def __call__(self, decoder_initial_states, encoder_output,
encoder_padding_mask, **kwargs):
output_layer = kwargs.pop("output_layer", None)
if self.decoding_strategy == "train_greedy":
# for teach-forcing MLE pre-training
helper = layers.TrainingHelper(**kwargs)
elif self.decoding_strategy == "infer_sample":
helper = layers.SampleEmbeddingHelper(**kwargs)
elif self.decoding_strategy == "infer_greedy":
helper = layers.GreedyEmbeddingHelper(**kwargs)
if self.decoding_strategy == "beam_search":
beam_size = kwargs.get("beam_size", 4)
encoder_output = layers.BeamSearchDecoder.tile_beam_merge_with_batch(
encoder_output, beam_size)
encoder_padding_mask = layers.BeamSearchDecoder.tile_beam_merge_with_batch(
encoder_padding_mask, beam_size)
decoder = layers.BeamSearchDecoder(
cell=self.decoder_cell, output_fn=output_layer, **kwargs)
else:
decoder = layers.BasicDecoder(
self.decoder_cell, helper, output_fn=output_layer)
(decoder_output, decoder_final_state,
dec_seq_lengths) = layers.dynamic_decode(
decoder,
inits=decoder_initial_states,
max_step_num=self.max_decoding_length,
encoder_output=encoder_output,
encoder_padding_mask=encoder_padding_mask,
impute_finished=False # for test coverage
if self.decoding_strategy == "beam_search" else True,
is_test=True if self.decoding_strategy == "beam_search" else False,
return_length=True)
return decoder_output, decoder_final_state, dec_seq_lengths
def _build_decoder(self, enc_final_state, mode='train', beam_size=10):
dec_cell = DecoderCell(self.num_layers, self.hidden_size, self.dropout,
self.init_scale)
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=uniform_initializer(
self.init_scale)),
bias_attr=False)
if mode == 'train':
dec_output, dec_final_state = rnn(cell=dec_cell,
inputs=self.tar_emb,
initial_states=enc_final_state)
dec_output = output_layer(dec_output)
return dec_output
elif mode == 'beam_search':
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)
outputs, _ = dynamic_decode(beam_search_decoder,
inits=enc_final_state,
max_step_num=self.beam_max_step_num)
return outputs
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 decoder(encoder_output,
encoder_output_proj,
encoder_state,
encoder_padding_mask,
trg=None,
is_train=True):
"""Decoder: GRU with Attention"""
decoder_cell = DecoderCell(hidden_size=decoder_size)
decoder_initial_states = layers.fc(encoder_state,
size=decoder_size,
act="tanh")
trg_embeder = lambda x: fluid.embedding(
input=x,
size=[target_dict_size, hidden_dim],
dtype="float32",
param_attr=fluid.ParamAttr(name="trg_emb_table"))
output_layer = lambda x: layers.fc(x,
size=target_dict_size,
num_flatten_dims=len(x.shape) - 1,
param_attr=fluid.ParamAttr(name=
"output_w"))
if is_train:
decoder_output, _ = layers.rnn(
cell=decoder_cell,
inputs=trg_embeder(trg),
initial_states=decoder_initial_states,
time_major=False,
encoder_output=encoder_output,
encoder_output_proj=encoder_output_proj,
encoder_padding_mask=encoder_padding_mask)
decoder_output = output_layer(decoder_output)
else:
encoder_output = layers.BeamSearchDecoder.tile_beam_merge_with_batch(
encoder_output, beam_size)
encoder_output_proj = layers.BeamSearchDecoder.tile_beam_merge_with_batch(
encoder_output_proj, beam_size)
encoder_padding_mask = layers.BeamSearchDecoder.tile_beam_merge_with_batch(
encoder_padding_mask, beam_size)
beam_search_decoder = layers.BeamSearchDecoder(
cell=decoder_cell,
start_token=bos_id,
end_token=eos_id,
beam_size=beam_size,
embedding_fn=trg_embeder,
output_fn=output_layer)
decoder_output, _ = layers.dynamic_decode(
decoder=beam_search_decoder,
inits=decoder_initial_states,
max_step_num=max_length,
output_time_major=False,
encoder_output=encoder_output,
encoder_output_proj=encoder_output_proj,
encoder_padding_mask=encoder_padding_mask)
return decoder_output
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
def _build_decoder(self,
z_mean=None,
z_log_var=None,
enc_output=None,
mode='train',
beam_size=10):
dec_input = layers.dropout(self.tar_emb,
dropout_prob=self.dec_dropout_in,
dropout_implementation="upscale_in_train")
# `output_layer` will be used within BeamSearchDecoder
output_layer = lambda x: layers.fc(x,
size=self.tar_vocab_size,
num_flatten_dims=len(x.shape) - 1,
