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
}
def _encode(self, input_dict):
inputs = input_dict['source_tensors'][0]
source_length = input_dict['source_tensors'][1]
with tf.variable_scope("encode"):
# prepare encoder graph
if len(self.layers) == 0:
knum_list = list(
zip(*self.params.get("conv_nchannels_kwidth")))[0]
kwidth_list = list(
zip(*self.params.get("conv_nchannels_kwidth")))[1]
with tf.variable_scope("embedding"):
self.embedding_softmax_layer = embedding_layer.EmbeddingSharedWeights(
vocab_size=self._src_vocab_size,
hidden_size=self._src_emb_size,
pad_vocab_to_eight=self._pad2eight,
init_var=0.1,
embed_scale=False,
pad_sym=self._pad_sym,
mask_paddings=True)
with tf.variable_scope("pos_embedding"):
self.position_embedding_layer = embedding_layer.EmbeddingSharedWeights(
vocab_size=self.params.get("max_input_length",
MAX_INPUT_LENGTH),
hidden_size=self._src_emb_size,
pad_vocab_to_eight=self._pad2eight,
init_var=0.1,
embed_scale=False,
pad_sym=self._pad_sym,
mask_paddings=True)
# linear projection before cnn layers
self.layers.append(
ffn_wn_layer.FeedFowardNetworkNormalized(
self._src_emb_size,
knum_list[0],
dropout=self.params["embedding_dropout_keep_prob"],
var_scope_name="linear_mapping_before_cnn_layers",
mode=self.mode,
normalization_type=self.normalization_type,
regularizer=self.regularizer,
init_var=self.init_var))
for i in range(len(knum_list)):
in_dim = knum_list[i] if i == 0 else knum_list[i - 1]
out_dim = knum_list[i]
# linear projection is needed for residual connections if
# input and output of a cnn layer do not match
if in_dim != out_dim:
linear_proj = ffn_wn_layer.FeedFowardNetworkNormalized(
in_dim,
out_dim,
var_scope_name="linear_mapping_cnn_" + str(i + 1),
dropout=1.0,
mode=self.mode,
normalization_type=self.normalization_type,
regularizer=self.regularizer,
init_var=self.init_var)
else:
linear_proj = None
conv_layer = conv_wn_layer.Conv1DNetworkNormalized(
in_dim,
out_dim,
kernel_width=kwidth_list[i],
mode=self.mode,
layer_id=i + 1,
hidden_dropout=self.params["hidden_dropout_keep_prob"],
conv_padding="SAME",
decode_padding=False,
activation=self.conv_activation,
normalization_type=self.normalization_type,
regularizer=self.regularizer,
init_var=self.init_var)
self.layers.append([linear_proj, conv_layer])
# linear projection after cnn layers
self.layers.append(
ffn_wn_layer.FeedFowardNetworkNormalized(
knum_list[-1],
self._src_emb_size,
dropout=1.0,
var_scope_name="linear_mapping_after_cnn_layers",
mode=self.mode,
normalization_type=self.normalization_type,
regularizer=self.regularizer,
init_var=self.init_var))
encoder_inputs = self.embedding_softmax_layer(inputs)
inputs_attention_bias = get_padding_bias(inputs,
res_rank=3,
pad_sym=self._pad_sym)
with tf.name_scope("add_pos_encoding"):
pos_input = tf.range(0,
tf.shape(encoder_inputs)[1],
delta=1,
dtype=tf.int32,
name='range')
pos_encoding = self.position_embedding_layer(pos_input)
encoder_inputs = encoder_inputs + tf.cast(
x=pos_encoding, dtype=encoder_inputs.dtype)
if self.mode == "train":
encoder_inputs = tf.nn.dropout(
encoder_inputs, self.params["embedding_dropout_keep_prob"])
# mask the paddings in the input given to cnn layers
inputs_padding = get_padding(inputs,
self._pad_sym,
dtype=encoder_inputs.dtype)
padding_mask = tf.expand_dims(1 - inputs_padding, 2)
encoder_inputs *= padding_mask
outputs, outputs_b, final_state = self._call(
encoder_inputs, padding_mask)
return {
'outputs': outputs,
'outputs_b': outputs_b,
'inputs_attention_bias_cs2s': inputs_attention_bias,
'state': final_state,
'src_lengths': source_length, # should it include paddings or not?
