None,
inputs_embeds: np.ndarray | tf.Tensor | None = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
labels: np.ndarray | tf.Tensor | None = None,
training: Optional[bool] = False,
) -> Union[TFMultipleChoiceModelOutput, Tuple[tf.Tensor]]:
r"""
labels (`tf.Tensor` or `np.ndarray` 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)
"""
if input_ids is not None:
num_choices = shape_list(input_ids)[1]
seq_length = shape_list(input_ids)[2]
else:
num_choices = shape_list(inputs_embeds)[1]
seq_length = shape_list(inputs_embeds)[2]
flat_input_ids = tf.reshape(tensor=input_ids, shape=(-1, seq_length)) if input_ids is not None else None
flat_attention_mask = (
tf.reshape(tensor=attention_mask, shape=(-1, seq_length)) if attention_mask is not None else None
)
flat_token_type_ids = (
tf.reshape(tensor=token_type_ids, shape=(-1, seq_length)) if token_type_ids is not None else None
)
flat_inputs_embeds = (
tf.reshape(tensor=inputs_embeds, shape=(-1, seq_length, shape_list(inputs_embeds)[3]))
if inputs_embeds is not None
else None
)
outputs = self.roformer(
input_ids=flat_input_ids,
attention_mask=flat_attention_mask,
token_type_ids=flat_token_type_ids,
head_mask=head_mask,
inputs_embeds=flat_inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
training=training,
)
logits = self.sequence_summary(inputs=outputs[0], training=training)
logits = self.classifier(inputs=logits)
reshaped_logits = tf.reshape(tensor=logits, shape=(-1, num_choices))
loss = None if labels is None else self.hf_compute_loss(labels=labels, logits=reshaped_logits)
if not 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,
)
@add_start_docstrings(
"""
RoFormer 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.
""",
ROFORMER_START_DOCSTRING,
)
class TFRoFormerForTokenClassification(TFRoFormerPreTrainedModel, TFTokenClassificationLoss):
def __init__(self, config: RoFormerConfig, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.num_labels = config.num_labels
self.roformer = TFRoFormerMainLayer(config, name="roformer")
self.dropout = tf.keras.layers.Dropout(rate=config.hidden_dropout_prob)
self.classifier = tf.keras.layers.Dense(
units=config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
)
@unpack_inputs
@add_start_docstrings_to_model_forward(ROFORMER_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFTokenClassifierOutput,
config_class=_CONFIG_FOR_DOC,
)
def call(
self,
input_ids: TFModelInputType | None = None,
attention_mask: np.ndarray | tf.Tensor | None = None,
token_type_ids: np.ndarray | tf.Tensor | None = None,
head_mask: np.ndarray | tf.Tensor | None = None,
inputs_embeds: np.ndarray | tf.Tensor | None = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
labels: np.ndarray | tf.Tensor | None = None,
training: Optional[bool] = False,
) -> Union[TFTokenClassifierOutput, Tuple[tf.Tensor]]:
r"""
labels (`tf.Tensor` or `np.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
"""
outputs = self.roformer(
input_ids=input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_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,
)
sequence_output = outputs[0]
sequence_output = self.dropout(inputs=sequence_output, training=training)
logits = self.classifier(inputs=sequence_output)
loss = None if labels is None else self.hf_compute_loss(labels=labels, logits=logits)
if not return_dict:
output = (logits,) + outputs[1:]
return ((loss,) + output) if loss is not None else output
return TFTokenClassifierOutput(
loss=loss,
logits=logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
@add_start_docstrings(
"""
RoFormer Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear
layer on top of the hidden-states output to compute `span start logits` and `span end logits`).
""",
ROFORMER_START_DOCSTRING,
)
class TFRoFormerForQuestionAnswering(TFRoFormerPreTrainedModel, TFQuestionAnsweringLoss):
def __init__(self, config: RoFormerConfig, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.num_labels = config.num_labels
self.roformer = TFRoFormerMainLayer(config, name="roformer")
self.qa_outputs = tf.keras.layers.Dense(
units=config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="qa_outputs"
)
@unpack_inputs
@add_start_docstrings_to_model_forward(ROFORMER_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
@add_code_sample_docstrings(
checkpoint=_CHECKPOINT_FOR_DOC,
output_type=TFQuestionAnsweringModelOutput,
config_class=_CONFIG_FOR_DOC,
)
def call(
self,
input_ids: TFModelInputType | None = None,
attention_mask: np.ndarray | tf.Tensor | None = None,
token_type_ids: np.ndarray | tf.Tensor | None = None,
head_mask: np.ndarray | tf.Tensor | None = None,
inputs_embeds: np.ndarray | tf.Tensor | None = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
start_positions: np.ndarray | tf.Tensor | None = None,
end_positions: np.ndarray | tf.Tensor | None = None,
training: Optional[bool] = False,
) -> Union[TFQuestionAnsweringModelOutput, Tuple[tf.Tensor]]:
r"""
start_positions (`tf.Tensor` or `np.ndarray` 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` or `np.ndarray` 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.
"""
outputs = self.roformer(
input_ids=input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_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,
)
sequence_output = outputs[0]
logits = self.qa_outputs(inputs=sequence_output)
start_logits, end_logits = tf.split(value=logits, num_or_size_splits=2, axis=-1)
start_logits = tf.squeeze(input=start_logits, axis=-1)
end_logits = tf.squeeze(input=end_logits, axis=-1)
loss = None
if start_positions is not None and end_positions is not None:
labels = {"start_position": start_positions, "end_position": end_positions}
loss = self.hf_compute_loss(labels=labels, logits=(start_logits, end_logits))
if not 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,
)