tensorflow训练中出现nan问题

        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}))