encoder.py

# Copyright 2022 The TensorFlow Authors. 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.

"""Transformer-based BERT encoder network."""
# pylint: disable=g-classes-have-attributes

from typing import Any, Callable, Optional, Union, Tuple
from absl import logging
import tensorflow as tf

from official.nlp.modeling import layers


_Initializer = Union[str, tf.keras.initializers.Initializer]
_Activation = Union[str, Callable[..., Any]]

_approx_gelu = lambda x: tf.keras.activations.gelu(x, approximate=True)


class TokenDropBertEncoder(tf.keras.layers.Layer):
  """Bi-directional Transformer-based encoder network with token dropping.

  During pretraining, we drop unimportant tokens starting from an intermediate
  layer in the model, to make the model focus on important tokens more
  efficiently with its limited computational resources. The dropped tokens are
  later picked up by the last layer of the model, so that the model still
  produces full-length sequences. This approach reduces the pretraining cost of
  BERT by 25% while achieving better overall fine-tuning performance on standard
  downstream tasks.

  Args:
    vocab_size: The size of the token vocabulary.
    hidden_size: The size of the transformer hidden layers.
    num_layers: The number of transformer layers.
    num_attention_heads: The number of attention heads for each transformer. The
      hidden size must be divisible by the number of attention heads.
    max_sequence_length: The maximum sequence length that this encoder can
      consume. If None, max_sequence_length uses the value from sequence length.
      This determines the variable shape for positional embeddings.
    type_vocab_size: The number of types that the 'type_ids' input can take.
    inner_dim: The output dimension of the first Dense layer in a two-layer
      feedforward network for each transformer.
    inner_activation: The activation for the first Dense layer in a two-layer
      feedforward network for each transformer.
    output_dropout: Dropout probability for the post-attention and output
      dropout.
    attention_dropout: The dropout rate to use for the attention layers within
      the transformer layers.
    token_loss_init_value: The default loss value of a token, when the token is
      never masked and predicted.
    token_loss_beta: How running average factor for computing the average loss
      value of a token.
    token_keep_k: The number of tokens you want to keep in the intermediate
      layers. The rest will be dropped in those layers.
    token_allow_list: The list of token-ids that should not be droped. In the
      BERT English vocab, token-id from 1 to 998 contains special tokens such
      as [CLS], [SEP]. By default, token_allow_list contains all of these
      special tokens.
    token_deny_list: The list of token-ids that should always be droped. In the
      BERT English vocab, token-id=0 means [PAD]. By default, token_deny_list
      contains and only contains [PAD].
    initializer: The initialzer to use for all weights in this encoder.
    output_range: The sequence output range, [0, output_range), by slicing the
      target sequence of the last transformer layer. `None` means the entire
      target sequence will attend to the source sequence, which yields the full
      output.
    embedding_width: The width of the word embeddings. If the embedding width is
      not equal to hidden size, embedding parameters will be factorized into two
      matrices in the shape of ['vocab_size', 'embedding_width'] and
      ['embedding_width', 'hidden_size'] ('embedding_width' is usually much
      smaller than 'hidden_size').
    embedding_layer: An optional Layer instance which will be called to generate
      embeddings for the input word IDs.
    norm_first: Whether to normalize inputs to attention and intermediate dense
      layers. If set False, output of attention and intermediate dense layers is
      normalized.
    with_dense_inputs: Whether to accept dense embeddings as the input.
  """

  def __init__(
      self,
      vocab_size: int,
      hidden_size: int = 768,
      num_layers: int = 12,
      num_attention_heads: int = 12,
      max_sequence_length: int = 512,
      type_vocab_size: int = 16,
      inner_dim: int = 3072,
      inner_activation: _Activation = _approx_gelu,
      output_dropout: float = 0.1,
      attention_dropout: float = 0.1,
      token_loss_init_value: float = 10.0,
      token_loss_beta: float = 0.995,
      token_keep_k: int = 256,
      token_allow_list: Tuple[int, ...] = (100, 101, 102, 103),
      token_deny_list: Tuple[int, ...] = (0,),
      initializer: _Initializer = tf.keras.initializers.TruncatedNormal(
          stddev=0.02),
      output_range: Optional[int] = None,
      embedding_width: Optional[int] = None,
      embedding_layer: Optional[tf.keras.layers.Layer] = None,
      norm_first: bool = False,
      with_dense_inputs: bool = False,
      **kwargs):
    # Pops kwargs that are used in V1 implementation.
    if 'dict_outputs' in kwargs:
      kwargs.pop('dict_outputs')
    if 'return_all_encoder_outputs' in kwargs:
      kwargs.pop('return_all_encoder_outputs')
    if 'intermediate_size' in kwargs:
      inner_dim = kwargs.pop('intermediate_size')
    if 'activation' in kwargs:
      inner_activation = kwargs.pop('activation')
    if 'dropout_rate' in kwargs:
      output_dropout = kwargs.pop('dropout_rate')
    if 'attention_dropout_rate' in kwargs:
      attention_dropout = kwargs.pop('attention_dropout_rate')
    super().__init__(**kwargs)