name="output_w")
# `sample_output_layer` samples an id from the logits distribution instead of argmax(logits)
# it will be used within BeamSearchDecoder
sample_output_layer = lambda x: layers.unsqueeze(
fluid.one_hot(layers.unsqueeze(
layers.sampling_id(layers.softmax(
layers.squeeze(output_layer(x), [1])),
dtype='int'), [1]),
depth=self.tar_vocab_size), [1])
if mode == 'train':
latent_z = self._sampling(z_mean, z_log_var)
else:
latent_z = layers.gaussian_random_batch_size_like(
self.tar, shape=[-1, self.latent_size])
dec_first_hidden_cell = layers.fc(latent_z,
2 * self.hidden_size *
self.num_layers,
name='fc_hc')
dec_first_hidden, dec_first_cell = layers.split(
dec_first_hidden_cell, 2)
if self.num_layers > 1:
dec_first_hidden = layers.split(dec_first_hidden, self.num_layers)
dec_first_cell = layers.split(dec_first_cell, self.num_layers)
else:
dec_first_hidden = [dec_first_hidden]
dec_first_cell = [dec_first_cell]
dec_initial_states = [[h, c]
for h, c in zip(dec_first_hidden, dec_first_cell)
]
dec_cell = DecoderCell(self.num_layers, self.hidden_size, latent_z,
self.param_attr_initializer,
self.param_attr_scale, self.dec_dropout_out)
if mode == 'train':
dec_output, _ = rnn(cell=dec_cell,
inputs=dec_input,
initial_states=dec_initial_states,
sequence_length=self.tar_sequence_length)
dec_output = output_layer(dec_output)
return dec_output
elif mode == 'greedy':
start_token = 1
end_token = 2
max_length = 100
beam_search_decoder = BeamSearchDecoder(
dec_cell,
start_token,
end_token,
beam_size=1,
embedding_fn=self.tar_embeder,
output_fn=output_layer)
outputs, _ = dynamic_decode(beam_search_decoder,
inits=dec_initial_states,
max_step_num=max_length)
return outputs
elif mode == 'sampling':
start_token = 1
end_token = 2
max_length = 100
beam_search_decoder = BeamSearchDecoder(
dec_cell,
start_token,
end_token,
beam_size=1,
embedding_fn=self.tar_embeder,
output_fn=sample_output_layer)
outputs, _ = dynamic_decode(beam_search_decoder,
inits=dec_initial_states,
max_step_num=max_length)
return outputs
else:
print("mode not supprt", mode)
def decoder(encoder_output,
encoder_output_proj,
encoder_state,
encoder_padding_mask,
trg=None,
is_train=True):
# 定义 RNN 所需要的组件
decoder_cell = DecoderCell(hidden_size=decoder_size)
decoder_initial_states = layers.fc(encoder_state,
size=decoder_size,
act="tanh")
trg_embeder = lambda x: fluid.embedding(
input=x,
size=[target_dict_size, hidden_dim],
dtype="float32",
param_attr=fluid.ParamAttr(name="trg_emb_table"))
output_layer = lambda x: layers.fc(input=x,
size=target_dict_size,
num_flatten_dims=len(x.shape) - 1,
param_attr=fluid.ParamAttr(name=
"output_w"))
if is_train: # 训练
# 训练时使用 `layers.rnn` 构造由 `cell` 指定的循环神经网络
# 循环的每一步从 `inputs` 中切片产生输入，并执行 `cell.call`
# [-1,-1,512,] , [-1,512,]
decoder_output, _ = layers.rnn(
cell=decoder_cell,
inputs=trg_embeder(trg),
initial_states=decoder_initial_states,
time_major=False,
encoder_output=encoder_output,
encoder_output_proj=encoder_output_proj,
encoder_padding_mask=encoder_padding_mask)
decoder_output = layers.fc(input=decoder_output,
size=target_dict_size,
num_flatten_dims=2,
param_attr=fluid.ParamAttr(name="output_w"))
else: # 基于 beam search 的预测生成
# beam search 时需要将用到的形为 `[batch_size, ...]` 的张量扩展为 `[batch_size* beam_size, ...]`
encoder_output = layers.BeamSearchDecoder.tile_beam_merge_with_batch(
encoder_output, beam_size)
encoder_output_proj = layers.BeamSearchDecoder.tile_beam_merge_with_batch(
encoder_output_proj, beam_size)
encoder_padding_mask = layers.BeamSearchDecoder.tile_beam_merge_with_batch(
encoder_padding_mask, beam_size)
# BeamSearchDecoder 定义了单步解码的操作：`cell.call` + `beam_search_step`
beam_search_decoder = layers.BeamSearchDecoder(
cell=decoder_cell,
start_token=bos_id,
end_token=eos_id,
beam_size=beam_size,
embedding_fn=trg_embeder,
output_fn=output_layer)
# 使用 layers.dynamic_decode 动态解码
# 重复执行 `decoder.step()` 直到其返回的表示完成状态的张量中的值全部为True或解码步骤达到 `max_step_num`
decoder_output, _ = layers.dynamic_decode(
decoder=beam_search_decoder,
inits=decoder_initial_states,
max_step_num=max_length,
output_time_major=False,
encoder_output=encoder_output,
encoder_output_proj=encoder_output_proj,
encoder_padding_mask=encoder_padding_mask)
return decoder_output
