learning_rate_policy.m

function lr_all = learning_rate_policy(init_lr, step, drop, min_lr, num_epochs)
% -------------------------------------------------------------------------
%   Description:
%       function to generate a set of learning rates for each epoch
%
%   Input:
%       - init_lr   : initial learning rate
%       - step      : number of epochs to drop learning rate
%       - drop      : learning rate drop ratio
%       - min_lr    : minimum learning rate
%       - num_epochs: total number of epochs
%
%   Output:
%       - lr_all    : learning rate for epochs
%
%   Citation: 
%       Deep Laplacian Pyramid Networks for Fast and Accurate Super-Resolution
%       Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang
%       IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017
%
%   Contact:
%       Wei-Sheng Lai
%       wlai24@ucmerced.edu
%       University of California, Merced
% -------------------------------------------------------------------------

    if( drop == 0 )
        lr_all = repmat(init_lr, 1, num_epochs);
    else
        num_drop = round(num_epochs / step) - 1;
        lr_all = init_lr * drop.^(0:num_drop);
        lr_all = repmat(lr_all, step, 1);
        lr_all = lr_all(:);
    end
    
    lr_all = max(lr_all, min_lr);
end