    activation = tf.keras.activations.get(inner_activation)
    initializer = tf.keras.initializers.get(initializer)

    if embedding_width is None:
      embedding_width = hidden_size

    if embedding_layer is None:
      self._embedding_layer = layers.OnDeviceEmbedding(
          vocab_size=vocab_size,
          embedding_width=embedding_width,
          initializer=initializer,
          name='word_embeddings')
    else:
      self._embedding_layer = embedding_layer

    self._position_embedding_layer = layers.PositionEmbedding(
        initializer=initializer,
        max_length=max_sequence_length,
        name='position_embedding')

    self._type_embedding_layer = layers.OnDeviceEmbedding(
        vocab_size=type_vocab_size,
        embedding_width=embedding_width,
        initializer=initializer,
        use_one_hot=True,
        name='type_embeddings')

    self._embedding_norm_layer = tf.keras.layers.LayerNormalization(
        name='embeddings/layer_norm', axis=-1, epsilon=1e-12, dtype=tf.float32)

    self._embedding_dropout = tf.keras.layers.Dropout(
        rate=output_dropout, name='embedding_dropout')

    # We project the 'embedding' output to 'hidden_size' if it is not already
    # 'hidden_size'.
    self._embedding_projection = None
    if embedding_width != hidden_size:
      self._embedding_projection = tf.keras.layers.experimental.EinsumDense(
          '...x,xy->...y',
          output_shape=hidden_size,
          bias_axes='y',
          kernel_initializer=initializer,
          name='embedding_projection')

    # The first 999 tokens are special tokens such as [PAD], [CLS], [SEP].
    # We want to always mask [PAD], and always not to maks [CLS], [SEP].
    init_importance = tf.constant(token_loss_init_value, shape=(vocab_size))
    if token_allow_list:
      init_importance = tf.tensor_scatter_nd_update(
          tensor=init_importance,
          indices=[[x] for x in token_allow_list],
          updates=[1.0e4 for x in token_allow_list])
    if token_deny_list:
      init_importance = tf.tensor_scatter_nd_update(
          tensor=init_importance,
          indices=[[x] for x in token_deny_list],
          updates=[-1.0e4 for x in token_deny_list])
    self._token_importance_embed = layers.TokenImportanceWithMovingAvg(
        vocab_size=vocab_size,
        init_importance=init_importance,
        moving_average_beta=token_loss_beta)

    self._token_separator = layers.SelectTopK(top_k=token_keep_k)
    self._transformer_layers = []
    self._num_layers = num_layers
    self._attention_mask_layer = layers.SelfAttentionMask(
        name='self_attention_mask')
    for i in range(num_layers):
      layer = layers.TransformerEncoderBlock(
          num_attention_heads=num_attention_heads,
          inner_dim=inner_dim,
          inner_activation=inner_activation,
          output_dropout=output_dropout,
          attention_dropout=attention_dropout,
          norm_first=norm_first,
          output_range=output_range if i == num_layers - 1 else None,
          kernel_initializer=initializer,
          name='transformer/layer_%d' % i)
      self._transformer_layers.append(layer)

    self._pooler_layer = tf.keras.layers.Dense(
        units=hidden_size,
        activation='tanh',
        kernel_initializer=initializer,
        name='pooler_transform')

    self._config = {
        'vocab_size': vocab_size,
        'hidden_size': hidden_size,
        'num_layers': num_layers,
        'num_attention_heads': num_attention_heads,
        'max_sequence_length': max_sequence_length,
        'type_vocab_size': type_vocab_size,
        'inner_dim': inner_dim,
        'inner_activation': tf.keras.activations.serialize(activation),
        'output_dropout': output_dropout,
        'attention_dropout': attention_dropout,
        'token_loss_init_value': token_loss_init_value,
        'token_loss_beta': token_loss_beta,
        'token_keep_k': token_keep_k,
        'token_allow_list': token_allow_list,
        'token_deny_list': token_deny_list,
        'initializer': tf.keras.initializers.serialize(initializer),
        'output_range': output_range,
        'embedding_width': embedding_width,
        'embedding_layer': embedding_layer,
        'norm_first': norm_first,
        'with_dense_inputs': with_dense_inputs,
    }
    if with_dense_inputs:
      self.inputs = dict(
          input_word_ids=tf.keras.Input(shape=(None,), dtype=tf.int32),
          input_mask=tf.keras.Input(shape=(None,), dtype=tf.int32),
          input_type_ids=tf.keras.Input(shape=(None,), dtype=tf.int32),
          dense_inputs=tf.keras.Input(
              shape=(None, embedding_width), dtype=tf.float32),
          dense_mask=tf.keras.Input(shape=(None,), dtype=tf.int32),
          dense_type_ids=tf.keras.Input(shape=(None,), dtype=tf.int32),
      )
    else:
      self.inputs = dict(
          input_word_ids=tf.keras.Input(shape=(None,), dtype=tf.int32),
          input_mask=tf.keras.Input(shape=(None,), dtype=tf.int32),
          input_type_ids=tf.keras.Input(shape=(None,), dtype=tf.int32))

