def _decode(self, input_dict):
if 'target_tensors' in input_dict:
targets = input_dict['target_tensors'][0]
else:
targets = None
encoder_outputs = input_dict['encoder_output']['outputs']
inputs_attention_bias = (
input_dict['encoder_output']['inputs_attention_bias'])
self.embedding_softmax_layer = (
input_dict['encoder_output']['embedding_softmax_layer'])
with tf.name_scope("decode"):
# prepare decoder layers
if len(self.layers) == 0:
for _ in range(self.params["num_hidden_layers"]):
self_attention_layer = attention_layer.SelfAttention(
self.params["hidden_size"],
self.params["num_heads"],
self.params["attention_dropout"],
self.mode == "train",
)
enc_dec_attention_layer = attention_layer.Attention(
self.params["hidden_size"],
self.params["num_heads"],
self.params["attention_dropout"],
self.mode == "train",
)
feed_forward_network = ffn_layer.FeedFowardNetwork(
self.params["hidden_size"],
self.params["filter_size"],
self.params["relu_dropout"],
self.mode == "train",
)
self.layers.append([
PrePostProcessingWrapper(self_attention_layer,
self.params,
self.mode == "train"),
PrePostProcessingWrapper(enc_dec_attention_layer,
self.params,
self.mode == "train"),
PrePostProcessingWrapper(feed_forward_network,
self.params,
self.mode == "train")
])
self.output_normalization = LayerNormalization(
self.params["hidden_size"])
if targets is None:
return self.predict(encoder_outputs, inputs_attention_bias)
else:
logits = self.decode_pass(targets, encoder_outputs,
inputs_attention_bias)
return {
"logits": logits,
"outputs": [tf.argmax(logits, axis=-1)],
"final_state": None,
"final_sequence_lengths": None
}
def _encode(self, input_dict):
if len(self.layers) == 0:
# prepare encoder graph
self.embedding_softmax_layer = embedding_layer.EmbeddingSharedWeights(
self.params["src_vocab_size"], self.params["hidden_size"],
pad_vocab_to_eight=self.params.get('pad_embeddings_2_eight', False))
for _ in range(self.params['encoder_layers']):
# Create sublayers for each layer.
self_attention_layer = attention_layer.SelfAttention(
self.params["hidden_size"], self.params["num_heads"],
self.params["attention_dropout"], self.mode == "train")
feed_forward_network = ffn_layer.FeedFowardNetwork(
self.params["hidden_size"], self.params["filter_size"],
self.params["relu_dropout"], self.mode == "train")
self.layers.append([
PrePostProcessingWrapper(self_attention_layer, self.params,
self.mode == "train"),
PrePostProcessingWrapper(feed_forward_network, self.params,
self.mode == "train")])
# Create final layer normalization layer.
self.output_normalization = LayerNormalization(self.params["hidden_size"])
# actual encoder part
with tf.name_scope("encode"):
#inputs = input_dict['src_sequence']
inputs = input_dict['source_tensors'][0]
# Prepare inputs to the layer stack by adding positional encodings and
# applying dropout.
embedded_inputs = self.embedding_softmax_layer(inputs)
inputs_padding = utils.get_padding(inputs)
inputs_attention_bias = utils.get_padding_bias(inputs)
#inputs_attention_bias = tf.cast(utils.get_padding_bias(inputs),
# dtype=self.params['dtype'])
with tf.name_scope("add_pos_encoding"):
length = tf.shape(embedded_inputs)[1]
pos_encoding = utils.get_position_encoding(
length, self.params["hidden_size"])
encoder_inputs = embedded_inputs + tf.cast(x=pos_encoding,
dtype=embedded_inputs.dtype)
if self.mode == "train":
encoder_inputs = tf.nn.dropout(
encoder_inputs, 1 - self.params["layer_postprocess_dropout"])
encoded = self._call(encoder_inputs, inputs_attention_bias,
inputs_padding)
return {'outputs': encoded,
'inputs_attention_bias': inputs_attention_bias,
'state': None,
'src_lengths': input_dict['source_tensors'][1],
'embedding_softmax_layer': self.embedding_softmax_layer,
'encoder_input': inputs}
def _encode(self, input_dict):
training = (self.mode == "train")
if len(self.layers) == 0:
# prepare encoder graph
self.embedding_softmax_layer = embedding_layer.EmbeddingSharedWeights(
self.params["src_vocab_size"],
self.params["hidden_size"],
pad_vocab_to_eight=self.params.get('pad_embeddings_2_eight',
False),
)
for _ in range(self.params['encoder_layers']):
# Create sublayers for each layer.
self_attention_layer = attention_layer.SelfAttention(
hidden_size=self.params["hidden_size"],
num_heads=self.params["num_heads"],
attention_dropout=self.params["attention_dropout"],
train=training,
regularizer=self.regularizer,
batch_size=self.batch_size,
num_feature=self.num_features)
feed_forward_network = ffn_layer.FeedFowardNetwork(
hidden_size=self.params["hidden_size"],
filter_size=self.params["filter_size"],
relu_dropout=self.params["relu_dropout"],
train=training,
#num_features=self.num_features,
#batch_size=self.batch_size,
regularizer=self.regularizer)
self.layers.append([
PrePostProcessingWrapper(self_attention_layer, self.params,
training),
PrePostProcessingWrapper(feed_forward_network, self.params,
training)
])
# final normalization layer.
print("Encoder:", self.norm_params["type"], self.mode)
if self.norm_params["type"] == "batch_norm":
self.output_normalization = Transformer_BatchNorm(
training=training, params=self.norm_params)
else:
self.output_normalization = LayerNormalization(
hidden_size=self.params["hidden_size"],
params=self.norm_params)
# actual encoder part
with tf.name_scope("encode"):
inputs, src_lengths = input_dict['source_tensors']
#inputs = input_dict['source_tensors'][0]
# Prepare inputs to the layer stack by adding positional encodings and
# applying dropout.
embedded_inputs = self.embedding_softmax_layer(inputs)
if self.params["remove_padding"]:
inputs_padding = utils.get_padding(inputs)
#inputs_padding = utils.get_padding(inputs,dtype=self._params["dtype"])
else:
inputs_padding = None
inputs_attention_bias = utils.get_padding_bias(inputs)
inputs_attention_bias = tf.transpose(inputs_attention_bias,
[0, 1, 3, 2, 4])
# inputs_attention_bias = utils.get_padding_bias(inputs, dtype=self._params["dtype"])
with tf.name_scope("add_pos_encoding"):
length = tf.shape(embedded_inputs)[1]
pos_encoding = utils.get_position_encoding(
length,
self.params["hidden_size"],
)
#encoder_inputs = embedded_inputs + tf.cast(x=pos_encoding,
# dtype=embedded_inputs.dtype)
pos_encoding = tf.cast(x=pos_encoding,
dtype=embedded_inputs.dtype)
pos_encoding_exp = pos_encoding[None, :, None, :]
encoder_inputs = embedded_inputs + pos_encoding_exp
if self.mode == "train":
encoder_inputs = tf.nn.dropout(
encoder_inputs,
keep_prob=1.0 - self.params["layer_postprocess_dropout"],
)
encoded = self._call(encoder_inputs, inputs_attention_bias,
inputs_padding)
return {
'outputs': encoded,
'inputs_attention_bias': inputs_attention_bias,
'state': None,
'src_lengths': src_lengths,
#'src_lengths': input_dict['source_tensors'][1],
'embedding_softmax_layer': self.embedding_softmax_layer,
'encoder_input': inputs
}
