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 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 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