# coding=utf-8 # Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team. # Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """TF 2.0 MobileBERT model.""" from __future__ import annotations import warnings from dataclasses import dataclass from typing import 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, TFMaskedLMOutput, TFMultipleChoiceModelOutput, TFNextSentencePredictorOutput, TFQuestionAnsweringModelOutput, TFSequenceClassifierOutput, TFTokenClassifierOutput, ) from ...modeling_tf_utils import ( TFMaskedLanguageModelingLoss, TFModelInputType, TFMultipleChoiceLoss, TFNextSentencePredictionLoss, TFPreTrainedModel, TFQuestionAnsweringLoss, TFSequenceClassificationLoss, TFTokenClassificationLoss, get_initializer, keras, keras_serializable, unpack_inputs, ) from ...tf_utils import check_embeddings_within_bounds, shape_list, stable_softmax from ...utils import ( ModelOutput, add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings, ) from .configuration_mobilebert import MobileBertConfig logger = logging.get_logger(__name__) _CHECKPOINT_FOR_DOC = "google/mobilebert-uncased" _CONFIG_FOR_DOC = "MobileBertConfig" # TokenClassification docstring _CHECKPOINT_FOR_TOKEN_CLASSIFICATION = "vumichien/mobilebert-finetuned-ner" _TOKEN_CLASS_EXPECTED_OUTPUT = "['I-ORG', 'I-ORG', 'O', 'O', 'O', 'O', 'O', 'I-LOC', 'O', 'I-LOC', 'I-LOC']" _TOKEN_CLASS_EXPECTED_LOSS = 0.03 # QuestionAnswering docstring _CHECKPOINT_FOR_QA = "vumichien/mobilebert-uncased-squad-v2" _QA_EXPECTED_OUTPUT = "'a nice puppet'" _QA_EXPECTED_LOSS = 3.98 _QA_TARGET_START_INDEX = 12 _QA_TARGET_END_INDEX = 13 # SequenceClassification docstring _CHECKPOINT_FOR_SEQUENCE_CLASSIFICATION = "vumichien/emo-mobilebert" _SEQ_CLASS_EXPECTED_OUTPUT = "'others'" _SEQ_CLASS_EXPECTED_LOSS = "4.72" # Copied from transformers.models.bert.modeling_tf_bert.TFBertPreTrainingLoss class TFMobileBertPreTrainingLoss: """ Loss function suitable for BERT-like pretraining, that is, the task of pretraining a language model by combining NSP + MLM. .. note:: Any label of -100 will be ignored (along with the corresponding logits) in the loss computation. """ def hf_compute_loss(self, labels: tf.Tensor, logits: tf.Tensor) -> tf.Tensor: loss_fn = keras.losses.SparseCategoricalCrossentropy(from_logits=True, reduction=keras.losses.Reduction.NONE) # Clip negative labels to zero here to avoid NaNs and errors - those positions will get masked later anyway unmasked_lm_losses = loss_fn(y_true=tf.nn.relu(labels["labels"]), y_pred=logits[0]) # make sure only labels that are not equal to -100 # are taken into account for the loss computation lm_loss_mask = tf.cast(labels["labels"] != -100, dtype=unmasked_lm_losses.dtype) masked_lm_losses = unmasked_lm_losses * lm_loss_mask reduced_masked_lm_loss = tf.reduce_sum(masked_lm_losses) / tf.reduce_sum(lm_loss_mask) # Clip negative labels to zero here to avoid NaNs and errors - those positions will get masked later anyway unmasked_ns_loss = loss_fn(y_true=tf.nn.relu(labels["next_sentence_label"]), y_pred=logits[1]) ns_loss_mask = tf.cast(labels["next_sentence_label"] != -100, dtype=unmasked_ns_loss.dtype) masked_ns_loss = unmasked_ns_loss * ns_loss_mask reduced_masked_ns_loss = tf.reduce_sum(masked_ns_loss) / tf.reduce_sum(ns_loss_mask) return tf.reshape(reduced_masked_lm_loss + reduced_masked_ns_loss, (1,)) class TFMobileBertIntermediate(keras.layers.Layer): def __init__(self, config, **kwargs): super().__init__(**kwargs) self.dense = keras.layers.Dense(config.intermediate_size, 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 self.config = config def call(self, hidden_states): hidden_states = self.dense(hidden_states) hidden_states = self.intermediate_act_fn(hidden_states) return hidden_states def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "dense", None) is not None: with tf.name_scope(self.dense.name): self.dense.build([None, None, self.config.true_hidden_size]) class TFLayerNorm(keras.layers.LayerNormalization): def __init__(self, feat_size, *args, **kwargs): self.feat_size = feat_size super().__init__(*args, **kwargs) def build(self, input_shape=None): super().build([None, None, self.feat_size]) class TFNoNorm(keras.layers.Layer): def __init__(self, feat_size, epsilon=None, **kwargs): super().__init__(**kwargs) self.feat_size = feat_size def build(self, input_shape): self.bias = self.add_weight("bias", shape=[self.feat_size], initializer="zeros") self.weight = self.add_weight("weight", shape=[self.feat_size], initializer="ones") super().build(input_shape) def call(self, inputs: tf.Tensor): return inputs * self.weight + self.bias NORM2FN = {"layer_norm": TFLayerNorm, "no_norm": TFNoNorm} class TFMobileBertEmbeddings(keras.layers.Layer): """Construct the embeddings from word, position and token_type embeddings.""" def __init__(self, config, **kwargs): super().__init__(**kwargs) self.trigram_input = config.trigram_input self.embedding_size = config.embedding_size self.config = config self.hidden_size = config.hidden_size self.max_position_embeddings = config.max_position_embeddings self.initializer_range = config.initializer_range self.embedding_transformation = keras.layers.Dense(config.hidden_size, name="embedding_transformation") # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load # any TensorFlow checkpoint file self.LayerNorm = NORM2FN[config.normalization_type]( config.hidden_size, epsilon=config.layer_norm_eps, name="LayerNorm" ) self.dropout = keras.layers.Dropout(rate=config.hidden_dropout_prob) self.embedded_input_size = self.embedding_size * (3 if self.trigram_input else 1) def build(self, input_shape=None): with tf.name_scope("word_embeddings"): self.weight = self.add_weight( name="weight", shape=[self.config.vocab_size, self.embedding_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.config.