解决tensorflow 的 Saver.restore()无法从本地读取变量的问题

最近做tensorflow 手写数字识别的时候遇到了一个问题，Saver的restore()方法无法从本地恢复变量，导致了每次都会重新训练。

原来代码

saver = tf.train.Saver(max_to_keep=5)epoch = tf.Variable(0, name='epoch', trainable=False)sess = tf.Session()sess.run(tf.global_variables_initializer())ckpt_dir = "./model/"if not os.path.exists(ckpt_dir):    os.makedirs(ckpt_dir)ckpt = tf.train.latest_checkpoint(ckpt_dir)if ckpt != None:    saver.restore(sess, ckpt)else:    print('Train from scratch')start = sess.run(epoch)

修改代码

epoch = tf.Variable(0, name='epoch', trainable=False)saver = tf.train.Saver(max_to_keep=5)sess = tf.Session()sess.run(tf.global_variables_initializer())ckpt_dir = "./model/"if not os.path.exists(ckpt_dir):    os.makedirs(ckpt_dir)ckpt = tf.train.latest_checkpoint(ckpt_dir)if ckpt != None:    saver.restore(sess, ckpt)else:    print('Train from scratch')start = sess.run(epoch)

其实主要改变的就是以下两行的顺序

epoch = tf.Variable(0, name='epoch', trainable=False)

saver = tf.train.Saver(max_to_keep=5)