learning_rate_scheduler

cosine_decay

  • paddle.fluid.layers.cosinedecay(_learning_rate, step_each_epoch, epochs)
  • 使用 cosine decay 的衰减方式进行学习率调整。

在训练模型时,建议一边进行训练一边降低学习率。 通过使用此方法,学习速率将通过如下cosine衰减策略进行衰减:

learning_rate_scheduler - 图1

  • 参数:
    • learning_rate (Variable | float) - 初始学习率。
    • step_each_epoch (int) - 一次迭代中的步数。
    • epochs - 总迭代次数。代码示例
  1. import paddle.fluid as fluid
  2. base_lr = 0.1
  3. lr = fluid.layers.cosine_decay( learning_rate = base_lr, step_each_epoch=10000, epochs=120)

exponential_decay

  • paddle.fluid.layers.exponentialdecay(_learning_rate, decay_steps, decay_rate, staircase=False)
  • 在学习率上运用指数衰减。训练模型时,推荐在训练过程中降低学习率。每次 decay_steps 步骤中用 decay_rate 衰减学习率。
  1. if staircase == True:
  2. decayed_learning_rate = learning_rate * decay_rate ^ floor(global_step / decay_steps)
  3. else:
  4. decayed_learning_rate = learning_rate * decay_rate ^ (global_step / decay_steps)
  • 参数:
    • learning_rate (Variable|float)-初始学习率
    • decay_steps (int)-见以上衰减运算
    • decay_rate (float)-衰减率。见以上衰减运算
    • staircase (Boolean)-若为True,按离散区间衰减学习率。默认:False返回:衰减的学习率

返回类型:变量(Variable)

代码示例

  1. import paddle.fluid as fluid
  2. base_lr = 0.1
  3. sgd_optimizer = fluid.optimizer.SGD(
  4. learning_rate=fluid.layers.exponential_decay(
  5. learning_rate=base_lr,
  6. decay_steps=10000,
  7. decay_rate=0.5,
  8. staircase=True))

inverse_time_decay

  • paddle.fluid.layers.inversetime_decay(_learning_rate, decay_steps, decay_rate, staircase=False)
  • 在初始学习率上运用逆时衰减。

训练模型时,最好在训练过程中降低学习率。通过执行该函数,将对初始学习率运用逆向衰减函数。

  1. if staircase == True:
  2. decayed_learning_rate = learning_rate / (1 + decay_rate * floor(global_step / decay_step))
  3. else:
  4. decayed_learning_rate = learning_rate / (1 + decay_rate * global_step / decay_step)
  • 参数:
    • learning_rate (Variable|float)-初始学习率
    • decay_steps (int)-见以上衰减运算
    • decay_rate (float)-衰减率。见以上衰减运算
    • staircase (Boolean)-若为True,按间隔区间衰减学习率。默认:False返回:衰减的学习率

返回类型:变量(Variable)

示例代码:

  1. import paddle.fluid as fluid
  2. base_lr = 0.1
  3. sgd_optimizer = fluid.optimizer.SGD(
  4. learning_rate=fluid.layers.natural_exp_decay(
  5. learning_rate=base_lr,
  6. decay_steps=10000,
  7. decay_rate=0.5,
  8. staircase=True))
  9. sgd_optimizer.minimize(avg_cost)

linear_lr_warmup

  • paddle.fluid.layers.linearlr_warmup(_learning_rate, warmup_steps, start_lr, end_lr)
  • 在正常学习率调整之前先应用线性学习率热身(warm up)进行初步调整。
  1. if global_step < warmup_steps:
  2. linear_step = end_lr - start_lr
  3. lr = start_lr + linear_step * (global_step / warmup_steps)
  • 参数:
    • learning_rate (float | Variable) - 学习率,类型为float值或变量。
    • warmup_steps (int) - 进行warm up过程的步数。
    • start_lr (float) - warm up的起始学习率
    • end_lr (float) - warm up的最终学习率。返回:进行热身衰减后的学习率。

