Notes - model training: Save reading using the model

Save Model

# 创建Saver()节点
saver = tf.train.Saver()

# 训练过程中保存节点
save_path = saver.save(sess, "./ckpt/my_model.ckpt", global_step=epoch)

# 保存最终节点
save_path = saver.save(sess, "./ckpt/my_model_final.ckpt")

Read the model

# 创建Saver()节点
saver = tf.train.Saver()

# 读取节点
ckpt = tf.train.get_checkpoint_state('./ckpt/')

# 读取模型
saver.restore(sess, ckpt.model_checkpoint_path)

Using the model

"""
前期需要将整个计算图构建出来
但不需要像训练时init参数
"""

saver = tf.train.Saver()

with tf.Session() as sess:
    saver.restore(sess, "./ckpt/my_model_final.ckpt")

    # 测试
    print(accuracy.eval({x: my_mnist.test.images, y_: my_mnist.test.labels}))

TODO: restore time, how the parameter corresponds to the network structure