tensorflow 1.x 实战教程(十一)—模型的保存与恢复

目标

本文旨在介绍 tensorflow 入门知识点及实战示例，希望各位新手同学能在学习之后娴熟 tensorflow 相关根本操作

模型保存

import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("MNIST", one_hot=True)
batch_size = 64
n_batches = mnist.train.num_examples // batch_size
x = tf.placeholder(tf.float32, [None, 784])
y = tf.placeholder(tf.float32, [None, 10])
w = tf.Variable(tf.random_normal([784, 10], stddev=0.1))
b = tf.Variable(tf.zeros([10]))
predict = tf.nn.softmax(tf.matmul(x, w) + b)
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=predict, labels=y))
opt = tf.train.AdamOptimizer(0.001).minimize(loss)
correct = tf.equal(tf.argmax(y, 1), tf.argmax(predict, 1))
accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
saver = tf.train.Saver()
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    total_batch = 0
    last = 0
    best = 0
    for epoch in range(100):
        for _ in range(n_batches):
            xx,yy = mnist.train.next_batch(batch_size)
            sess.run(opt, {x:xx, y:yy})
        acc, l = sess.run([accuracy, loss], {x:mnist.test.images, y:mnist.test.labels})
        if acc > best:
            best = acc
            last = total_batch
            saver.save(sess, 'saved_model/model') # 每次只保存最好的成果
            print(epoch, acc, l)
        if total_batch - last > 5:
            print('early stop')
            break
        total_batch += 1

成果输出

0 0.9035 1.5953374
1 0.9147 1.5688152
2 0.9212 1.5580758
3 0.9234 1.552525
4 0.9239 1.5495663
5 0.9264 1.5462393
6 0.9271 1.5441632
7 0.9288 1.5419955
8 0.9302 1.5403246
12 0.9308 1.5376735
14 0.9324 1.5360526
19 0.9333 1.534032
25 0.9338 1.5329739
26 0.934 1.5326717
early stop

模型读取

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    saver.restore(sess, 'saved_model/model')
    acc, l = sess.run([accuracy, loss], {x:mnist.test.images, y:mnist.test.labels})
    print(acc, l)

成果打印

0.934 1.5326717

要点一

由于咱们只保存效果最好的模型，所以咱们在读取模型，使用相同数据进行测验的成果和练习的最后一次是相同的。

本文参阅

本文参阅：blog.csdn.net/qq_19672707…