Tensorflow手写识别

代码

from tensorflow.examples.tutorials.mnist import input_data

import tensorflow as tf
import matplotlib as mp
import matplotlib.pyplot as plot

mnist = input_data.read_data_sets("MnisData", one_hot=True)

#定义卷积核
filter = tf.Variable(tf.random_normal(shape=[5,5,1,3], stddev=1))

inputdataPlaceHolder = tf.placeholder(dtype=tf.float32, shape=[None, 28,28, 1])
LabelPlaceHolder = tf.placeholder(dtype=tf.float32, shape=[None,  10])
#定义网络结构
level1 =   tf.nn.conv2d(inputdataPlaceHolder, filter, strides=[1,1,1,1], padding='SAME')

level1 = tf.reduce_sum(level1,3)


level1 = tf.reshape(level1,[-1,784])
level2 = tf.layers.dense(level1, 250, activation=tf.nn.sigmoid , kernel_initializer= tf.random_normal_initializer(stddev=0.1))
level3 = tf.layers.dense(level2, 300, activation=tf.nn.sigmoid , kernel_initializer= tf.random_normal_initializer(stddev=0.1))
output = tf.layers.dense(level3, 10, activation=tf.nn.sigmoid , kernel_initializer= tf.random_normal_initializer(stddev=0.1))

#加入softmax层
output = tf.nn.softmax(output)
loss = tf.reduce_mean(  tf.square(output - LabelPlaceHolder))

global_step = tf.Variable(0)
learning_rate = tf.train.exponential_decay(1e-3, global_step, 1000, 0.9, staircase=False)

train = tf.train.AdamOptimizer(learning_rate).minimize(loss, global_step=global_step)
setpAry = []
lossAry = []
saver = tf.train.Saver()

with tf.Session() as sess:
    tf.global_variables_initializer().run()
    saver.restore(sess, "save/data-1800")
    while True:
        data, label = mnist.train.next_batch(1000)
        data = data.reshape([-1,28,28,1])
        _, step = sess.run([train, global_step], feed_dict={
            inputdataPlaceHolder:data,
            LabelPlaceHolder:label
        })

        if step % 100 == 0:
            data, label = mnist.test.next_batch(1)
            data = data.reshape([-1, 28, 28, 1])
            lossTemp,Temp = sess.run([loss, output], feed_dict={
            inputdataPlaceHolder:data,
            LabelPlaceHolder:label})
            setpAry.append(step)
            lossAry.append(lossTemp*100)
            print(lossTemp)
            mp.pyplot.clf()
            mp.pyplot.cla()
            mp.pyplot.plot(setpAry, lossAry)
            mp.pyplot.show()
            mp.pyplot.close()
            print(Temp)

            print(label)
            print("-----------------")
            saver.save(sess, "save/data", global_step=step )
            if lossTemp < 1e-6:
                break

准确率

Mnist.validation 97.88%

Mnist.test 97.55%

Mnist.train 99.69%