def __init__(self, params, is_train, mode=None):
self.is_train = is_train
self.params = params
if mode is not None:
self.mode = mode
elif self.is_train:
self.mode = ModeKeys.TRAIN
else:
self.mode = ModeKeys.PREDICT
if params.shared_embedding_softmax_weights:
print("sharing embedding!!!")
self.embedding_softmax_layer = embedding_layer.EmbeddingSharedWeights(
params.vocab_size, params.hidden_size)
self.encoder_embedding_layer = self.embedding_softmax_layer
self.decoder_embedding_layer = self.embedding_softmax_layer
self.decoder_softmax_layer = self.embedding_softmax_layer
else:
print("not sharing embedding!!!")
self.encoder_embedding_layer = embedding_layer.EmbeddingWeights(
params.source_vocab_size, params.hidden_size, "source_embedding")
self.decoder_embedding_layer = embedding_layer.EmbeddingWeights(
params.target_vocab_size, params.hidden_size, "target_embedding")
self.decoder_softmax_layer = embedding_layer.EmbeddingWeights(
params.target_vocab_size, params.hidden_size, 'soft_max')
# done
self.encoder_stack = EncoderDecoder.EncoderStack(params, is_train, self.mode)
self.decoder_stack = EncoderDecoder.DecoderStack(params, is_train, self.mode)
self._initializer = tf.variance_scaling_initializer(
self.params.initializer_gain, mode="fan_avg", distribution="uniform")
def __init__(self, params, is_train, mode=None):
self.is_train = is_train
self.params = params
if mode is not None:
self.mode = mode
elif self.is_train:
self.mode = ModeKeys.TRAIN
else:
self.mode = ModeKeys.PREDICT
self.encoder_stack = EncoderDecoder.EncoderStack(params, is_train, self.mode)
self.decoder_stack = EncoderDecoder.DecoderStack(params, is_train, self.mode)
self._initializer = tf.variance_scaling_initializer(
self.params.initializer_gain, mode="fan_avg", distribution="uniform")
def __init__(self, params, is_train, mode=None):
self.is_train = is_train
self.params = params
if mode is not None:
self.mode = mode
elif self.is_train:
self.mode = ModeKeys.TRAIN
else:
self.mode = ModeKeys.PREDICT
#with tf.device('/cpu:0'):
# self.dropout_pl = tf.placeholder(dtype=tf.float32, shape=[], name="dropout_pl")
# self.params.layer_postprocess_dropout = self.dropout_pl
# self.params.attention_dropout = self.dropout_pl
# self.relu_dropout = self.dropout_pl
if params.shared_embedding_softmax_weights:
print("sharing embedding!!!")
self.embedding_softmax_layer = embedding_layer.EmbeddingSharedWeights(
params.vocab_size, params.hidden_size)
self.encoder_embedding_layer = self.embedding_softmax_layer
self.decoder_embedding_layer = self.embedding_softmax_layer
self.decoder_softmax_layer = self.embedding_softmax_layer
else:
print("not sharing embedding!!!")
self.encoder_embedding_layer = embedding_layer.EmbeddingWeights(
params.source_vocab_size, params.hidden_size, "source_embedding")
self.decoder_embedding_layer = embedding_layer.EmbeddingWeights(
params.target_vocab_size, params.hidden_size, "target_embedding")
self.decoder_softmax_layer = embedding_layer.EmbeddingWeights(
params.target_vocab_size, params.hidden_size, 'soft_max')
# done
self.encoder_stack = EncoderDecoder.EncoderStack(params, is_train, self.mode)
self.decoder_stack = EncoderDecoder.DecoderStack(params, is_train, self.mode)
self._initializer = tf.variance_scaling_initializer(
self.params.initializer_gain, mode="fan_avg", distribution="uniform")