LookaheadOptimizer

class paddle.fluid.optimizer.LookaheadOptimizer(inner_optimizer, alpha=0.5, k=5)[源代码]

本类实现了Lookahead优化算法:https://arxiv.org/abs/1907.08610。Lookahead优化算法在内存中保存两部分参数:快参数和慢参数。每个训练步次,inner_optimizer都更新快参数;每隔k个训练步次,Lookahead更新慢参数,如下:

\[ \begin{align}\begin{aligned}& slow\_param_t = slow\_param_{t-1} + \alpha * (fast\_param_{t-1} - slow\_param_{t-1})\\& fast\_param_t = slow\_param_t\end{aligned}\end{align} \]
参数:
  • inner_optimizer (Optimizer) - 基础优化器,如SGD
  • alpha (float) - Lookahead 的学习率
  • k (int) - 慢参数更新的频率:k次一更新

代码示例

import paddle
import paddle.fluid as fluid
import numpy as np

x = fluid.layers.data(name='x', shape=[2], dtype='float32')
label = fluid.layers.data(name="label", shape=[1], dtype="int64")
y = fluid.layers.fc(input=[x], size=2, act="softmax")
loss = fluid.layers.cross_entropy(input=y, label=label)
loss = fluid.layers.mean(x=loss)
sgd = fluid.optimizer.SGD(learning_rate=0.01)
optimizer = fluid.optimizer.LookaheadOptimizer(sgd,
                                alpha=0.5,
                                k=5)
optimizer.minimize(loss)
main_program = fluid.default_main_program()
place = fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())

feeder = fluid.DataFeeder(feed_list=[x, label], place=place)

step = 0
while(step < 10):
    step += 1
    exe.run(fluid.default_main_program(),
    feed=feeder.feed(batch_data))