示例代码

  1. import paddle.fluid as fluid
  2. boundaries = [100, 200]
  3. lr_steps = [0.1, 0.01, 0.001]
  4. warmup_steps = 50
  5. start_lr = 1. / 3.
  6. end_lr = 0.1
  7. decayed_lr = fluid.layers.linear_lr_warmup(
  8. fluid.layers.piecewise_decay(boundaries, lr_steps),
  9. warmup_steps, start_lr, end_lr)

natural_exp_decay

  • paddle.fluid.layers.naturalexp_decay(_learning_rate, decay_steps, decay_rate, staircase=False)
  • 将自然指数衰减运用到初始学习率上。
  1. if not staircase:
  2. decayed_learning_rate = learning_rate * exp(- decay_rate * (global_step / decay_steps))
  3. else:
  4. decayed_learning_rate = learning_rate * exp(- decay_rate * (global_step / decay_steps))
  • 参数:
    • learning_rate - 标量float32值或变量。是训练过程中的初始学习率。
    • decay_steps - Python int32数
    • decay_rate - Python float数
    • staircase - Boolean.若设为true,每个decay_steps衰减学习率返回:衰减的学习率

示例代码:

  1. import paddle.fluid as fluid
  2. base_lr = 0.1
  3. sgd_optimizer = fluid.optimizer.SGD(
  4. learning_rate=fluid.layers.natural_exp_decay(
  5. learning_rate=base_lr,
  6. decay_steps=10000,
  7. decay_rate=0.5,
  8. staircase=True))

noam_decay

  • paddle.fluid.layers.noamdecay(_d_model, warmup_steps)
  • Noam衰减方法。noam衰减的numpy实现如下。
  1. import padde.fluid as fluid
  2. import numpy as np
  3. # 设置超参数
  4. d_model = 2
  5. current_steps = 20
  6. warmup_steps = 200
  7. # 计算
  8. lr_value = np.power(d_model, -0.5) * np.min([
  9. np.power(current_steps, -0.5),
  10. np.power(warmup_steps, -1.5) * current_steps])

请参照 attention is all you need

  • 参数:
    • d_model (Variable)-模型的输入和输出维度
    • warmup_steps (Variable)-超参数返回:衰减的学习率

代码示例

  1. import padde.fluid as fluid
  2. warmup_steps = 100
  3. learning_rate = 0.01
  4. lr = fluid.layers.learning_rate_scheduler.noam_decay(
  5. 1/(warmup_steps *(learning_rate ** 2)),
  6. warmup_steps)

piecewise_decay

  • paddle.fluid.layers.piecewisedecay(_boundaries, values)
  • 对初始学习率进行分段衰减。

该算法可用如下代码描述。

  1. boundaries = [10000, 20000]
  2. values = [1.0, 0.5, 0.1]
  3. if step < 10000:
  4. learning_rate = 1.0
  5. elif 10000 <= step < 20000:
  6. learning_rate = 0.5
  7. else:
  8. learning_rate = 0.1
  • 参数:
    • boundaries -一列代表步数的数字
    • values -一列学习率的值,从不同的步边界中挑选返回:衰减的学习率

代码示例

  1. import paddle.fluid as fluid
  2. boundaries = [10000, 20000]
  3. values = [1.0, 0.5, 0.1]
  4. optimizer = fluid.optimizer.Momentum(
  5. momentum=0.9,
  6. learning_rate=fluid.layers.piecewise_decay(boundaries=boundaries, values=values),
  7. regularization=fluid.regularizer.L2Decay(1e-4))

polynomial_decay

  • paddle.fluid.layers.polynomialdecay(_learning_rate, decay_steps, end_learning_rate=0.0001, power=1.0, cycle=False)
  • 对初始学习率使用多项式衰减
  1. if cycle:
  2. decay_steps = decay_steps * ceil(global_step / decay_steps)
  3. else:
  4. global_step = min(global_step, decay_steps)
  5. decayed_learning_rate = (learning_rate - end_learning_rate) *
  6. (1 - global_step / decay_steps) ^ power + end_learning_rate
  • 参数:
    • learning_rate (Variable|float32)-标量float32值或变量。是训练过程中的初始学习率。
    • decay_steps (int32)-Python int32数
    • end_learning_rate (float)-Python float数
    • power (float)-Python float数
    • cycle (bool)-若设为true,每decay_steps衰减学习率返回:衰减的学习率

返回类型:变量(Variable)

代码示例

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