Question

我正在尝试在tensorflow-1.0.0中实现一个简单的rnn，并且每次在第一次时加载脚本后重新加载脚本时，我都会收到此错误：

ValueError: Variable rnn/rnn/basic_lstm_cell/weights already exists, disallowed. Did you mean to set reuse=True in VarScope? Originally defined at:

我的脚本是：

from __future__ import print_function

import os
import nltk
import pickle
import numpy as np

import tensorflow as tf
from tensorflow.contrib import rnn

import app
from prelude import *
from utils import *

os.system('clear')


'''
    Rough idea, determine Pr[ rating | h_T]
        - experiment with:
            * one hot encoding
            * pretrained word vectors
'''
############################################################
'''
    Load data
'''
root     = os.getcwd()
data_dir = os.path.join(root, 'data/aclImdb/')
out_dir  = os.path.join(root, 'tutorials/imdb/output/')

'''
    Settings 
'''
SETTING = {'UNK'             : '<unk>'
          ,'PAD'             : '_'
          ,'End-of-Paragraph': '<EOP>'

          ,'VOCAB_SIZE'      : 6000
          ,'min-length'      : 5
          ,'max-length'      : 25}

imdb = Imdb(SETTING, data_dir, out_dir)

############################################################
'''
    RNN

    training parameters
'''
learn_rate   = 0.001
train_iters  = 100000
batch_size   = 128
display_step = 10


'''
    network parameters
'''
n_input   = SETTING['VOCAB_SIZE'] # one hot vector for each word
n_steps   = SETTING['max-length'] # maximum of 25 words per review
n_hidden  = 128
n_classes = 2


'''
    graph input
'''
X = tf.placeholder(tf.float32, [None, n_input, n_steps])
Y = tf.placeholder(tf.float32, [None, n_classes]       )

'''
    network parameters
'''
theta = {
     'W': tf.Variable(tf.random_normal([n_hidden, n_classes]))
    ,'b': tf.Variable(tf.random_normal([n_classes]))
}


'''
    @Use: given input X and parameters theta, 
          output unormalized response to 
'''
def RNN(X, theta):
    '''
        conform data shape to rnn function requirements
        X shape       : batch-size * col * row
        required shape: col * batch_size * row
    '''
    X = tf.reshape  (X  , [-1, n_input])
    X = tf.split    (X , n_steps, 0   )

    with tf.variable_scope('lstm'):
        # define instance of lstm cell
        lstm_cell = rnn.BasicLSTMCell(n_hidden, forget_bias = 1.0)
        outputs, states = rnn.static_rnn(lstm_cell, X, dtype=tf.float32)

    yhat = tf.matmul(outputs[-1],theta['W']) + theta['b']

    return yhat

Yhat = RNN(X, theta)

我知道它与变量范围有关，但我不确定如何调整变量的范围，以便每次重新加载脚本时都不会抛出错误？

Answer 1

这是整个剧本吗？你能尝试在lstm单元周围设置变量范围吗？

with tf.variable_scope('lstm'):
    outputs, states = rnn.static_rnn(lstm_cell, X, dtype=tf.float32)

关于变量范围的更多here。

张量流中的变量范围错误

1 个答案: