def forward(self, src, src_length):
# encoding
encoder_output, encoder_final_state = self.encoder(src, src_length)
# decoder initial states
decoder_initial_states = [
encoder_final_state,
self.decoder.lstm_attention.cell.get_initial_states(
batch_ref=encoder_output, shape=[self.hidden_size])
]
# attention mask to avoid paying attention on padddings
src_mask = layers.sequence_mask(
src_length,
maxlen=layers.shape(src)[1],
dtype=encoder_output.dtype)
encoder_padding_mask = (src_mask - 1.0) * 1e9
encoder_padding_mask = layers.unsqueeze(encoder_padding_mask, [1])
# Tile the batch dimension with beam_size
encoder_output = BeamSearchDecoder.tile_beam_merge_with_batch(
encoder_output, self.beam_size)
encoder_padding_mask = BeamSearchDecoder.tile_beam_merge_with_batch(
encoder_padding_mask, self.beam_size)
# dynamic decoding with beam search
rs, _ = self.beam_search_decoder(
inits=decoder_initial_states,
encoder_output=encoder_output,
encoder_padding_mask=encoder_padding_mask)
return rs
def _build_decoder(self, enc_final_state, mode='train', beam_size=10):
output_layer = lambda x: layers.fc(
x,
size=self.tar_vocab_size,
num_flatten_dims=len(x.shape) - 1,
param_attr=fluid.ParamAttr(
name="output_w",
initializer=fluid.initializer.UniformInitializer(
low=-self.init_scale, high=self.init_scale)),
bias_attr=False)
dec_cell = AttentionDecoderCell(self.num_layers, self.hidden_size,
self.dropout, self.init_scale)
dec_initial_states = [
enc_final_state,
dec_cell.get_initial_states(batch_ref=self.enc_output,
shape=[self.hidden_size])
]
max_src_seq_len = layers.shape(self.src)[1]
src_mask = layers.sequence_mask(self.src_sequence_length,
maxlen=max_src_seq_len,
dtype='float32')
enc_padding_mask = (src_mask - 1.0)
if mode == 'train':
dec_output, _ = rnn(cell=dec_cell,
inputs=self.tar_emb,
initial_states=dec_initial_states,
sequence_length=None,
enc_output=self.enc_output,
enc_padding_mask=enc_padding_mask)
dec_output = output_layer(dec_output)
elif mode == 'beam_search':
output_layer = lambda x: layers.fc(
x,
size=self.tar_vocab_size,
num_flatten_dims=len(x.shape) - 1,
param_attr=fluid.ParamAttr(name="output_w"),
bias_attr=False)
beam_search_decoder = BeamSearchDecoder(
dec_cell,
self.beam_start_token,
self.beam_end_token,
beam_size,
embedding_fn=self.tar_embeder,
output_fn=output_layer)
enc_output = beam_search_decoder.tile_beam_merge_with_batch(
self.enc_output, beam_size)
enc_padding_mask = beam_search_decoder.tile_beam_merge_with_batch(
enc_padding_mask, beam_size)
outputs, _ = dynamic_decode(beam_search_decoder,
inits=dec_initial_states,
max_step_num=self.beam_max_step_num,
enc_output=enc_output,
enc_padding_mask=enc_padding_mask)
return outputs
return dec_output
def forward(self, inputs, *args):
gru_backward, encoded_vector, encoded_proj = self.encoder(inputs)
decoder_boot = self.fc(gru_backward[:, 0])
if self.beam_size:
# Tile the batch dimension with beam_size
encoded_vector = BeamSearchDecoder.tile_beam_merge_with_batch(
encoded_vector, self.beam_size)
encoded_proj = BeamSearchDecoder.tile_beam_merge_with_batch(
encoded_proj, self.beam_size)
# dynamic decoding with beam search
rs, _ = self.infer_decoder(inits=decoder_boot,
encoder_vec=encoded_vector,
encoder_proj=encoded_proj)
return rs
def _build_net(self):
self.seq_len = fluid.layers.data(name="seq_len",
shape=[1],
dtype='int64',
lod_level=0)
self.seq_len_used = fluid.layers.squeeze(self.seq_len, axes=[1])
src_mask = fluid.layers.sequence_mask(self.seq_len_used,
maxlen=self.max_seq_len,
dtype='float32')
enc_padding_mask = (src_mask - 1.0)
# Define decoder and initialize it.
dec_cell = AttentionDecoderCell(self.num_layers, self.hidden_size,
self.dropout)
dec_init_hidden = fluid.layers.fc(
input=self.feature,
size=self.hidden_size,
num_flatten_dims=1,
param_attr=fluid.ParamAttr(
name="dec_init_hidden_w",
initializer=fluid.initializer.TruncatedNormal(scale=0.02)),
bias_attr=fluid.ParamAttr(
name="dec_init_hidden_b",
initializer=fluid.initializer.Constant(0.)))
dec_initial_states = [[[
dec_init_hidden,
dec_cell.get_initial_states(batch_ref=self.feature,
shape=[self.hidden_size])
]] * self.num_layers,
dec_cell.get_initial_states(
batch_ref=self.feature,
shape=[self.hidden_size])]
tar_vocab_size = len(self._label_list)
tar_embeder = lambda x: fluid.embedding(
input=x,
size=[tar_vocab_size, self.hidden_size],
dtype='float32',
is_sparse=False,
param_attr=fluid.ParamAttr(name='target_embedding',
initializer=fluid.initializer.
UniformInitializer(low=-0.1, high=0.1)))
start_token_id = self._label_list.index(self.start_token)
end_token_id = self._label_list.index(self.end_token)
if not self.is_predict_phase:
self.dec_input = fluid.layers.data(name="dec_input",
shape=[self.max_seq_len],
dtype='int64')
tar_emb = tar_embeder(self.dec_input)
dec_output, _ = rnn(cell=dec_cell,
inputs=tar_emb,
initial_states=dec_initial_states,
sequence_length=None,
enc_output=self.token_feature,
enc_padding_mask=enc_padding_mask)
self.logits = fluid.layers.fc(
dec_output,
size=tar_vocab_size,
num_flatten_dims=len(dec_output.shape) - 1,
param_attr=fluid.ParamAttr(
name="output_w",
initializer=fluid.initializer.UniformInitializer(
low=-0.1, high=0.1)))
self.ret_infers = fluid.layers.reshape(x=fluid.layers.argmax(
self.logits, axis=2),
shape=[-1, 1])
logits = self.logits
logits = fluid.layers.softmax(logits)
return [logits]
else:
output_layer = lambda x: fluid.layers.fc(
x,
size=tar_vocab_size,
num_flatten_dims=len(x.shape) - 1,
param_attr=fluid.ParamAttr(name="output_w"))
beam_search_decoder = BeamSearchDecoder(dec_cell,
start_token_id,
end_token_id,
self.beam_size,
embedding_fn=tar_embeder,
output_fn=output_layer)
enc_output = beam_search_decoder.tile_beam_merge_with_batch(
self.token_feature, self.beam_size)
enc_padding_mask = beam_search_decoder.tile_beam_merge_with_batch(
enc_padding_mask, self.beam_size)
self.ret_infers, _ = dynamic_decode(
beam_search_decoder,
inits=dec_initial_states,
max_step_num=self.beam_max_step_num,
enc_output=enc_output,
enc_padding_mask=enc_padding_mask)
return self.ret_infers