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), ) if self.built: return self.built = True if getattr(self, "embedding_transformation", None) is not None: with tf.name_scope(self.embedding_transformation.name): self.embedding_transformation.build([None, None, self.embedded_input_size]) if getattr(self, "LayerNorm", None) is not None: with tf.name_scope(self.LayerNorm.name): self.LayerNorm.build(None) def call(self, input_ids=None, position_ids=None, token_type_ids=None, inputs_embeds=None, training=False): """ Applies embedding based on inputs tensor. Returns: final_embeddings (`tf.Tensor`): output embedding tensor. """ assert not (input_ids is None and inputs_embeds is None) if input_ids is not None: 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) if self.trigram_input: # From the paper MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited # Devices (https://arxiv.org/abs/2004.02984) # # The embedding table in BERT models accounts for a substantial proportion of model size. To compress # the embedding layer, we reduce the embedding dimension to 128 in MobileBERT. # Then, we apply a 1D convolution with kernel size 3 on the raw token embedding to produce a 512 # dimensional output. inputs_embeds = tf.concat( [ tf.pad(inputs_embeds[:, 1:], ((0, 0), (0, 1), (0, 0))), inputs_embeds, tf.pad(inputs_embeds[:, :-1], ((0, 0), (1, 0), (0, 0))), ], axis=2, ) if self.trigram_input or self.embedding_size != self.hidden_size: inputs_embeds = self.embedding_transformation(inputs_embeds) if position_ids is None: position_ids = tf.expand_dims(tf.range(start=0, limit=input_shape[-1]), axis=0) position_embeds = tf.gather(params=self.position_embeddings, indices=position_ids) token_type_embeds = tf.gather(params=self.token_type_embeddings, indices=token_type_ids) final_embeddings = inputs_embeds + position_embeds + token_type_embeds final_embeddings = self.LayerNorm(inputs=final_embeddings) final_embeddings = self.dropout(inputs=final_embeddings, training=training) return final_embeddings class TFMobileBertSelfAttention(keras.layers.Layer): def __init__(self, config, **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 of attention " f"heads ({config.num_attention_heads}" ) self.num_attention_heads = config.num_attention_heads self.output_attentions = config.output_attentions assert config.hidden_size % config.num_attention_heads == 0 self.attention_head_size = int(config.true_hidden_size / config.num_attention_heads) self.all_head_size = self.num_attention_heads * self.attention_head_size self.query = keras.layers.Dense( self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="query" ) self.key = keras.layers.Dense( self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="key" ) self.value = keras.layers.Dense( self.all_head_size, kernel_initializer=get_initializer(config.initializer_range), name="value" ) self.dropout = keras.layers.Dropout(config.attention_probs_dropout_prob) self.config = config def transpose_for_scores(self, x, batch_size): # Reshape from [batch_size, seq_length, all_head_size] to [batch_size, seq_length, num_attention_heads, attention_head_size] x = tf.reshape(x, (batch_size, -1, self.num_attention_heads, self.attention_head_size)) return tf.transpose(x, perm=[0, 2, 1, 3]) def call( self, query_tensor, key_tensor, value_tensor, attention_mask, head_mask, output_attentions, training=False ): batch_size = shape_list(attention_mask)[0] mixed_query_layer = self.query(query_tensor) mixed_key_layer = self.key(key_tensor) mixed_value_layer = self.value(value_tensor) 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. attention_scores = tf.matmul( query_layer, key_layer, transpose_b=True ) # (batch size, num_heads, seq_len_q, seq_len_k) dk = tf.cast(shape_list(key_layer)[-1], dtype=attention_scores.dtype) # scale attention_scores attention_scores = attention_scores / tf.math.sqrt(dk) if attention_mask is not None: # Apply the attention mask is (precomputed for all layers in TFMobileBertModel call() function) attention_mask = tf.cast(attention_mask, dtype=attention_scores.dtype) attention_scores = attention_scores + attention_mask # Normalize the attention scores to probabilities. attention_probs = stable_softmax(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(attention_probs, training=training) # Mask heads if we want to if head_mask is not None: attention_probs = attention_probs * head_mask context_layer = tf.matmul(attention_probs, value_layer) context_layer = tf.transpose(context_layer, perm=[0, 2, 1, 3]) context_layer = tf.reshape( context_layer, (batch_size, -1, self.all_head_size) ) # (batch_size, seq_len_q, all_head_size) outputs = (context_layer, attention_probs) if output_attentions else (context_layer,) return outputs def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "query", None) is not None: with tf.name_scope(self.query.name): self.query.build([None, None, self.config.true_hidden_size]) if getattr(self, "key", None) is not None: with tf.name_scope(self.key.name): self.key.build([None, None, self.config.true_hidden_size]) if getattr(self, "value", None) is not None: with tf.name_scope(self.value.name): self.value.build( [ None, None, self.config.true_hidden_size if self.config.use_bottleneck_attention else self.config.hidden_size, ] ) class TFMobileBertSelfOutput(keras.layers.Layer): def __init__(self, config, **kwargs): super().__init__(**kwargs) self.use_bottleneck = config.use_bottleneck self.dense = keras.layers.Dense( config.true_hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="dense" ) self.LayerNorm = NORM2FN[config.normalization_type]( config.true_hidden_size, epsilon=config.layer_norm_eps, name="LayerNorm" ) if not self.use_bottleneck: self.dropout = keras.layers.Dropout(config.hidden_dropout_prob) self.config = config def call(self, hidden_states, residual_tensor, training=False): hidden_states = self.dense(hidden_states) if not self.use_bottleneck: hidden_states = self.dropout(hidden_states, training=training) hidden_states = self.LayerNorm(hidden_states + residual_tensor) return hidden_states def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "dense", None) is not None: with tf.name_scope(self.dense.name): self.dense.build([None, None, self.config.true_hidden_size]) if getattr(self, "LayerNorm", None) is not None: with tf.name_scope(self.LayerNorm.name): self.LayerNorm.build(None) class TFMobileBertAttention(keras.layers.Layer): def __init__(self, config, **kwargs): super().