时间戳

首页 > 解决方案 > TensorFlow Checkpoint 恢复学习率

python-3.x - TensorFlow Checkpoint 恢复学习率

问题描述

我正在尝试使用TensorFlow checkpoint，除了Learning rate. 每次我运行时它都会重新初始化，并且不会从之前恢复。

这是我试图复制问题的玩具示例：

import numpy as np
import tensorflow as tf
  

X = tf.range(10.)
Y = 50.*X

class CGMM(tf.Module):
    def __init__(self):
        super(CGMM, self).__init__(name='CGMM')
        self.beta =  tf.Variable(1. , dtype=np.float32)
        self.learning_rate =  tf.Variable(1. , dtype=np.float32)

    @tf.function
    def objfun(self):
        beta = self.beta
        obj = tf.reduce_mean(tf.square(beta*self.X - self.Y))
        return obj

    def build_model(self,X,Y,decay_steps):

        self.X,self.Y=X,Y
        starter_learning_rate = 0.05 #0.05
        global_step = tf.Variable(1, trainable=False)
        self.learning_rate = tf.compat.v1.train.exponential_decay(starter_learning_rate, global_step,decay_steps, 0.96, staircase=True)

        optimizer = tf.compat.v1.train.RMSPropOptimizer(self.learning_rate)

        ckpt = tf.train.Checkpoint(step=tf.Variable(1) ,model=self, optimizer=optimizer)
        manager = tf.train.CheckpointManager(ckpt, './tf_ckpts_cg', max_to_keep=3)

        ckpt.restore(manager.latest_checkpoint)
        if manager.latest_checkpoint:
            print("Restored from {}".format(manager.latest_checkpoint))
        else:
            print("Initializing from scratch.")

        for i in range(500):
            optimizer.minimize(self.objfun, global_step=global_step,  var_list =  [self.beta])
            loss, beta, learning_rate = self.objfun(), self.beta, self.learning_rate().numpy()

            ckpt.step.assign_add(1)
            if (int(ckpt.step)-1) % 100 == 0:
                save_path = manager.save()
                print("Saved checkpoint for step {}: {}".format(int(ckpt.step), save_path))
                print("learning_rate : " + str(learning_rate))

        return beta


model = CGMM()
opt_beta = model.build_model(X,Y,100)

第一次运行结果：

Initializing from scratch.
Saved checkpoint for step 101: ./tf_ckpts_cg/ckpt-1
learning_rate : 0.048
Saved checkpoint for step 201: ./tf_ckpts_cg/ckpt-2
learning_rate : 0.04608
Saved checkpoint for step 301: ./tf_ckpts_cg/ckpt-3
learning_rate : 0.044236798
Saved checkpoint for step 401: ./tf_ckpts_cg/ckpt-4
learning_rate : 0.042467322
Saved checkpoint for step 501: ./tf_ckpts_cg/ckpt-5
learning_rate : 0.04076863

第二次运行结果：

Restored from ./tf_ckpts_cg/ckpt-5
Saved checkpoint for step 601: ./tf_ckpts_cg/ckpt-6
learning_rate : 0.048
Saved checkpoint for step 701: ./tf_ckpts_cg/ckpt-7
learning_rate : 0.04608
Saved checkpoint for step 801: ./tf_ckpts_cg/ckpt-8
learning_rate : 0.044236798
Saved checkpoint for step 901: ./tf_ckpts_cg/ckpt-9
learning_rate : 0.042467322
Saved checkpoint for step 1001: ./tf_ckpts_cg/ckpt-10
learning_rate : 0.04076863

如您所见Learning Rate，两次运行都重复了相同的操作，但其他变量运行良好。你能帮我解决这个问题吗？

标签： python-3.xtensorflowtensorflow2.0checkpoint

解决方案

我找到了解决方法并想离开这篇文章，也许它可能会在将来对某人有所帮助。

不得不添加self.global_step = tf.Variable(1, trainable=False)

这是完整的脚本；

import numpy as np
import tensorflow as tf
from matplotlib import pyplot as plt


X = tf.range(10.)
Y = 50.*X

class CGMM(tf.Module):
    def __init__(self):
        super(CGMM, self).__init__(name='CGMM')
        self.beta =  tf.Variable(1. , dtype=np.float32)
        self.global_step = tf.Variable(1, trainable=False)

    @tf.function
    def objfun(self):
        beta = self.beta
        obj = tf.reduce_mean(tf.square(beta*self.X - self.Y))

        return obj

    def build_model(self,X,Y,decay_steps):
        self.X,self.Y=X,Y
        starter_learning_rate = 0.05 

        learning_rate = tf.compat.v1.train.exponential_decay(starter_learning_rate, self.global_step,decay_steps, 0.96, staircase=True)

        optimizer = tf.compat.v1.train.RMSPropOptimizer(learning_rate)

        ckpt = tf.train.Checkpoint(step=tf.Variable(1) ,model=self, optimizer=optimizer)
        manager = tf.train.CheckpointManager(ckpt, './tf_ckpts_cg', max_to_keep=3)

        ckpt.restore(manager.latest_checkpoint)
        if manager.latest_checkpoint:
            print("Restored from {}".format(manager.latest_checkpoint))
        else:
            print("Initializing from scratch.")

        for i in range(500):
            optimizer.minimize(self.objfun, global_step=self.global_step,  var_list =  [self.beta])
            loss, beta = self.objfun(), self.beta

            ckpt.step.assign_add(1)
            lr_vector.append(learning_rate().numpy())
            if (int(ckpt.step)-1) % 100 == 0:
                save_path = manager.save()
                print("Saved checkpoint for step {}: {}".format(int(ckpt.step), save_path))
                print("learning_rate : " + str(learning_rate().numpy()))

        return loss_vec,beta,lr_vector

model = CGMM()
loss_vec, opt_beta,lr_vector = model.build_model(X,Y,100)

推荐阅读