Mnist手写体识别是tensorflow的入门经典教程,此处的mnist的手写体识别率达到了91%,优化算法为梯度下降算法,激活函数为softmax回归,没有中间层,基本步骤可以分为七步。
1、设置变量
2、设置数据与结果的计算关系(设置图)
3、设置优化算法(梯度下降,train_step)
4、设置激活函数(sotfmax)
5、初始化数据
6、开始训练
7、模型评估
代码如下:
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
INPUT_NODE = 784
OUTPUT_NODE = 10
#梯度下降率
GDDOWN = 0.01
#训练轮数
TRAINING_TIMES = 1000
#每次训练数量
TRAINING_STEPS = 100
if __name__ == '__main__':
mnist = input_data.read_data_sets("/path/to/MNIST_data/", one_hot=True)
##首先打印出mnist图片数据的大小
print("input_data\'s train size: " , mnist.train.num_examples)
print("input_data\'s validation size: " , mnist.validation.num_examples)
print("input_data\'s test size: " , mnist.test.num_examples)
# 各个变量
x = tf.placeholder("float", shape=[None, INPUT_NODE])
y_ = tf.placeholder("float", shape=[None,OUTPUT_NODE])
W = tf.Variable(tf.zeros([INPUT_NODE, OUTPUT_NODE]))
b = tf.Variable(tf.zeros([OUTPUT_NODE]))
# 变量初始化
init = tf.initialize_all_variables()
sess = tf.InteractiveSession()
sess.run(init)
# 初始化图
y = tf.nn.softmax(tf.matmul(x, W) + b)
# 优化算法
cross_entropy = -tf.reduce_sum(y_ * tf.log(y))
train_step =tf.train.GradientDescentOptimizer(GDDOWN).minimize(cross_entropy)
for i in range(TRAINING_TIMES):
# 训练
batch_xs, batch_ys =mnist.train.next_batch(TRAINING_STEPS)
sess.run(train_step, feed_dict={x: batch_xs, y_: batch_ys})
# 模型评估
correct_prediction = tf.equal(tf.argmax(y, 1),tf.argmax(y_, 1))
accuracy =tf.reduce_mean(tf.cast(correct_prediction, "float"))
j = i+1
print("第%d轮训练,训练个数%d个" % (j,j*TRAINING_STEPS))
#print("正确率预测: " + correct_prediction + "\n")
print("当前正确率: ")
print(sess.run(accuracy,feed_dict={x: mnist.test.images, y_: mnist.test.labels}))