Created by: peterzhang2029

配置如下：

def lstm_step(x_t, hidden_t_prev, cell_t_prev, size):
    def linear(inputs):
        return fluid.layers.fc(input=inputs, size=size, bias_attr=True)

    forget_gate = fluid.layers.sigmoid(x=linear([hidden_t_prev, x_t]))
    input_gate = fluid.layers.sigmoid(x=linear([hidden_t_prev, x_t]))
    output_gate = fluid.layers.sigmoid(x=linear([hidden_t_prev, x_t]))
    cell_tilde = fluid.layers.tanh(x=linear([hidden_t_prev, x_t]))

    cell_t = fluid.layers.sums(input=[
        fluid.layers.elementwise_mul(
            x=forget_gate, y=cell_t_prev), fluid.layers.elementwise_mul(
                x=input_gate, y=cell_tilde)
    ])

    hidden_t = fluid.layers.elementwise_mul(
        x=output_gate, y=fluid.layers.tanh(x=cell_t))

    return hidden_t, cell_t


def seq_to_seq_net(embedding_dim, encoder_size, decoder_size, source_dict_dim,
                   target_dict_dim, is_generating, beam_size, max_length):
    """Construct a seq2seq network."""

    def bi_lstm_encoder(input_seq, gate_size):
        # Linear transformation part for input gate, output gate, forget gate
        # and cell activation vectors need be done outside of dynamic_lstm.
        # So the output size is 4 times of gate_size.
        input_forward_proj = fluid.layers.fc(input=input_seq,
                                             size=gate_size * 4,
                                             act='tanh',
                                             bias_attr=True)
        forward, _ = fluid.layers.dynamic_lstm(
            input=input_forward_proj, size=gate_size * 4, use_peepholes=False)
        input_reversed_proj = fluid.layers.fc(input=input_seq,
                                              size=gate_size * 4,
                                              act='tanh',
                                              bias_attr=True)
        reversed, _ = fluid.layers.dynamic_lstm(
            input=input_reversed_proj,
            size=gate_size * 4,
            is_reverse=True,
            use_peepholes=False)
        return forward, reversed

    src_word_idx = fluid.layers.data(
        name='source_sequence', shape=[1], dtype='int64', lod_level=1)

    src_embedding = fluid.layers.embedding(
        input=src_word_idx,
        size=[source_dict_dim, embedding_dim],
        dtype='float32')

    src_forward, src_reversed = bi_lstm_encoder(
        input_seq=src_embedding, gate_size=encoder_size)

    encoded_vector = fluid.layers.concat(
        input=[src_forward, src_reversed], axis=1)
    '''
    encoded_proj = fluid.layers.fc(input=encoded_vector,
                                   size=decoder_size,
                                   bias_attr=False)
    '''
    context = fluid.layers.sequence_pool(
        input=encoded_vector, pool_type='last')

    backward_first = fluid.layers.sequence_pool(
        input=src_reversed, pool_type='first')
    
    decoder_boot = fluid.layers.fc(input=backward_first,
                                   size=decoder_size,
                                   bias_attr=False,
                                   act='tanh')

    def lstm_decoder_with_attention(target_embedding, decoder_boot, context, decoder_size):
        
        rnn = fluid.layers.DynamicRNN()

        cell_init = fluid.layers.fill_constant_batch_size_like(
            input=decoder_boot,
            value=0.0,
            shape=[-1, decoder_size],
            dtype='float32')
        cell_init.stop_gradient = False

        with rnn.block():
            current_word = rnn.step_input(target_embedding)
            context = rnn.static_input(context)
            
            hidden_mem = rnn.memory(init=decoder_boot, need_reorder=True)
            cell_mem = rnn.memory(init=cell_init)
            decoder_inputs = fluid.layers.concat(
                input=[context, current_word], axis=1)
            h, c = lstm_step(decoder_inputs, hidden_mem, cell_mem, decoder_size)
            rnn.update_memory(hidden_mem, h)
            rnn.update_memory(cell_mem, c)
            out = fluid.layers.fc(input=h,
                                  size=target_dict_dim,
                                  bias_attr=True,
                                  act='softmax')
            rnn.output(out)
        return rnn()

    if not is_generating:
        trg_word_idx = fluid.layers.data(
            name='target_sequence', shape=[1], dtype='int64', lod_level=1)

        trg_embedding = fluid.layers.embedding(
            input=trg_word_idx,
            size=[target_dict_dim, embedding_dim],
            dtype='float32')

        prediction = lstm_decoder_with_attention(trg_embedding,
                                                 decoder_boot,
                                                 context,
                                                 decoder_size)
        label = fluid.layers.data(
            name='label_sequence', shape=[1], dtype='int64', lod_level=1)
        cost = fluid.layers.cross_entropy(input=prediction, label=label)
        avg_cost = fluid.layers.mean(x=cost)

        feeding_list = ["source_sequence", "target_sequence", "label_sequence"]

        return avg_cost, feeding_list

当上面去掉上面语句的注释时：

    encoded_proj = fluid.layers.fc(input=encoded_vector,
                                   size=decoder_size,
                                   bias_attr=False)

会报错：

Traceback (most recent call last):
  File "nmt.py", line 316, in <module>
    train()
  File "nmt.py", line 239, in train
    optimizer.minimize(avg_cost)
  File "/home/zhangchao/.jumbo/lib/python2.7/site-packages/paddle/v2/fluid/optimizer.py", line 214, in minimize
    error_clip_callback)
  File "/home/zhangchao/.jumbo/lib/python2.7/site-packages/paddle/v2/fluid/backward.py", line 401, in append_backward
    raise ValueError("param %s is not in map" % param)
ValueError: param fc_2.w_0 is not in map

encoded_proj 变量后面并没有被用到，如果将这句注释掉，程序是可以正确运行的。如果出现了这类悬挂变量，程序也应该是正常运行才对吧。