__init__(**kwargs) self.self = TFMobileBertSelfAttention(config, name="self") self.mobilebert_output = TFMobileBertSelfOutput(config, name="output") def prune_heads(self, heads): raise NotImplementedError def call( self, query_tensor, key_tensor, value_tensor, layer_input, attention_mask, head_mask, output_attentions, training=False, ): self_outputs = self.self( query_tensor, key_tensor, value_tensor, attention_mask, head_mask, output_attentions, training=training ) attention_output = self.mobilebert_output(self_outputs[0], layer_input, training=training) outputs = (attention_output,) + self_outputs[1:] # add attentions if we output them return outputs def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "self", None) is not None: with tf.name_scope(self.self.name): self.self.build(None) if getattr(self, "mobilebert_output", None) is not None: with tf.name_scope(self.mobilebert_output.name): self.mobilebert_output.build(None) class TFOutputBottleneck(keras.layers.Layer): def __init__(self, config, **kwargs): super().__init__(**kwargs) self.dense = keras.layers.Dense(config.hidden_size, name="dense") self.LayerNorm = NORM2FN[config.normalization_type]( config.hidden_size, epsilon=config.layer_norm_eps, name="LayerNorm" ) self.dropout = keras.layers.Dropout(config.hidden_dropout_prob) self.config = config def call(self, hidden_states, residual_tensor, training=False): layer_outputs = self.dense(hidden_states) layer_outputs = self.dropout(layer_outputs, training=training) layer_outputs = self.LayerNorm(layer_outputs + residual_tensor) return layer_outputs def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "dense", None) is not None: with tf.name_scope(self.dense.name): self.dense.build([None, None, self.config.true_hidden_size]) if getattr(self, "LayerNorm", None) is not None: with tf.name_scope(self.LayerNorm.name): self.LayerNorm.build(None) class TFMobileBertOutput(keras.layers.Layer): def __init__(self, config, **kwargs): super().__init__(**kwargs) self.use_bottleneck = config.use_bottleneck self.dense = keras.layers.Dense( config.true_hidden_size, kernel_initializer=get_initializer(config.initializer_range), name="dense" ) self.LayerNorm = NORM2FN[config.normalization_type]( config.true_hidden_size, epsilon=config.layer_norm_eps, name="LayerNorm" ) if not self.use_bottleneck: self.dropout = keras.layers.Dropout(config.hidden_dropout_prob) else: self.bottleneck = TFOutputBottleneck(config, name="bottleneck") self.config = config def call(self, hidden_states, residual_tensor_1, residual_tensor_2, training=False): hidden_states = self.dense(hidden_states) if not self.use_bottleneck: hidden_states = self.dropout(hidden_states, training=training) hidden_states = self.LayerNorm(hidden_states + residual_tensor_1) else: hidden_states = self.LayerNorm(hidden_states + residual_tensor_1) hidden_states = self.bottleneck(hidden_states, residual_tensor_2) return hidden_states def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "dense", None) is not None: with tf.name_scope(self.dense.name): self.dense.build([None, None, self.config.intermediate_size]) if getattr(self, "LayerNorm", None) is not None: with tf.name_scope(self.LayerNorm.name): self.LayerNorm.build(None) if getattr(self, "bottleneck", None) is not None: with tf.name_scope(self.bottleneck.name): self.bottleneck.build(None) class TFBottleneckLayer(keras.layers.Layer): def __init__(self, config, **kwargs): super().__init__(**kwargs) self.dense = keras.layers.Dense(config.intra_bottleneck_size, name="dense") self.LayerNorm = NORM2FN[config.normalization_type]( config.intra_bottleneck_size, epsilon=config.layer_norm_eps, name="LayerNorm" ) self.config = config def call(self, inputs): hidden_states = self.dense(inputs) hidden_states = self.LayerNorm(hidden_states) return hidden_states def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "dense", None) is not None: with tf.name_scope(self.dense.name): self.dense.build([None, None, self.config.hidden_size]) if getattr(self, "LayerNorm", None) is not None: with tf.name_scope(self.LayerNorm.name): self.LayerNorm.build(None) class TFBottleneck(keras.layers.Layer): def __init__(self, config, **kwargs): super().__init__(**kwargs) self.key_query_shared_bottleneck = config.key_query_shared_bottleneck self.use_bottleneck_attention = config.use_bottleneck_attention self.bottleneck_input = TFBottleneckLayer(config, name="input") if self.key_query_shared_bottleneck: self.attention = TFBottleneckLayer(config, name="attention") def call(self, hidden_states): # This method can return three different tuples of values. These different values make use of bottlenecks, # which are linear layers used to project the hidden states to a lower-dimensional vector, reducing memory # usage. These linear layer have weights that are learned during training. # # If `config.use_bottleneck_attention`, it will return the result of the bottleneck layer four times for the # key, query, value, and "layer input" to be used by the attention layer. # This bottleneck is used to project the hidden. This last layer input will be used as a residual tensor # in the attention self output, after the attention scores have been computed. # # If not `config.use_bottleneck_attention` and `config.key_query_shared_bottleneck`, this will return # four values, three of which have been passed through a bottleneck: the query and key, passed through the same # bottleneck, and the residual layer to be applied in the attention self output, through another bottleneck. # # Finally, in the last case, the values for the query, key and values are the hidden states without bottleneck, # and the residual layer will be this value passed through a bottleneck. bottlenecked_hidden_states = self.bottleneck_input(hidden_states) if self.use_bottleneck_attention: return (bottlenecked_hidden_states,) * 4 elif self.key_query_shared_bottleneck: shared_attention_input = self.attention(hidden_states) return (shared_attention_input, shared_attention_input, hidden_states, bottlenecked_hidden_states) else: return (hidden_states, hidden_states, hidden_states, bottlenecked_hidden_states) def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "bottleneck_input", None) is not None: with tf.name_scope(self.bottleneck_input.name): self.bottleneck_input.build(None) if getattr(self, "attention", None) is not None: with tf.name_scope(self.attention.name): self.attention.build(None) class TFFFNOutput(keras.layers.Layer): def __init__(self, config, **kwargs): super().