tensorflow模型保存、加载之变量重命名

话不多说，干就完了。

---

变量重命名的用处？

简单定义：简单来说就是将模型A中的参数parameter_A赋给模型B中的parameter_B

使用场景：当需要使用已经训练好的模型参数，尤其是使用别人训练好的模型参数时，往往别人模型中的参数命名方式与自己当前的命名方式不同，所以在加载模型参数时需要对参数进行重命名，使得代码更简洁易懂。

实现方法：

1）、模型保存

import os
import tensorflow as tf

weights = tf.Variable(initial_value=tf.truncated_normal(shape=[1024, 2],
                                                        mean=0.0,
                                                        stddev=0.1),
                      dtype=tf.float32,
                      name="weights")
biases = tf.Variable(initial_value=tf.zeros(shape=[2]),
                     dtype=tf.float32,
                     name="biases")

weights_2 = tf.Variable(initial_value=weights.initialized_value(),
                        dtype=tf.float32,
                        name="weights_2")

# saver checkpoint
if os.path.exists("checkpoints") is False:
    os.makedirs("checkpoints")

saver = tf.train.Saver()
with tf.Session() as sess:
    init_op = [tf.global_variables_initializer()]
    sess.run(init_op)
    saver.save(sess=sess, save_path="checkpoints/variable.ckpt")

2）、模型加载（变量名称保持不变）

import tensorflow as tf
from matplotlib import pyplot as plt
import os

current_path = os.path.dirname(os.path.abspath(__file__))

def restore_variable(sess):
    # need not initilize variable, but need to define the same variable like checkpoint
    weights = tf.Variable(initial_value=tf.truncated_normal(shape=[1024, 2],
                                                            mean=0.0,
                                                            stddev=0.1),
                          dtype=tf.float32,
                          name="weights")
    biases = tf.Variable(initial_value=tf.zeros(shape=[2]),
                         dtype=tf.float32,
                         name="biases")

    weights_2 = tf.Variable(initial_value=weights.initialized_value(),
                            dtype=tf.float32,
                            name="weights_2")

    saver = tf.train.Saver()

    ckpt_path = os.path.join(current_path, "checkpoints", "variable.ckpt")
    saver.restore(sess=sess, save_path=ckpt_path)

    weights_val, weights_2_val = sess.run(
        [
            tf.reshape(weights, shape=[2048]),
            tf.reshape(weights_2, shape=[2048])
        ]
    )

    plt.subplot(1, 2, 1)
    plt.scatter([i for i in range(len(weights_val))], weights_val)
    plt.subplot(1, 2, 2)
    plt.scatter([i for i in range(len(weights_2_val))], weights_2_val)
    plt.show()


if __name__ == '__main__':
    with tf.Session() as sess:
        restore_variable(sess)

3）、模型加载（变量重命名）

import tensorflow as tf
from matplotlib import pyplot as plt
import os

current_path = os.path.dirname(os.path.abspath(__file__))


def restore_variable_renamed(sess):
    conv1_w = tf.Variable(initial_value=tf.truncated_normal(shape=[1024, 2],
                                                            mean=0.0,
                                                            stddev=0.1),
                          dtype=tf.float32,
                          name="conv1_w")
    conv1_b = tf.Variable(initial_value=tf.zeros(shape=[2]),
                          dtype=tf.float32,
                          name="conv1_b")

    conv2_w = tf.Variable(initial_value=conv1_w.initialized_value(),
                          dtype=tf.float32,
                          name="conv2_w")

    # variable named 'weights' in ckpt assigned to current variable conv1_w
    # variable named 'biases' in ckpt assigned to current variable conv1_b
    # variable named 'weights_2' in ckpt assigned to current variable conv2_w
    saver = tf.train.Saver({
        "weights": conv1_w,
        "biases": conv1_b,
        "weights_2": conv2_w
    })

    ckpt_path = os.path.join(current_path, "checkpoints", "variable.ckpt")
    saver.restore(sess=sess, save_path=ckpt_path)

    conv1_w__val, conv2_w__val = sess.run(
        [
            tf.reshape(conv1_w, shape=[2048]),
            tf.reshape(conv2_w, shape=[2048])
        ]
    )

    plt.subplot(1, 2, 1)
    plt.scatter([i for i in range(len(conv1_w__val))], conv1_w__val)
    plt.subplot(1, 2, 2)
    plt.scatter([i for i in range(len(conv2_w__val))], conv2_w__val)
    plt.show()


if __name__ == '__main__':
    with tf.Session() as sess:
        restore_variable_renamed(sess)

总结：

    # 之前模型中叫 'weights'的变量赋值给当前的conv1_w变量

    # 之前模型中叫 'biases' 的变量赋值给当前的conv1_b变量

    # 之前模型中叫 'weights_2'的变量赋值给当前的conv2_w变量

    saver = tf.train.Saver({

        "weights": conv1_w,

        "biases": conv1_b,

        "weights_2": conv2_w

    })