  def call(self, inputs):
    if isinstance(inputs, dict):
      word_ids = inputs.get('input_word_ids')
      mask = inputs.get('input_mask')
      type_ids = inputs.get('input_type_ids')

      dense_inputs = inputs.get('dense_inputs', None)
      dense_mask = inputs.get('dense_mask', None)
      dense_type_ids = inputs.get('dense_type_ids', None)
    else:
      raise ValueError('Unexpected inputs type to %s.' % self.__class__)

    word_embeddings = self._embedding_layer(word_ids)

    if dense_inputs is not None:
      # Concat the dense embeddings at sequence end.
      word_embeddings = tf.concat([word_embeddings, dense_inputs], axis=1)
      type_ids = tf.concat([type_ids, dense_type_ids], axis=1)
      mask = tf.concat([mask, dense_mask], axis=1)

    # absolute position embeddings.
    position_embeddings = self._position_embedding_layer(word_embeddings)
    type_embeddings = self._type_embedding_layer(type_ids)

    embeddings = word_embeddings + position_embeddings + type_embeddings
    embeddings = self._embedding_norm_layer(embeddings)
    embeddings = self._embedding_dropout(embeddings)

    if self._embedding_projection is not None:
      embeddings = self._embedding_projection(embeddings)

    attention_mask = self._attention_mask_layer(embeddings, mask)

    encoder_outputs = []
    x = embeddings

    # Get token routing.
    token_importance = self._token_importance_embed(word_ids)
    selected, not_selected = self._token_separator(token_importance)

    # For a 12-layer BERT:
    #   1. All tokens fist go though 5 transformer layers, then
    #   2. Only important tokens go through 1 transformer layer with cross
    #      attention to unimportant tokens, then
    #   3. Only important tokens go through 5 transformer layers without cross
    #      attention.
    #   4. Finally, all tokens go through the last layer.

    # Step 1.
    for layer in self._transformer_layers[:self._num_layers // 2 - 1]:
      x = layer([x, attention_mask])
      encoder_outputs.append(x)

    # Step 2.
    # First, separate important and non-important tokens.
    x_selected = tf.gather(x, selected, batch_dims=1, axis=1)
    mask_selected = tf.gather(mask, selected, batch_dims=1, axis=1)
    attention_mask_token_drop = self._attention_mask_layer(
        x_selected, mask_selected)

    x_not_selected = tf.gather(x, not_selected, batch_dims=1, axis=1)
    mask_not_selected = tf.gather(mask, not_selected, batch_dims=1, axis=1)
    attention_mask_token_pass = self._attention_mask_layer(
        x_selected, tf.concat([mask_selected, mask_not_selected], axis=1))
    x_all = tf.concat([x_selected, x_not_selected], axis=1)

    # Then, call transformer layer with cross attention.
    x_selected = self._transformer_layers[self._num_layers // 2 - 1](
        [x_selected, x_all, attention_mask_token_pass])
    encoder_outputs.append(x_selected)

    # Step 3.
    for layer in self._transformer_layers[self._num_layers // 2:-1]:
      x_selected = layer([x_selected, attention_mask_token_drop])
      encoder_outputs.append(x_selected)

    # Step 4.
    # First, merge important and non-important tokens.
    x_not_selected = tf.cast(x_not_selected, dtype=x_selected.dtype)
    x = tf.concat([x_selected, x_not_selected], axis=1)
    indices = tf.concat([selected, not_selected], axis=1)
    reverse_indices = tf.argsort(indices)
    x = tf.gather(x, reverse_indices, batch_dims=1, axis=1)

    # Then, call transformer layer with all tokens.
    x = self._transformer_layers[-1]([x, attention_mask])
    encoder_outputs.append(x)

    last_encoder_output = encoder_outputs[-1]
    first_token_tensor = last_encoder_output[:, 0, :]
    pooled_output = self._pooler_layer(first_token_tensor)

    return dict(
        sequence_output=encoder_outputs[-1],
        pooled_output=pooled_output,
        encoder_outputs=encoder_outputs)

  def record_mlm_loss(self, mlm_ids: tf.Tensor, mlm_losses: tf.Tensor):
    self._token_importance_embed.update_token_importance(
        token_ids=mlm_ids, importance=mlm_losses)

  def get_embedding_table(self):
    return self._embedding_layer.embeddings

  def get_embedding_layer(self):
    return self._embedding_layer

  def get_config(self):
    return dict(self._config)

  @property
  def transformer_layers(self):
    """List of Transformer layers in the encoder."""
    return self._transformer_layers

  @property
  def pooler_layer(self):
    """The pooler dense layer after the transformer layers."""
    return self._pooler_layer

  @classmethod
  def from_config(cls, config, custom_objects=None):
    if 'embedding_layer' in config and config['embedding_layer'] is not None:
      warn_string = (
          'You are reloading a model that was saved with a '
          'potentially-shared embedding layer object. If you contine to '
          'train this model, the embedding layer will no longer be shared. '
          'To work around this, load the model outside of the Keras API.')
      print('WARNING: ' + warn_string)
      logging.warn(warn_string)

    return cls(**config)