使用 RNN 生成一系列张量

Question

我正在构建一个 rnn，我使用 tf.nn.dynamic_rnn 来生成输出和状态。代码如下（tf版本1.3）：

import tensorflow as tf    
def lstm_cell():
        return tf.contrib.rnn.DropoutWrapper(tf.contrib.rnn.BasicLSTMCell(128), output_keep_prob=0.7)

cell= tf.contrib.rnn.MultiRNNCell([lstm_cell() for _ in range(3)])

initial_state= cell.zero_state(1, tf.float32)

layer = tf.placeholder(tf.float32, [1,1,36])

outputs, state=tf.nn.dynamic_rnn(cell=cell, inputs=layer, initial_state=initial_state)

由于输入张量的批量大小始终为 1，因此 initial_state 和状态的批量大小也为 1。 layer 也是 batch_size=1 的输入，每个单元格有 36 个节点（嵌入序列的大小）。每层有lstm_size128.

循环 rnn 单元格时出现问题。

rnn_outputs_sequence=outputs
for i in range(1, num_pics, 1):
    outputs, state=tf.nn.dynamic_rnn(cell=cell, inputs=outputs, initial_state=state)
    rnn_outputs_sequence=tf.concat((rnn_outputs_sequence, outputs),axis=1)

rnn_outputs_sequence 应具有 [1, num_pics, 36] 的形状。但是，这会触发错误：

Trying to share variable rnn/multi_rnn_cell/cell_0/basic_lstm_cell/kernel, but specified shape (256, 512) and found shape (164, 512).

这个形状我想不出来[164, 512]。谁能帮我解决这个问题？谢谢

Answer 1

import tensorflow as tf
def lstm_cell():
    return tf.contrib.rnn.DropoutWrapper(tf.contrib.rnn.BasicLSTMCell(128), output_keep_prob=0.7)

cell= tf.contrib.rnn.MultiRNNCell([lstm_cell() for _ in range(2)])

initial_state= cell.zero_state(1, tf.float32)

layer = tf.placeholder(tf.float32, [1,1,36])

outputs, state=tf.nn.dynamic_rnn(cell=cell, inputs=layer, initial_state=initial_state)

outputs = tf.reshape(outputs, shape=[1, -1])
outputs = tf.layers.dense(outputs, 36,\
                                 kernel_initializer=tf.contrib.layers.xavier_initializer(uniform=False))
outputs = tf.reshape(outputs, shape=[1, 1, -1])
rnn_outputs_sequence=outputs
print(outputs)
for i in range(1, 16, 1):
    outputs, state=tf.nn.dynamic_rnn(cell=cell, inputs=outputs, initial_state=state)
    outputs = tf.reshape(outputs, shape=[1, -1])
    outputs = tf.layers.dense(outputs, 36,\
                                 kernel_initializer=tf.contrib.layers.xavier_initializer(uniform=False))
    outputs = tf.reshape(outputs, shape=[1, 1, -1])
    print(outputs)
    rnn_outputs_sequence=tf.concat((rnn_outputs_sequence, outputs),axis=1)
print(rnn_outputs_sequence)

使用 RNN 生成一系列张量

Generate a series of tensors using RNN

lstm

tensorflow

rnn