tensorflow实现seq2seq模型中decoder出现ValueError: Dimensions must be equal, but are 512 and 277 for 'Decode

参考资料：https://www.xszz.org/faq-2/question-2018101955896.html

原代码

def decoder_lstm_based(h_decoder_in, encoder_final_state):
    cell_1 = tf.contrib.rnn.BasicLSTMCell(config.n_hidden, activation=tf.nn.leaky_relu)
    lstm_cell = tf.contrib.rnn.MultiRNNCell([cell_1 for _ in range(config.num_layers_encoder)], state_is_tuple=True)
    init_state = ( encoder_final_state)

    outputs, final_state = tf.nn.dynamic_rnn(lstm_cell, h_decoder_in, initial_state=init_state, time_major=False) #(batch_size,10,256)
    outputs_reshaped = tf.reshape(outputs,[-1,config.n_hidden]) #(batch_size*10,256)

    fc_1 = tf.layers.dense(outputs_reshaped,config.vocabulary_size)#(batch_size*10,21)
    logits = tf.reshape(fc_1,[-1,config.sentence_length+1,config.vocabulary_size]) #(batch_size,10,21)
    s_decoded = tf.nn.softmax(logits=logits, axis=2)

    return s_decoded,logits

其中第二行和第三行出现错误，按照“https://www.xszz.org/faq-2/question-2018101955896.html”的指点：

You should not reuse the same cell for the first and deeper layers, because their inputs are different, hence kernel matrices are different. Try this:

# Extra function is for readability. No problem to inline it. 
def make_cell(lstm_size): 
return tf.nn.rnn_cell.BasicLSTMCell(lstm_size, state_is_tuple=True) 
network = rnn_cell.MultiRNNCell([make_cell(num_units) for _ in range(num_layers)], state_is_tuple=True)

也就是说原来的代码是在第一层和第二层中使用了同一个cell，按照这个想法，我修改代码为：

def make_cell(lstm_size):
    return tf.contrib.rnn.BasicLSTMCell(lstm_size)

def decoder_lstm_based(h_decoder_in, encoder_final_state):
    #num_units = [config.n_hidden,config.n_hidden]

    #cell_1 = [tf.contrib.rnn.BasicLSTMCell(num_units=n, activation=tf.nn.leaky_relu) for n in num_units]
    lstm_cell = tf.contrib.rnn.MultiRNNCell([make_cell(config.n_hidden) for _ in range(config.num_layers_encoder)] , state_is_tuple=True)
    init_state = ( encoder_final_state)

    outputs, final_state = tf.nn.dynamic_rnn(lstm_cell, h_decoder_in, initial_state=init_state, time_major=False) #(batch_size,10,256)
    outputs_reshaped = tf.reshape(outputs,[-1,config.n_hidden])#(batch_size*10,256)

    fc_1 = tf.layers.dense(outputs_reshaped,config.vocabulary_size)#(batch_size*10,21)
    logits = tf.reshape(fc_1,[-1,config.sentence_length+1,config.vocabulary_size]) #(batch_size,10,21)
    s_decoded = tf.nn.softmax(logits=logits, axis=2)

    return s_decoded,logits

成功通过！

当然，也可以选择另一种方式：

def decoder_lstm_based(h_decoder_in, encoder_final_state):
    num_units = [config.n_hidden,config.n_hidden]
    cell_1 = [tf.contrib.rnn.BasicLSTMCell(num_units=n, activation=tf.nn.leaky_relu) for n in num_units]
    lstm_cell = tf.contrib.rnn.MultiRNNCell(cell_1, state_is_tuple=True)
    init_state = encoder_final_state

    outputs, final_state = tf.nn.dynamic_rnn(lstm_cell, h_decoder_in, initial_state=init_state, time_major=False) #(batch_size,10,256)
    outputs_reshaped = tf.reshape(outputs,[-1,config.n_hidden])#(batch_size*10,256)

    fc_1 = tf.layers.dense(outputs_reshaped,config.vocabulary_size)#(batch_size*10,21)
    logits = tf.reshape(fc_1,[-1,config.sentence_length+1,config.vocabulary_size]) #(batch_size,10,21)
    s_decoded = tf.nn.softmax(logits=logits, axis=2)

    return s_decoded,logits