Tensorflow GRU 变量已经存在 ValueError

Question

我正在尝试使用带有 GRU 单元的 RNN 来生成文本，但是当我在文本上调用 RNN 函数来生成文本时，出现错误。

Traceback (most recent call last):
  File "trump.py", line 93, in <module>
    var = RNN(emb(new[-3:]), weights, biases)
  File "trump.py", line 58, in RNN
    outputs, states = tf.nn.static_rnn(stack, lists, dtype = tf.float32)
  File "C:\Users\Abhinav Kumar\Miniconda3\lib\site-packages\tensorflow\python\ops\rnn.py", line 1212, in static_rnn
    (output, state) = call_cell()
  File "C:\Users\Abhinav Kumar\Miniconda3\lib\site-packages\tensorflow\python\ops\rnn.py", line 1199, in <lambda>
    call_cell = lambda: cell(input_, state)
  File "C:\Users\Abhinav Kumar\Miniconda3\lib\site-packages\tensorflow\python\ops\rnn_cell_impl.py", line 180, in __call__
    return super(RNNCell, self).__call__(inputs, state)
  File "C:\Users\Abhinav Kumar\Miniconda3\lib\site-packages\tensorflow\python\layers\base.py", line 441, in __call__
    outputs = self.call(inputs, *args, **kwargs)
  File "C:\Users\Abhinav Kumar\Miniconda3\lib\site-packages\tensorflow\python\ops\rnn_cell_impl.py", line 916, in call
    cur_inp, new_state = cell(cur_inp, cur_state)
  File "C:\Users\Abhinav Kumar\Miniconda3\lib\site-packages\tensorflow\python\ops\rnn_cell_impl.py", line 752, in __call__
    output, new_state = self._cell(inputs, state, scope)
  File "C:\Users\Abhinav Kumar\Miniconda3\lib\site-packages\tensorflow\python\ops\rnn_cell_impl.py", line 180, in __call__
    return super(RNNCell, self).__call__(inputs, state)
  File "C:\Users\Abhinav Kumar\Miniconda3\lib\site-packages\tensorflow\python\layers\base.py", line 441, in __call__
    outputs = self.call(inputs, *args, **kwargs)
  File "C:\Users\Abhinav Kumar\Miniconda3\lib\site-packages\tensorflow\python\ops\rnn_cell_impl.py", line 295, in call
    self._kernel_initializer))
  File "C:\Users\Abhinav Kumar\Miniconda3\lib\site-packages\tensorflow\python\ops\rnn_cell_impl.py", line 1017, in _linear
    initializer=kernel_initializer)
  File "C:\Users\Abhinav Kumar\Miniconda3\lib\site-packages\tensorflow\python\ops\variable_scope.py", line 1065, in get_variabl
e
    use_resource=use_resource, custom_getter=custom_getter)
  File "C:\Users\Abhinav Kumar\Miniconda3\lib\site-packages\tensorflow\python\ops\variable_scope.py", line 962, in get_variable

    use_resource=use_resource, custom_getter=custom_getter)
  File "C:\Users\Abhinav Kumar\Miniconda3\lib\site-packages\tensorflow\python\ops\variable_scope.py", line 360, in get_variable

    validate_shape=validate_shape, use_resource=use_resource)
  File "C:\Users\Abhinav Kumar\Miniconda3\lib\site-packages\tensorflow\python\ops\variable_scope.py", line 1405, in wrapped_cus
tom_getter
    *args, **kwargs)
  File "C:\Users\Abhinav Kumar\Miniconda3\lib\site-packages\tensorflow\python\ops\rnn_cell_impl.py", line 183, in _rnn_get_vari
able
    variable = getter(*args, **kwargs)
  File "C:\Users\Abhinav Kumar\Miniconda3\lib\site-packages\tensorflow\python\ops\rnn_cell_impl.py", line 183, in _rnn_get_vari
able
    variable = getter(*args, **kwargs)
  File "C:\Users\Abhinav Kumar\Miniconda3\lib\site-packages\tensorflow\python\ops\variable_scope.py", line 352, in _true_getter

    use_resource=use_resource)
  File "C:\Users\Abhinav Kumar\Miniconda3\lib\site-packages\tensorflow\python\ops\variable_scope.py", line 664, in _get_single_
variable
    name, "".join(traceback.format_list(tb))))
ValueError: Variable rnn/multi_rnn_cell/cell_0/gru_cell/gates/kernel already exists, disallowed. Did you mean to set reuse=True
 in VarScope? Originally defined at:

  File "C:\Users\Abhinav Kumar\Miniconda3\lib\site-packages\tensorflow\python\framework\ops.py", line 1269, in __init__
    self._traceback = _extract_stack()
  File "C:\Users\Abhinav Kumar\Miniconda3\lib\site-packages\tensorflow\python\framework\ops.py", line 2506, in create_op
    original_op=self._default_original_op, op_def=op_def)
  File "C:\Users\Abhinav Kumar\Miniconda3\lib\site-packages\tensorflow\python\framework\op_def_library.py", line 767, in apply_
op
    op_def=op_def)