__init__(**kwargs) self.dense = keras.layers.Dense(config.true_hidden_size, name="dense") self.LayerNorm = NORM2FN[config.normalization_type]( config.true_hidden_size, epsilon=config.layer_norm_eps, name="LayerNorm" ) self.config = config def call(self, hidden_states, residual_tensor): hidden_states = self.dense(hidden_states) hidden_states = self.LayerNorm(hidden_states + residual_tensor) return hidden_states def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "dense", None) is not None: with tf.name_scope(self.dense.name): self.dense.build([None, None, self.config.intermediate_size]) if getattr(self, "LayerNorm", None) is not None: with tf.name_scope(self.LayerNorm.name): self.LayerNorm.build(None) class TFFFNLayer(keras.layers.Layer): def __init__(self, config, **kwargs): super().__init__(**kwargs) self.intermediate = TFMobileBertIntermediate(config, name="intermediate") self.mobilebert_output = TFFFNOutput(config, name="output") def call(self, hidden_states): intermediate_output = self.intermediate(hidden_states) layer_outputs = self.mobilebert_output(intermediate_output, hidden_states) return layer_outputs def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "intermediate", None) is not None: with tf.name_scope(self.intermediate.name): self.intermediate.build(None) if getattr(self, "mobilebert_output", None) is not None: with tf.name_scope(self.mobilebert_output.name): self.mobilebert_output.build(None) class TFMobileBertLayer(keras.layers.Layer): def __init__(self, config, **kwargs): super().__init__(**kwargs) self.use_bottleneck = config.use_bottleneck self.num_feedforward_networks = config.num_feedforward_networks self.attention = TFMobileBertAttention(config, name="attention") self.intermediate = TFMobileBertIntermediate(config, name="intermediate") self.mobilebert_output = TFMobileBertOutput(config, name="output") if self.use_bottleneck: self.bottleneck = TFBottleneck(config, name="bottleneck") if config.num_feedforward_networks > 1: self.ffn = [TFFFNLayer(config, name=f"ffn.{i}") for i in range(config.num_feedforward_networks - 1)] def call(self, hidden_states, attention_mask, head_mask, output_attentions, training=False): if self.use_bottleneck: query_tensor, key_tensor, value_tensor, layer_input = self.bottleneck(hidden_states) else: query_tensor, key_tensor, value_tensor, layer_input = [hidden_states] * 4 attention_outputs = self.attention( query_tensor, key_tensor, value_tensor, layer_input, attention_mask, head_mask, output_attentions, training=training, ) attention_output = attention_outputs[0] s = (attention_output,) if self.num_feedforward_networks != 1: for i, ffn_module in enumerate(self.ffn): attention_output = ffn_module(attention_output) s += (attention_output,) intermediate_output = self.intermediate(attention_output) layer_output = self.mobilebert_output(intermediate_output, attention_output, hidden_states, training=training) outputs = ( (layer_output,) + attention_outputs[1:] + ( tf.constant(0), query_tensor, key_tensor, value_tensor, layer_input, attention_output, intermediate_output, ) + s ) # add attentions if we output them return outputs def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "attention", None) is not None: with tf.name_scope(self.attention.name): self.attention.build(None) if getattr(self, "intermediate", None) is not None: with tf.name_scope(self.intermediate.name): self.intermediate.build(None) if getattr(self, "mobilebert_output", None) is not None: with tf.name_scope(self.mobilebert_output.name): self.mobilebert_output.build(None) if getattr(self, "bottleneck", None) is not None: with tf.name_scope(self.bottleneck.name): self.bottleneck.build(None) if getattr(self, "ffn", None) is not None: for layer in self.ffn: with tf.name_scope(layer.name): layer.build(None) class TFMobileBertEncoder(keras.layers.Layer): def __init__(self, config, **kwargs): super().__init__(**kwargs) self.output_attentions = config.output_attentions self.output_hidden_states = config.output_hidden_states self.layer = [TFMobileBertLayer(config, name=f"layer_._{i}") for i in range(config.num_hidden_layers)] def call( self, hidden_states, attention_mask, head_mask, output_attentions, output_hidden_states, return_dict, training=False, ): all_hidden_states = () if output_hidden_states else None all_attentions = () if output_attentions else 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, attention_mask, head_mask[i], 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 ) def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "layer", None) is not None: for layer in self.layer: with tf.name_scope(layer.name): layer.build(None) class TFMobileBertPooler(keras.layers.Layer): def __init__(self, config, **kwargs): super().__init__(**kwargs) self.do_activate = config.classifier_activation if self.do_activate: self.dense = keras.layers.Dense( config.hidden_size, kernel_initializer=get_initializer(config.initializer_range), activation="tanh", name="dense", ) self.config = config def call(self, hidden_states): # We "pool" the model by simply taking the hidden state corresponding # to the first token. first_token_tensor = hidden_states[:, 0] if not self.do_activate: return first_token_tensor else: pooled_output = self.dense(first_token_tensor) return pooled_output def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "dense", None) is not None: with tf.name_scope(self.dense.name): self.dense.build([None, None, self.config.hidden_size]) class TFMobileBertPredictionHeadTransform(keras.layers.Layer): def __init__(self, config, **kwargs): super().__init__(**kwargs) self.dense = keras.layers.Dense( config.hidden_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 = NORM2FN["layer_norm"](config.hidden_size, epsilon=config.layer_norm_eps, name="LayerNorm") self.config = config def call(self, hidden_states): hidden_states = self.dense(hidden_states) hidden_states = self.transform_act_fn(hidden_states) hidden_states = self.LayerNorm(hidden_states) return hidden_states def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "dense", None) is not None: with tf.name_scope(self.dense.name): self.dense.build([None, None, self.config.hidden_size]) if getattr(self, "LayerNorm", None) is not None: with tf.name_scope(self.LayerNorm.name): self.LayerNorm.build(None) class TFMobileBertLMPredictionHead(keras.layers.Layer): def __init__(self, config, **kwargs): super().