def __init__(self,
vocab_size=33708,
embedding_width=512,
dropout_rate=0.0,
padded_decode=False,
decode_max_length=None,
extra_decode_length=0,
beam_size=4,
alpha=0.6,
encoder_layer=None,
decoder_layer=None,
dtype=tf.float32,
eos_id=EOS_ID,
**kwargs):
"""Initialize layers to build Transformer model.
Args:
vocab_size: Size of vocabulary.
embedding_width: Size of hidden layer for embedding.
dropout_rate: Dropout probability.
padded_decode: Whether to max_sequence_length padding is used. If set
False, max_sequence_length padding is not used.
decode_max_length: maximum number of steps to decode a sequence.
extra_decode_length: Beam search will run extra steps to decode.
beam_size: Number of beams for beam search
alpha: The strength of length normalization for beam search.
encoder_layer: An initialized encoder layer.
decoder_layer: An initialized decoder layer.
dtype: float dtype.
eos_id: Id of end of sentence token.
**kwargs: other keyword arguments.
"""
super(Seq2SeqTransformer, self).__init__(**kwargs)
self._vocab_size = vocab_size
self._embedding_width = embedding_width
self._dropout_rate = dropout_rate
self._padded_decode = padded_decode
self._decode_max_length = decode_max_length
self._extra_decode_length = extra_decode_length
self._beam_size = beam_size
self._alpha = alpha
self._dtype = dtype
self._eos_id = eos_id
self.embedding_lookup = keras_nlp.layers.OnDeviceEmbedding(
vocab_size=self._vocab_size,
embedding_width=self._embedding_width,
initializer=tf.random_normal_initializer(
mean=0., stddev=self._embedding_width**-0.5),
scale_factor=self._embedding_width**0.5)
self.encoder_layer = encoder_layer
self.decoder_layer = decoder_layer
self.position_embedding = layers.RelativePositionEmbedding(
hidden_size=self._embedding_width)
self.encoder_dropout = tf.keras.layers.Dropout(rate=self._dropout_rate)
self.decoder_dropout = tf.keras.layers.Dropout(rate=self._dropout_rate)
def __init__(self, hparams):
super(ScreenCaptionModel, self).__init__()
self._hparams = hparams
with tf.name_scope('captioning'):
self._word_embedding_layer = EmbeddingLayer(
name='word',
hidden_dim=self._hparams['hidden_size'],
embedding_file=self._hparams['embedding_file'],
vocab_size=self._hparams['vocab_size'],
embedding_dim=self._hparams['hidden_size'], # not used
trainable=self._hparams['glove_trainable'])
self._position_embedding_layer = layers.RelativePositionEmbedding(
self._hparams['hidden_size'])
self._encoder = EncoderLayer(self._hparams,
self._word_embedding_layer)
self._decoder = DecoderLayer(self._hparams,
self._word_embedding_layer,
self._position_embedding_layer)
self._word_layer = tf.keras.layers.Dense(
units=self._hparams['vocab_size'])
self.model_metrics = {
'loss': tf.keras.metrics.Mean(name='loss'),
'caption_loss': tf.keras.metrics.Mean(name='caption_loss'),
'global_norm': tf.keras.metrics.Mean(name='global_norm'),
}
self.caption_metrics = {}
for score_name in self._SCORE_NAMES:
scoped_name = 'COCO/{}'.format(score_name)
self.caption_metrics[scoped_name] = tf.keras.metrics.Mean(
name=scoped_name)
self._word_vocab = []
with tf.io.gfile.Open(self._hparams['word_vocab_path']) as f:
for index, line in enumerate(f):
if index >= self._hparams['vocab_size']:
break
self._word_vocab.append(line.strip())
def __init__(self,
vocab_size=33708,
hidden_size=512,
dropout_rate=0.0,
padded_decode=False,
num_replicas=1,
decode_batch_size=2048,
decode_max_length=97,
dtype=tf.float32,
extra_decode_length=0,
num_heads=8,
num_layers=6,
beam_size=4,
alpha=0.6,
encoder_layer=None,
decoder_layer=None,
name=None,
**kwargs):
"""Initialize layers to build Transformer model.
Arguments:
vocab_size: Size of vocabulary.
hidden_size: Size of hidden layer for embedding.
dropout_rate: Dropout probability.
padded_decode: Whether to max_sequence_length padding is used. If set
False, max_sequence_length padding is not used.
num_replicas: Number of replicas for distribution strategy.
decode_batch_size: batch_size for decoding.
decode_max_length: maximum number of steps to decode a sequence.
dtype: data type.
extra_decode_length: Beam search will run extra steps to decode.
num_heads: Number of attention heads.
num_layers: Number of identical layers for Transformer architecture.
beam_size: Number of beams for beam search
alpha: The strength of length normalization for beam search.
encoder_layer: An initialized encoder layer.
decoder_layer: An initialized decoder layer.
name: name of the model.
**kwargs: other keyword arguments.
"""
super(Seq2SeqTransformer, self).__init__(**kwargs)
self._vocab_size = vocab_size
self._hidden_size = hidden_size
self._dropout_rate = dropout_rate
self._padded_decode = padded_decode
self._num_replicas = num_replicas
self._decode_batch_size = decode_batch_size
self._decode_max_length = decode_max_length
self._dtype = dtype
self._extra_decode_length = extra_decode_length
self._num_heads = num_heads
self._num_layers = num_layers
self._beam_size = beam_size
self._alpha = alpha
self.embedding_lookup = layers.OnDeviceEmbedding(
vocab_size=self._vocab_size,
embedding_width=self._hidden_size,
initializer=tf.random_normal_initializer(
mean=0., stddev=self._hidden_size**-0.5),
use_scale=True)
self.encoder_layer = encoder_layer
self.decoder_layer = decoder_layer
self.position_embedding = layers.RelativePositionEmbedding(
hidden_size=self._hidden_size)
self.encoder_dropout = tf.keras.layers.Dropout(rate=self._dropout_rate)
self.decoder_dropout = tf.keras.layers.Dropout(rate=self._dropout_rate)
