tensorflow训练中出现nan问题

青旅半醒 2022-03-10 07:20 224阅读 0赞

深度学习中对于网络的训练是参数更新的过程，需要注意一种情况就是输入数据未做归一化时，如果前向传播结果已经是\[0,0,0,1,0,0,0,0\]这种形式，而真实结果是\[1,0,0,0,0,0,0,0,0\]，此时由于得出的结论不惧有概率性，而是错误的估计值，此时反向传播会使得权重和偏置值变的无穷大，导致数据溢出，也就出现了nan的问题。

解决办法：

1、对输入数据进行归一化处理，如将输入的图片数据除以255将其转化成0-1之间的数据;

2、对于层数较多的情况，各层都做batch\_nomorlization;

3、对设置Weights权重使用tf.truncated\_normal(0, 0.01, \[3,3,1,64\])生成，同时值的均值为0，方差要小一些;

4、激活函数可以使用tanh;

5、减小学习率lr。

实例：

import tensorflow as tf
    from tensorflow.examples.tutorials.mnist import input_data
    
    mnist = input_data.read_data_sets('data',one_hot = True)
    
    def add_layer(input_data,in_size, out_size,activation_function=None):
        Weights = tf.Variable(tf.random_normal([in_size,out_size]))
        Biases = tf.Variable(tf.zeros([1, out_size])+0.1)
        Wx_plus_b = tf.add(tf.matmul(input_data, Weights), Biases)
        if activation_function==None:
            outputs = Wx_plus_b
        else:
            outputs = activation_function(Wx_plus_b)
        #return outputs#, Weights
        return { 'outdata':outputs, 'w':Weights}
    
    def get_accuracy(t_y):
    #global l1
    #accu = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(l1['outdata'],1),tf.argmax(t_y,1)), dtype = tf.float32))
        global prediction
        accu = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(prediction['outdata'],1),tf.argmax(t_y,1)), dtype = tf.float32))
        return accu
    
    X = tf.placeholder(tf.float32, [None, 784])
    Y = tf.placeholder(tf.float32, [None, 10])
    
    #l1 = add_layer(X, 784, 10, tf.nn.softmax)
    #cross_entropy = tf.reduce_mean(-tf.reduce_sum(Y*tf.log(l1['outdata']), reduction_indices= [1]))
    #l1 = add_layer(X, 784, 1024, tf.nn.relu)
    
    l1 = add_layer(X, 784, 1024, None)
    prediction = add_layer(l1['outdata'], 1024, 10, tf.nn.softmax)
    cross_entropy = tf.reduce_mean(-tf.reduce_sum(Y*tf.log(prediction['outdata']), reduction_indices= [1]))
    
    optimizer = tf.train.GradientDescentOptimizer(0.000001)
    train = optimizer.minimize(cross_entropy)
    
    
    newW = tf.Variable(tf.random_normal([1024,10]))
    newOut = tf.matmul(l1['outdata'],newW)
    newSoftMax = tf.nn.softmax(newOut)
    init = tf.global_variables_initializer()
    with tf.Session() as sess:
        sess.run(init)
        #print(sess.run(l1_Weights))
        for i in range(2):
            X_train, y_train = mnist.train.next_batch(1)
            X_train = X_train/255   #需要进行归一化处理
            #print(sess.run(l1['w'],feed_dict={X:X_train}))
            #print(sess.run(prediction['w'],feed_dict={X:X_train, Y:y_train}))
            #print(sess.run(l1['outdata'],feed_dict={X:X_train, Y:y_train}).shape)
            print(sess.run(prediction['outdata'],feed_dict={ X:X_train, Y:y_train}))
            print(sess.run(newOut, feed_dict={ X:X_train}))
            print(sess.run(newSoftMax, feed_dict={ X:X_train}))
            print(y_train)
            #print(sess.run(l1['outdata'], feed_dict={X:X_train}))
            sess.run(train, feed_dict={ X:X_train, Y:y_train})
            if i%100 == 0:
                #print(sess.run(cross_entropy, feed_dict={X:X_train, Y:y_train}))
                accuracy = get_accuracy(mnist.test.labels)
                print(sess.run(accuracy,feed_dict={ X:mnist.test.images}))
            
            #if i%100==0:
            #print(sess.run(prediction, feed_dict={X:X_train}))
            #print(sess.run(cross_entropy, feed_dict={X:X_train,Y:y_train}))

转载作者：你不来我不老  
来源：CSDN  
原文：[https://blog.csdn.net/fireflychh/article/details/73691373][https_blog.csdn.net_fireflychh_article_details_73691373]  
版权声明：本文为博主原创文章，转载请附上博文链接！

[https_blog.csdn.net_fireflychh_article_details_73691373]: https://blog.csdn.net/fireflychh/article/details/73691373