__init__(**kwargs) self.transform = TFMobileBertPredictionHeadTransform(config, name="transform") self.config = config def build(self, input_shape=None): self.bias = self.add_weight(shape=(self.config.vocab_size,), initializer="zeros", trainable=True, name="bias") self.dense = self.add_weight( shape=(self.config.hidden_size - self.config.embedding_size, self.config.vocab_size), initializer="zeros", trainable=True, name="dense/weight", ) self.decoder = self.add_weight( shape=(self.config.vocab_size, self.config.embedding_size), initializer="zeros", trainable=True, name="decoder/weight", ) if self.built: return self.built = True if getattr(self, "transform", None) is not None: with tf.name_scope(self.transform.name): self.transform.build(None) def get_output_embeddings(self): return self def set_output_embeddings(self, value): self.decoder = value self.config.vocab_size = shape_list(value)[0] def get_bias(self): return {"bias": self.bias} def set_bias(self, value): self.bias = value["bias"] self.config.vocab_size = shape_list(value["bias"])[0] def call(self, hidden_states): hidden_states = self.transform(hidden_states) hidden_states = tf.matmul(hidden_states, tf.concat([tf.transpose(self.decoder), self.dense], axis=0)) hidden_states = hidden_states + self.bias return hidden_states class TFMobileBertMLMHead(keras.layers.Layer): def __init__(self, config, **kwargs): super().__init__(**kwargs) self.predictions = TFMobileBertLMPredictionHead(config, name="predictions") def call(self, sequence_output): prediction_scores = self.predictions(sequence_output) return prediction_scores def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "predictions", None) is not None: with tf.name_scope(self.predictions.name): self.predictions.build(None) @keras_serializable class TFMobileBertMainLayer(keras.layers.Layer): config_class = MobileBertConfig def __init__(self, config, add_pooling_layer=True, **kwargs): super().__init__(**kwargs) self.config = config self.num_hidden_layers = config.num_hidden_layers self.output_attentions = config.output_attentions self.output_hidden_states = config.output_hidden_states self.return_dict = config.use_return_dict self.embeddings = TFMobileBertEmbeddings(config, name="embeddings") self.encoder = TFMobileBertEncoder(config, name="encoder") self.pooler = TFMobileBertPooler(config, name="pooler") if add_pooling_layer else None 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 @unpack_inputs 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, ): 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(input_shape, 1) if token_type_ids is None: token_type_ids = tf.fill(input_shape, 0) embedding_output = self.embeddings(input_ids, position_ids, token_type_ids, inputs_embeds, 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.num_hidden_layers encoder_outputs = self.encoder( embedding_output, extended_attention_mask, head_mask, output_attentions, output_hidden_states, return_dict, training=training, ) sequence_output = encoder_outputs[0] pooled_output = self.pooler(sequence_output) if self.pooler is not None else None if not return_dict: return ( sequence_output, pooled_output, ) + encoder_outputs[1:] return TFBaseModelOutputWithPooling( last_hidden_state=sequence_output, pooler_output=pooled_output, hidden_states=encoder_outputs.hidden_states, attentions=encoder_outputs.attentions, ) def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "embeddings", None) is not None: with tf.name_scope(self.embeddings.name): self.embeddings.build(None) if getattr(self, "encoder", None) is not None: with tf.name_scope(self.encoder.name): self.encoder.build(None) if getattr(self, "pooler", None) is not None: with tf.name_scope(self.pooler.name): self.pooler.build(None) class TFMobileBertPreTrainedModel(TFPreTrainedModel): """ An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models. """ config_class = MobileBertConfig base_model_prefix = "mobilebert" @dataclass class TFMobileBertForPreTrainingOutput(ModelOutput): """ Output type of [`TFMobileBertForPreTraining`]. Args: prediction_logits (`tf.Tensor` of shape `(batch_size, sequence_length, config.vocab_size)`): Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax). seq_relationship_logits (`tf.Tensor` of shape `(batch_size, 2)`): Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation 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. """ loss: tf.Tensor | None = None prediction_logits: Optional[tf.Tensor] = None seq_relationship_logits: Optional[tf.Tensor] = None hidden_states: Tuple[tf.Tensor] | None = None attentions: Tuple[tf.Tensor] | None = None MOBILEBERT_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 [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! Parameters: config ([`MobileBertConfig`]): 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. """ MOBILEBERT_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 [`AutoTokenizer`]. See [`PreTrainedTokenizer.__call__`] and [`PreTrainedTokenizer.encode`] for details. [What are input IDs?](../glossary#input-ids) attention_mask (`Numpy array` or `tf.Tensor` of shape `({0})`, *optional*): Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. [What are attention masks?](../glossary#attention-mask) token_type_ids (`Numpy array` or `tf.Tensor` of shape `({0})`, *optional*): Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0, 1]`: - 0 corresponds to a *sentence A* token, - 1 corresponds to a *sentence B* token. [What are token type IDs?](../glossary#token-type-ids) position_ids (`Numpy array` or `tf.Tensor` of shape `({0})`, *optional*): Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0, config.max_position_embeddings - 1]`. [What are position IDs?](../glossary#position-ids) head_mask (`Numpy array` or `tf.Tensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*): Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`: - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. inputs_embeds (`tf.Tensor` of shape `({0}, hidden_size)`, *optional*): Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert `input_ids` indices into associated vectors than the model's internal embedding lookup matrix. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. This argument can be used only in eager mode, in graph mode the value in the config will be used instead. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. This argument can be used only in eager mode, in graph mode the value in the config will be used instead. return_dict (`bool`, *optional*): Whether or not to return a [`~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 MobileBert Model transformer outputting raw hidden-states without any specific head on top.", MOBILEBERT_START_DOCSTRING, ) class TFMobileBertModel(TFMobileBertPreTrainedModel): def __init__(self, config, *inputs, **kwargs): super().__init__(config, *inputs, **kwargs) self.mobilebert = TFMobileBertMainLayer(config, name="mobilebert") @unpack_inputs @add_start_docstrings_to_model_forward(MOBILEBERT_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, position_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[Tuple, TFBaseModelOutputWithPooling]: outputs = self.mobilebert( 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, ) return outputs def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "mobilebert", None) is not None: with tf.name_scope(self.mobilebert.name): self.mobilebert.build(None) @add_start_docstrings( """ MobileBert Model with two heads on top as done during the pretraining: a `masked language modeling` head and a `next sentence prediction (classification)` head. """, MOBILEBERT_START_DOCSTRING, ) class TFMobileBertForPreTraining(TFMobileBertPreTrainedModel, TFMobileBertPreTrainingLoss): def __init__(self, config, *inputs, **kwargs): super().__init__(config, *inputs, **kwargs) self.mobilebert = TFMobileBertMainLayer(config, name="mobilebert") self.predictions = TFMobileBertMLMHead(config, name="predictions___cls") self.seq_relationship = TFMobileBertOnlyNSPHead(config, name="seq_relationship___cls") def get_lm_head(self): return self.predictions.predictions 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.predictions.name + "/" + self.predictions.predictions.name @unpack_inputs @add_start_docstrings_to_model_forward(MOBILEBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length")) @replace_return_docstrings(output_type=TFMobileBertForPreTrainingOutput, 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, position_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, next_sentence_label: np.ndarray | tf.Tensor | None = None, training: Optional[bool] = False, ) -> Union[Tuple, TFMobileBertForPreTrainingOutput]: r""" Return: Examples: ```python >>> import tensorflow as tf >>> from transformers import AutoTokenizer, TFMobileBertForPreTraining >>> tokenizer = AutoTokenizer.from_pretrained("google/mobilebert-uncased") >>> model = TFMobileBertForPreTraining.from_pretrained("google/mobilebert-uncased") >>> input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1 >>> outputs = model(input_ids) >>> prediction_scores, seq_relationship_scores = outputs[:2] ```""" outputs = self.mobilebert( 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, ) sequence_output, pooled_output = outputs[:2] prediction_scores = self.predictions(sequence_output) seq_relationship_score = self.seq_relationship(pooled_output) total_loss = None if labels is not None and next_sentence_label is not None: d_labels = {"labels": labels} d_labels["next_sentence_label"] = next_sentence_label total_loss = self.hf_compute_loss(labels=d_labels, logits=(prediction_scores, seq_relationship_score)) if not return_dict: output = (prediction_scores, seq_relationship_score) + outputs[2:] return ((total_loss,) + output) if total_loss is not None else output return TFMobileBertForPreTrainingOutput( loss=total_loss, prediction_logits=prediction_scores, seq_relationship_logits=seq_relationship_score, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "mobilebert", None) is not None: with tf.name_scope(self.mobilebert.name): self.mobilebert.build(None) if getattr(self, "predictions", None) is not None: with tf.name_scope(self.predictions.name): self.predictions.build(None) if getattr(self, "seq_relationship", None) is not None: with tf.name_scope(self.seq_relationship.name): self.seq_relationship.build(None) def tf_to_pt_weight_rename(self, tf_weight): if tf_weight == "cls.predictions.decoder.weight": return tf_weight, "mobilebert.embeddings.word_embeddings.weight" else: return (tf_weight,) @add_start_docstrings("""MobileBert Model with a `language modeling` head on top.""", MOBILEBERT_START_DOCSTRING) class TFMobileBertForMaskedLM(TFMobileBertPreTrainedModel, TFMaskedLanguageModelingLoss): # names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model _keys_to_ignore_on_load_unexpected = [ r"pooler", r"seq_relationship___cls", r"cls.seq_relationship", ] def __init__(self, config, *inputs, **kwargs): super().__init__(config, *inputs, **kwargs) self.mobilebert = TFMobileBertMainLayer(config, add_pooling_layer=False, name="mobilebert") self.predictions = TFMobileBertMLMHead(config, name="predictions___cls") def get_lm_head(self): return self.predictions.predictions 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.mlm.name + "/" + self.mlm.predictions.name @unpack_inputs @add_start_docstrings_to_model_forward(MOBILEBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length")) @add_code_sample_docstrings( checkpoint=_CHECKPOINT_FOR_DOC, output_type=TFMaskedLMOutput, config_class=_CONFIG_FOR_DOC, expected_output="'paris'", expected_loss=0.57, ) 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, position_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[Tuple, TFMaskedLMOutput]: 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 """ outputs = self.mobilebert( 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, ) sequence_output = outputs[0] prediction_scores = self.predictions(sequence_output, training=training) loss = None if labels is None else self.hf_compute_loss(labels, 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, ) def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "mobilebert", None) is not None: with tf.name_scope(self.mobilebert.name): self.mobilebert.build(None) if getattr(self, "predictions", None) is not None: with tf.name_scope(self.predictions.name): self.predictions.build(None) def tf_to_pt_weight_rename(self, tf_weight): if tf_weight == "cls.predictions.decoder.weight": return tf_weight, "mobilebert.embeddings.word_embeddings.weight" else: return (tf_weight,) class TFMobileBertOnlyNSPHead(keras.layers.Layer): def __init__(self, config, **kwargs): super().