InitNetParam.m

function [net_param] = InitNetParam
train_param.lr = 0.001;
train_param.gamma = 0.1;
train_param.mm = 0.9;
train_param.weight_decay = 0.005;
train_param.max_iter = 4 * 100000;
% train_param.train_num = 60000;
train_param.batch_size = 64;
% train_param.step_size = 108465;
train_param.step_size = 100000;
train_param.display_interval = 5000;
test_param.batch_size = 1000;
test_param.test_interval = 2000000;
% network architecture. Even entries indicates the feature dimension of input and output, respectively. 
% architecture = {'input',[784], 'fc', [784, 64], 'relu', [64],...
%     'fc', [64, 64], 'relu', [64],...
%      'fc', [64, 32], 'relu', [32],...
%      'fc', [32, 32], 'relu', [32],...
%     'fc', [32, 10], 'euclidean', [10,1]};
% architecture = {'input',[784], 'fc', [784, 64], 'relu', [64],...
%     'fc', [64, 64], 'relu', [64],...
%     'fc', [64, 10], 'euclidean', [10,1]};
architecture = {'input',[405], 'fc', [405, 128], 'relu', [128],...
    'fc', [128,128], 'relu', [128],...
    'fc', [128, 81], 'euclidean', [81,1]};
% architecture = {'input',[405], 'fc', [405, 128], 'relu', [128],...
%     'fc', [128,128], 'relu', [128],...
%     'fc', [128,128], 'relu', [128],...
%     'fc', [128, 81], 'euclidean', [81,1]};
net_param.train_param = train_param;
net_param.test_param = test_param;
net_param.architecture = architecture;
end