anna_lstm 两个版本的都出现问题

[1mrliu]

运行两个版本的LSTM代码的时候，都是出现这个问题，一直没找到这个bug，所以想请教一下。 问题出现的代码位置：

运行RNN

outputs, state = tf.nn.dynamic_rnn(cell, x_one_hot, initial_state=self.initial_state)

提示出现的错误信息：（貌似是Tensor的形状问题，但是检查了代码，没有发现这个问题出在哪里了） Dimensions must be equal, but are 1024 and 595 for 'rnn/while/rnn/multi_rnn_cell/cell_0/basic_lstm_cell/MatMul_1' (op: 'MatMul') with input shapes: [1,1024], [595,2048].

[PengInGitHub]

我的也是，没看出是为什么: ValueError: Dimensions must be equal, but are 1024 and 595 for 'rnn/while/rnn/multi_rnn_cell/cell_0/basic_lstm_cell/MatMul_1' (op: 'MatMul') with input shapes: [100,1024], [595,2048].

[snow123321]

你好，请问问题解决了吗？我也遇到了这个问题，不知道是为什么

[stud2008]

按以下代码重写build_lstm函数，TF1.11测试通过 def lstm_cell(): lstm = tf.nn.rnn_cell.BasicLSTMCell(lstm_size) return tf.nn.rnn_cell.DropoutWrapper(lstm, output_keep_prob=keep_prob)

def build_lstm(lstm_size, num_layers, batch_size, keep_prob): ''' 构建lstm层
keep_prob lstm_size: lstm隐层中结点数目 num_layers: lstm的隐层数目 batch_size: batch_size ''' cell = tf.nn.rnn_cell.MultiRNNCell([lstm_cell() for _ in range(num_layers)]) initial_state = cell.zero_state(batch_size, tf.float32)
return cell, initial_state

[li-aolong]

按以下代码重写build_lstm函数，TF1.11测试通过 def lstm_cell(): lstm = tf.nn.rnn_cell.BasicLSTMCell(lstm_size) return tf.nn.rnn_cell.DropoutWrapper(lstm, output_keep_prob=keep_prob)

def build_lstm(lstm_size, num_layers, batch_size, keep_prob): ''' 构建lstm层 keep_prob lstm_size: lstm隐层中结点数目 num_layers: lstm的隐层数目 batch_size: batch_size ''' cell = tf.nn.rnn_cell.MultiRNNCell([lstm_cell() for _ in range(num_layers)]) initial_state = cell.zero_state(batch_size, tf.float32) return cell, initial_state

没错，因为原始代码相当于只构建了一个lstm单元lstm = tf.contrib.rnn.BasicLSTMCell(lstm_size)，然后在cell = tf.contrib.rnn.MultiRNNCell([drop for _ in range(num_layers)])这里相当于将刚刚构建的lstm单元重复添加了，所以需要定义一个函数来保证每次添加lstm单元时都是新生成的