__init__(**kwargs) self.seq_relationship = keras.layers.Dense(2, name="seq_relationship") self.config = config def call(self, pooled_output): seq_relationship_score = self.seq_relationship(pooled_output) return seq_relationship_score def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "seq_relationship", None) is not None: with tf.name_scope(self.seq_relationship.name): self.seq_relationship.build([None, None, self.config.hidden_size]) @add_start_docstrings( """MobileBert Model with a `next sentence prediction (classification)` head on top.""", MOBILEBERT_START_DOCSTRING, ) class TFMobileBertForNextSentencePrediction(TFMobileBertPreTrainedModel, TFNextSentencePredictionLoss): # names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model _keys_to_ignore_on_load_unexpected = [r"predictions___cls", r"cls.predictions"] def __init__(self, config, *inputs, **kwargs): super().__init__(config, *inputs, **kwargs) self.mobilebert = TFMobileBertMainLayer(config, name="mobilebert") self.cls = TFMobileBertOnlyNSPHead(config, name="seq_relationship___cls") @unpack_inputs @add_start_docstrings_to_model_forward(MOBILEBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length")) @replace_return_docstrings(output_type=TFNextSentencePredictorOutput, 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, position_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, next_sentence_label: np.ndarray | tf.Tensor | None = None, training: Optional[bool] = False, ) -> Union[Tuple, TFNextSentencePredictorOutput]: r""" Return: Examples: ```python >>> import tensorflow as tf >>> from transformers import AutoTokenizer, TFMobileBertForNextSentencePrediction >>> tokenizer = AutoTokenizer.from_pretrained("google/mobilebert-uncased") >>> model = TFMobileBertForNextSentencePrediction.from_pretrained("google/mobilebert-uncased") >>> prompt = "In Italy, pizza served in formal settings, such as at a restaurant, is presented unsliced." >>> next_sentence = "The sky is blue due to the shorter wavelength of blue light." >>> encoding = tokenizer(prompt, next_sentence, return_tensors="tf") >>> logits = model(encoding["input_ids"], token_type_ids=encoding["token_type_ids"])[0] ```""" outputs = self.mobilebert( 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, ) pooled_output = outputs[1] seq_relationship_scores = self.cls(pooled_output) next_sentence_loss = ( None if next_sentence_label is None else self.hf_compute_loss(labels=next_sentence_label, logits=seq_relationship_scores) ) if not return_dict: output = (seq_relationship_scores,) + outputs[2:] return ((next_sentence_loss,) + output) if next_sentence_loss is not None else output return TFNextSentencePredictorOutput( loss=next_sentence_loss, logits=seq_relationship_scores, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "mobilebert", None) is not None: with tf.name_scope(self.mobilebert.name): self.mobilebert.build(None) if getattr(self, "cls", None) is not None: with tf.name_scope(self.cls.name): self.cls.build(None) @add_start_docstrings( """ MobileBert Model transformer with a sequence classification/regression head on top (a linear layer on top of the pooled output) e.g. for GLUE tasks. """, MOBILEBERT_START_DOCSTRING, ) class TFMobileBertForSequenceClassification(TFMobileBertPreTrainedModel, TFSequenceClassificationLoss): # names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model _keys_to_ignore_on_load_unexpected = [ r"predictions___cls", r"seq_relationship___cls", r"cls.predictions", r"cls.seq_relationship", ] _keys_to_ignore_on_load_missing = [r"dropout"] def __init__(self, config, *inputs, **kwargs): super().__init__(config, *inputs, **kwargs) self.num_labels = config.num_labels self.mobilebert = TFMobileBertMainLayer(config, name="mobilebert") classifier_dropout = ( config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob ) self.dropout = keras.layers.Dropout(classifier_dropout) self.classifier = keras.layers.Dense( config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="classifier" ) self.config = config @unpack_inputs @add_start_docstrings_to_model_forward(MOBILEBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length")) @add_code_sample_docstrings( checkpoint=_CHECKPOINT_FOR_SEQUENCE_CLASSIFICATION, output_type=TFSequenceClassifierOutput, config_class=_CONFIG_FOR_DOC, expected_output=_SEQ_CLASS_EXPECTED_OUTPUT, expected_loss=_SEQ_CLASS_EXPECTED_LOSS, ) 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, position_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[Tuple, TFSequenceClassifierOutput]: 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). """ outputs = self.mobilebert( 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, ) pooled_output = outputs[1] pooled_output = self.dropout(pooled_output, training=training) logits = self.classifier(pooled_output) loss = None if labels is None else self.hf_compute_loss(labels, logits) if not return_dict: output = (logits,) + outputs[2:] return ((loss,) + output) if loss is not None else output return TFSequenceClassifierOutput( loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "mobilebert", None) is not None: with tf.name_scope(self.mobilebert.name): self.mobilebert.build(None) if getattr(self, "classifier", None) is not None: with tf.name_scope(self.classifier.name): self.classifier.build([None, None, self.config.hidden_size]) @add_start_docstrings( """ MobileBert 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`). """, MOBILEBERT_START_DOCSTRING, ) class TFMobileBertForQuestionAnswering(TFMobileBertPreTrainedModel, TFQuestionAnsweringLoss): # names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model _keys_to_ignore_on_load_unexpected = [ r"pooler", r"predictions___cls", r"seq_relationship___cls", r"cls.predictions", r"cls.seq_relationship", ] def __init__(self, config, *inputs, **kwargs): super().