learning rate很重要,如何设置它有很多种方法,在Caffe源码的caffe-master\src\caffe\solvers\sgd_solver.cpp中的GetLearningRate函数注释中有对应的介绍,如下:
// policies are as follows:
// - fixed: always return base_lr.
// - step: return base_lr * gamma ^ (floor(iter / step))
// - exp: return base_lr * gamma ^ iter
// - inv: return base_lr * (1 + gamma * iter) ^ (- power)
// - multistep: similar to step but it allows non uniform steps defined by
// stepvalue
// - poly: the effective learning rate follows a polynomial decay, to be
// zero by the max_iter. return base_lr (1 - iter/max_iter) ^ (power)
// - sigmoid: the effective learning rate follows a sigmod decay
// return base_lr ( 1/(1 + exp(-gamma * (iter - stepsize))))
//
废话不多说,直接看它们的图像一目了然:
对应matlab代码是:
iter=1:50000;
max_iter=50000;
base_lr=0.01;
gamma=0.0001;
power=0.75;
step_size=5000;
% - fixed: always return base_lr.
lr=base_lr*ones(1,50000);
subplot(2,3,1)
plot(lr)
title('fixed')
% - step: return base_lr * gamma ^ (floor(iter / step))
lr=base_lr .* gamma.^(floor(iter./10000));
subplot(2,3,2)
plot(lr)
title('step')
% - exp: return base_lr * gamma ^ iter
lr=base_lr * gamma .^ iter;
subplot(2,3,3)
plot(lr)
title('exp')
% - inv: return base_lr * (1 + gamma * iter) ^ (- power)
lr=base_lr.*(1./(1+gamma.*iter).^power);
subplot(2,3,4)
plot(lr)
title('inv')
% - multistep: similar to step but it allows non uniform steps defined by
% stepvalue
% - poly: the effective learning rate follows a polynomial decay, to be
% zero by the max_iter. return base_lr (1 - iter/max_iter) ^ (power)
lr=base_lr *(1 - iter./max_iter) .^ (power);
subplot(2,3,5)
plot(lr)
title('poly')
% - sigmoid: the effective learning rate follows a sigmod decay
% return base_lr ( 1/(1 + exp(-gamma * (iter - stepsize))))
lr=base_lr *( 1./(1 + exp(-gamma * (iter - step_size))));
subplot(2,3,6)
plot(lr)
title('sigmoid')