Tensorflow中的手写数字识别

mnist = input_data.read_data_sets('MNIST_data',one_hot=True)
batch_size = 100n_batch = mnist.train.num_examples // batch_size
x =tf.placeholder(tf.float32,([None,784]))
y =tf.placeholder(tf.float32,([None,10]))

weight = tf.Variable(tf.truncated_normal([784,100],stddev=0.2))
bias = tf.Variable(tf.zeros([1,100])+0.1)
output1 = tf.nn.tanh(tf.matmul(x,weight)+bias)

weight3 = tf.Variable(tf.truncated_normal([100,100],stddev=0.2))
bias3 = tf.Variable(tf.zeros([1,100])+0.1)
output = tf.nn.relu(tf.matmul(output1,weight3)+bias3)

weight2 = tf.Variable(tf.truncated_normal([100,10],stddev=0.2))
bias2 = tf.Variable(tf.zeros([1,10])+0.1)# predict = tf.nn.softmax(tf.matmul(output,weight2)+bias2)predict = tf.matmul(output,weight2)+bias2# loss = tf.reduce_mean(tf.square(predict-y))loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=predict,labels=y))
train_step = tf.train.GradientDescentOptimizer(0.2).minimize(loss)

correct = tf.equal(tf.argmax(y,1),tf.argmax(predict,1))
accuracy = tf.reduce_mean(tf.cast(correct,tf.float32))with tf.Session(config=config) as sess:
    sess.run(tf.global_variables_initializer())    for epoch in range(21):        for i in range(n_batch):
            x_data,y_data = mnist.train.next_batch(batch_size)
            sess.run(train_step,feed_dict={x:x_data,y:y_data})
        acc = sess.run(accuracy, feed_dict={x: mnist.train.images, y: mnist.train.labels})
        print('train-acc' + str(acc))

    acc = sess.run(accuracy,feed_dict={x:mnist.test.images,y:mnist.test.labels})
    print('test-acc'+str(acc))

测试结果：
train-acc 0.9999273
test-acc  0.9717

版本2：

import tensorflow as tffrom tensorflow.examples.tutorials.mnist import input_data

mnist = input_data.read_data_sets('./MNIST_data',one_hot=True)
batch_size = 100batch = mnist.train.num_examples//batch_size
x = tf.placeholder(tf.float32,[None,28*28])
y = tf.placeholder(tf.float32,[None,10])
l1 = tf.layers.dense(x,1000)
l2 = tf.nn.sigmoid(l1)
l2 = tf.layers.dense(x,500,activation='relu')# l2 = tf.layers.dropout(l2,rate=0.2)l3 = tf.layers.dense(x,50,activation='relu')
pre = tf.layers.dense(l3,10)
loss = tf.nn.softmax_cross_entropy_with_logits(logits=pre,labels=y)
train = tf.train.GradientDescentOptimizer(0.001).minimize(loss)

accrate = tf.equal(tf.argmax(y,1),tf.argmax(pre,1))

acc = tf.reduce_mean(tf.cast(accrate,tf.float32))with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())    for j in range(20):        for i in range(batch):
            x_data,y_data = mnist.train.next_batch(batch_size)
            sess.run(train,feed_dict={x:x_data,y:y_data})            if i %10 ==0:
                print('train',sess.run(acc, feed_dict={x: x_data, y: y_data}))                # print('test',sess.run(acc,feed_dict={x:mnist.test.images,y:mnist.test.labels}))
    print('test', sess.run(acc, feed_dict={x: mnist.test.images, y: mnist.test.labels}))

结果: test 0.9727

作者：米特尼克大粉丝
链接：https://www.jianshu.com/p/800df125b906