__init__(config, *inputs, **kwargs) self.num_labels = config.num_labels self.mobilebert = TFMobileBertMainLayer(config, add_pooling_layer=False, name="mobilebert") self.qa_outputs = keras.layers.Dense( config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="qa_outputs" ) self.config = config @unpack_inputs @add_start_docstrings_to_model_forward(MOBILEBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length")) @add_code_sample_docstrings( checkpoint=_CHECKPOINT_FOR_QA, output_type=TFQuestionAnsweringModelOutput, config_class=_CONFIG_FOR_DOC, qa_target_start_index=_QA_TARGET_START_INDEX, qa_target_end_index=_QA_TARGET_END_INDEX, expected_output=_QA_EXPECTED_OUTPUT, expected_loss=_QA_EXPECTED_LOSS, ) 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, position_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[Tuple, TFQuestionAnsweringModelOutput]: 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. """ outputs = self.mobilebert( 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, ) sequence_output = outputs[0] logits = self.qa_outputs(sequence_output) start_logits, end_logits = tf.split(logits, 2, axis=-1) start_logits = tf.squeeze(start_logits, axis=-1) end_logits = tf.squeeze(end_logits, axis=-1) loss = None if 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, (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, ) def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "mobilebert", None) is not None: with tf.name_scope(self.mobilebert.name): self.mobilebert.build(None) if getattr(self, "qa_outputs", None) is not None: with tf.name_scope(self.qa_outputs.name): self.qa_outputs.build([None, None, self.config.hidden_size]) @add_start_docstrings( """ MobileBert 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. """, MOBILEBERT_START_DOCSTRING, ) class TFMobileBertForMultipleChoice(TFMobileBertPreTrainedModel, TFMultipleChoiceLoss): # names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model _keys_to_ignore_on_load_unexpected = [ r"predictions___cls", r"seq_relationship___cls", r"cls.predictions", r"cls.seq_relationship", ] _keys_to_ignore_on_load_missing = [r"dropout"] def __init__(self, config, *inputs, **kwargs): super().__init__(config, *inputs, **kwargs) self.mobilebert = TFMobileBertMainLayer(config, name="mobilebert") self.dropout = keras.layers.Dropout(config.hidden_dropout_prob) self.classifier = keras.layers.Dense( 1, kernel_initializer=get_initializer(config.initializer_range), name="classifier" ) self.config = config @unpack_inputs @add_start_docstrings_to_model_forward( MOBILEBERT_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, position_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[Tuple, TFMultipleChoiceModelOutput]: 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) """ 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(input_ids, (-1, seq_length)) if input_ids is not None else None flat_attention_mask = tf.reshape(attention_mask, (-1, seq_length)) if attention_mask is not None else None flat_token_type_ids = tf.reshape(token_type_ids, (-1, seq_length)) if token_type_ids is not None else None flat_position_ids = tf.reshape(position_ids, (-1, seq_length)) if position_ids is not None else None flat_inputs_embeds = ( tf.reshape(inputs_embeds, (-1, seq_length, shape_list(inputs_embeds)[3])) if inputs_embeds is not None else None ) outputs = self.mobilebert( flat_input_ids, flat_attention_mask, flat_token_type_ids, flat_position_ids, head_mask, flat_inputs_embeds, output_attentions, output_hidden_states, return_dict=return_dict, training=training, ) pooled_output = outputs[1] pooled_output = self.dropout(pooled_output, training=training) logits = self.classifier(pooled_output) reshaped_logits = tf.reshape(logits, (-1, num_choices)) loss = None if labels is None else self.hf_compute_loss(labels, reshaped_logits) if not return_dict: output = (reshaped_logits,) + outputs[2:] return ((loss,) + output) if loss is not None else output return TFMultipleChoiceModelOutput( loss=loss, logits=reshaped_logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "mobilebert", None) is not None: with tf.name_scope(self.mobilebert.name): self.mobilebert.build(None) if getattr(self, "classifier", None) is not None: with tf.name_scope(self.classifier.name): self.classifier.build([None, None, self.config.hidden_size]) @add_start_docstrings( """ MobileBert 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. """, MOBILEBERT_START_DOCSTRING, ) class TFMobileBertForTokenClassification(TFMobileBertPreTrainedModel, TFTokenClassificationLoss): # names with a '.' represents the authorized unexpected/missing layers when a TF model is loaded from a PT model _keys_to_ignore_on_load_unexpected = [ r"pooler", r"predictions___cls", r"seq_relationship___cls", r"cls.predictions", r"cls.seq_relationship", ] _keys_to_ignore_on_load_missing = [r"dropout"] def __init__(self, config, *inputs, **kwargs): super().__init__(config, *inputs, **kwargs) self.num_labels = config.num_labels self.mobilebert = TFMobileBertMainLayer(config, add_pooling_layer=False, name="mobilebert") classifier_dropout = ( config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob ) self.dropout = keras.layers.Dropout(classifier_dropout) self.classifier = keras.layers.Dense( config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="classifier" ) self.config = config @unpack_inputs @add_start_docstrings_to_model_forward(MOBILEBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length")) @add_code_sample_docstrings( checkpoint=_CHECKPOINT_FOR_TOKEN_CLASSIFICATION, output_type=TFTokenClassifierOutput, config_class=_CONFIG_FOR_DOC, expected_output=_TOKEN_CLASS_EXPECTED_OUTPUT, expected_loss=_TOKEN_CLASS_EXPECTED_LOSS, ) 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, position_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[Tuple, TFTokenClassifierOutput]: 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]`. """ outputs = self.mobilebert( 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, ) sequence_output = outputs[0] sequence_output = self.dropout(sequence_output, training=training) logits = self.classifier(sequence_output) loss = None if labels is None else self.hf_compute_loss(labels, logits) if not return_dict: output = (logits,) + outputs[2:] return ((loss,) + output) if loss is not None else output return TFTokenClassifierOutput( loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) def build(self, input_shape=None): if self.built: return self.built = True if getattr(self, "mobilebert", None) is not None: with tf.name_scope(self.mobilebert.name): self.mobilebert.build(None) if getattr(self, "classifier", None) is not None: with tf.name_scope(self.classifier.name): self.classifier.build([None, None, self.config.hidden_size]) __all__ = [ "TFMobileBertForMaskedLM", "TFMobileBertForMultipleChoice", "TFMobileBertForNextSentencePrediction", "TFMobileBertForPreTraining", "TFMobileBertForQuestionAnswering", "TFMobileBertForSequenceClassification", "TFMobileBertForTokenClassification", "TFMobileBertMainLayer", "TFMobileBertModel", "TFMobileBertPreTrainedModel", ]