代码，从RNN函数开始，如下：

# -*- coding: UTF-8 -*-
import tensorflow as tf
import os
import numpy as np
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'


wordp = ""
with open("test.txt", encoding = "UTF-8") as file:
    reader = file.readlines()
    for row in reader:
        wordp += row + " "

words = set(wordp.split())

dictionary = {i: j for i, j in enumerate(words)}
reverse = {j: i for i, j in enumerate(words)}
training_epochs = 1
learning_rate = .001
input_dim = len(words)
output_dim = 1
num_units = 300
num_units1 = 300
output_prob = .5
train = []

for line in reader:
    line = line.split()
    for i in range(len(line)):
        x = line[i]
        line[i] = np.zeros(len(dictionary))
        line[i][reverse[x]] = 1
    train.append(line)

def emb(param):
    train = []
    global dictionary
    global reverse
    for i in range(len(param)):
        x = param[i]
        param[i] = np.zeros(len(dictionary))
        param[i][reverse[x]] = 1
    return np.asarray(param, dtype = np.float32)

def RNN(x, weights, biases):
    cells = []

    LSTM = tf.contrib.rnn.GRUCell(num_units, activation=tf.nn.relu)
    LSTM = tf.contrib.rnn.DropoutWrapper(LSTM, output_keep_prob=output_prob)
    cells.append(LSTM)
    LSTM1 = tf.contrib.rnn.GRUCell(num_units1, activation=tf.nn.relu)
    LSTM1 = tf.contrib.rnn.DropoutWrapper(LSTM1, output_keep_prob=output_prob)
    cells.append(LSTM1)
    stack = tf.contrib.rnn.MultiRNNCell(cells, state_is_tuple=True)
    lists = tf.split(x, num_or_size_splits = 3, axis = 0)


    outputs, states = tf.nn.static_rnn(stack, lists, dtype = tf.float32)
    return tf.matmul(outputs[-1], weights)+biases

weights = tf.Variable(tf.random_normal([num_units,input_dim]))
biases = tf.Variable(tf.random_normal([input_dim]))
x_vals = tf.placeholder(tf.float32, shape = [None, input_dim])
y_vals = tf.placeholder(tf.float32, shape = [input_dim])
pred = RNN(x_vals,weights,biases)
cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=pred, labels=y_vals))
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cross_entropy)
init = tf.global_variables_initializer()

with tf.Session() as sess:
    sess.run(init)
    for j in range(training_epochs):
        accu = 0
        c = 0
        count = 0
        for line in train:
            i = 0
            while i+4 < len(line):
                _, acc, cost = sess.run([optimizer, accuracy, cross_entropy], feed_dict={x_vals: line[i:i+3], y_vals: line[i+3]})
                i+=1
                accu += acc
                count += 1
                c+= cost
        print("Accuracy")
        print(accu/count)
        print("Cost")
        print(c/count)

    new = ["badly", "broken", "big"]
    while len(" ".join(new)) <= 140:  
        var = RNN(emb(new[-3:]), weights, biases)
        sess.run(var)
        new.append(dictionary[var.index(max(var))])

训练部分工作正常，但是当我第二次尝试调用 RNN 函数时出现问题。我尝试将 RNN 函数内部和外部的各种变量放在不同的变量范围内，但随后我收到一条错误消息，提示 GRU 子变量未初始化。

感谢您的帮助！

Answer 1

您的代码中有几个问题：

• accuracy 未定义（我假设您发布了脚本的中间版本）

• RNN 是一个生成图表的函数（LSTM、LSTM1 等是节点）。

  它在训练中正常工作，但在测试时第二次调用会在图中创建新节点，因此出现此错误消息。

我有这样的内联 RNN 函数：

weights = tf.Variable(tf.random_normal([num_units,input_dim]))
biases = tf.Variable(tf.random_normal([input_dim]))
x_vals = tf.placeholder(tf.float32, shape = [None, input_dim])
y_vals = tf.placeholder(tf.float32, shape = [input_dim])

cells = []
LSTM = tf.contrib.rnn.GRUCell(num_units, activation=tf.nn.relu)
LSTM = tf.contrib.rnn.DropoutWrapper(LSTM, output_keep_prob=output_prob)
cells.append(LSTM)
LSTM1 = tf.contrib.rnn.GRUCell(num_units1, activation=tf.nn.relu)
LSTM1 = tf.contrib.rnn.DropoutWrapper(LSTM1, output_keep_prob=output_prob)
cells.append(LSTM1)
stack = tf.contrib.rnn.MultiRNNCell(cells, state_is_tuple=True)
lists = tf.split(x_vals, num_or_size_splits = 3, axis = 0)
outputs, states = tf.nn.static_rnn(stack, lists, dtype = tf.float32)
pred = tf.matmul(outputs[-1], weights)+biases

并将测试目标更改为tf.argmax(pred, dimension=1)：

while len(" ".join(new)) <= 140:
    test_input = emb(new[-3:])
    max_word = tf.argmax(pred, dimension=1)[0]
    word_index = sess.run(max_word, feed_dict={x_vals: test_input})
    print dictionary[word_index]

它为我生成了一些单词。我没有大的训练文件，所以输出对我来说毫无意义，但希望对你来说会更好。