tf.train.exponential_decay（）的理解与实践

self.learning_rate = tf.train.exponential_decay(
            self.initial_learning_rate, 
            self.global_step, 
            self.decay_steps,
            self.decay_rate, 
            self.staircase = Ture, name='learning_rate')

采用了学习率衰减，主要是根据训练过程中训练步数来改变学习率，可以每步都改变，也可以每间隔steps来改变，上面的例子就是每经过decay_steps步，原始的学习率将变为decay_steps倍。具体详情请看下面的代码和结果图。

import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
import yolo.config as cfg

class LearningRate(object):
    def __init__(self):
        self.init_learningrate = cfg.LEARNING_RATE
        self.decay_rate = cfg.DECAY_RATE
        self.global_steps = 1000
        self.decay_steps = cfg.DECAY_STEPS
        self.global_ = tf.Variable(tf.constant(0))
		#每decay_steps对学习率进行衰减
        self.learning_rate1 = tf.train.exponential_decay(
            self.init_learningrate, self.global_, self.decay_steps,
            self.decay_rate, staircase = True, name='learning_rate1')
            #每步都对学习率进行衰减
        self.learning_rate2 = tf.train.exponential_decay(
            self.init_learningrate, self.global_, self.decay_steps,
            self.decay_rate, staircase=False, name='learning_rate2')
T = []
F = []

def main():
    with tf.Session() as sess:
        a = LearningRate()
        for i in range(a.global_steps):
            F_c = sess.run(a.learning_rate2,feed_dict={a.global_:i})
            F.append(F_c)
            T_c = sess.run(a.learning_rate1,feed_dict={a.global_: i})
            T.append(T_c)
    plt.figure(1)
    plt.plot(range(a.global_steps),T,'r-')
    plt.plot(range(a.global_steps), F, 'b-')

    plt.show()

if __name__ == '__main__':
    main()

结果：