'embedding_softmax_layer': self.embedding_softmax_layer,
'encoder_input': inputs
}
def _decode(self, input_dict):
targets = input_dict['target_tensors'][0] \
if 'target_tensors' in input_dict else None
encoder_outputs = input_dict['encoder_output']['outputs']
encoder_outputs_b = input_dict['encoder_output'].get(
'outputs_b', encoder_outputs)
inputs_attention_bias = input_dict['encoder_output'].get(
'inputs_attention_bias_cs2s', None)
with tf.name_scope("decode"):
# prepare decoder layers
if len(self.layers) == 0:
knum_list = list(zip(*self.params.get("conv_nchannels_kwidth")))[0]
kwidth_list = list(zip(*self.params.get("conv_nchannels_kwidth")))[1]
normalization_type = self.params.get("normalization_type",
"weight_norm")
conv_activation = self.params.get("conv_activation", gated_linear_units)
# preparing embedding layers
with tf.variable_scope("embedding"):
if 'embedding_softmax_layer' in input_dict['encoder_output'] \
and self.params['shared_embed']:
self.embedding_softmax_layer = \
input_dict['encoder_output']['embedding_softmax_layer']
else:
self.embedding_softmax_layer = embedding_layer.EmbeddingSharedWeights(
vocab_size=self._tgt_vocab_size,
hidden_size=self._tgt_emb_size,
pad_vocab_to_eight=self._pad2eight,
init_var=0.1,
embed_scale=False,
pad_sym=self._pad_sym,
mask_paddings=True)
if self.params.get("pos_embed", True):
with tf.variable_scope("pos_embedding"):
if 'position_embedding_layer' in input_dict['encoder_output'] \
and self.params['shared_embed']:
self.position_embedding_layer = \
input_dict['encoder_output']['position_embedding_layer']
else:
self.position_embedding_layer = embedding_layer.EmbeddingSharedWeights(
vocab_size=self.params.get("max_input_length", MAX_INPUT_LENGTH),
hidden_size=self._tgt_emb_size,
pad_vocab_to_eight=self._pad2eight,
init_var=0.1,
embed_scale=False,
pad_sym=self._pad_sym,
mask_paddings=True)
else:
self.position_embedding_layer = None
# linear projection before cnn layers
self.layers.append(
ffn_wn_layer.FeedFowardNetworkNormalized(
self._tgt_emb_size,
knum_list[0],
dropout=self.params["embedding_dropout_keep_prob"],
var_scope_name="linear_mapping_before_cnn_layers",
mode=self.mode,
normalization_type=normalization_type))
for i in range(self.params['decoder_layers']):
in_dim = knum_list[i] if i == 0 else knum_list[i - 1]
out_dim = knum_list[i]
# linear projection is needed for residual connections if
# input and output of a cnn layer do not match
if in_dim != out_dim:
linear_proj = ffn_wn_layer.FeedFowardNetworkNormalized(
in_dim,
out_dim,
var_scope_name="linear_mapping_cnn_" + str(i + 1),
dropout=1.0,
mode=self.mode,
normalization_type=normalization_type)
else:
linear_proj = None
conv_layer = conv_wn_layer.Conv1DNetworkNormalized(
in_dim,
out_dim,
kernel_width=kwidth_list[i],
mode=self.mode,
layer_id=i + 1,
hidden_dropout=self.params["hidden_dropout_keep_prob"],
conv_padding="VALID",
decode_padding=True,
activation=conv_activation,
normalization_type=normalization_type)
att_layer = attention_wn_layer.AttentionLayerNormalized(
out_dim,
embed_size=self._tgt_emb_size,
layer_id=i + 1,
add_res=True,
mode=self.mode)
self.layers.append([linear_proj, conv_layer, att_layer])
# linear projection after cnn layers
self.layers.append(
ffn_wn_layer.FeedFowardNetworkNormalized(
knum_list[self.params['decoder_layers'] - 1],
self.params.get("out_emb_size", self._tgt_emb_size),
dropout=1.0,
var_scope_name="linear_mapping_after_cnn_layers",
mode=self.mode,
normalization_type=normalization_type))
if not self.params['shared_embed']:
self.layers.append(
ffn_wn_layer.FeedFowardNetworkNormalized(
self.params.get("out_emb_size", self._tgt_emb_size),
self._tgt_vocab_size,
dropout=self.params["out_dropout_keep_prob"],
var_scope_name="linear_mapping_to_vocabspace",
mode=self.mode,
normalization_type=normalization_type))
else:
# if embedding is shared,
# the shared embedding is used as the final linear projection to vocab space
self.layers.append(None)
if targets is None:
return self.predict(encoder_outputs, encoder_outputs_b,
inputs_attention_bias)
else:
logits = self.decode_pass(targets, encoder_outputs, encoder_outputs_b,
inputs_attention_bias)
return {
"logits": logits,
"outputs": [tf.argmax(logits, axis=-1)],
"final_state": None,
"final_sequence_lengths": None
}