polynomial_decay

paddle.fluid.layers. polynomial_decay ( learning_rate, decay_steps, end_learning_rate=0.0001, power=1.0, cycle=False ) [源代码]

对初始学习率使用多项式衰减

if cycle:
    decay_steps = decay_steps * ceil(global_step / decay_steps)
else:
    global_step = min(global_step, decay_steps)
    decayed_learning_rate = (learning_rate - end_learning_rate) *
        (1 - global_step / decay_steps) ^ power + end_learning_rate

参数

  • learning_rate (Variable|float) - 训练过程中的初始学习率,数据类型为float的常数或变量。

  • decay_steps (int) - 衰减步数

  • end_learning_rate (float) - 训练过程的最终学习率

  • power (float) - 多项式衰减系数

  • cycle (bool) - step 超出 decay_steps 后是否继续循环,默认为False

返回

衰减的学习率

返回类型

变量(Variable)

代码示例

import paddle.fluid as fluid
start_lr = 0.01
total_step = 5000
end_lr = 0
lr = fluid.layers.polynomial_decay(
    start_lr, total_step, end_lr, power=1)