main.m

% by 郑煜伟 Ewing 2016-04
clc;clear
%% 读取 image 及 label
[images_train, labels_train] = load_MNIST_data('dataSet/train-images.idx3-ubyte',...
    'dataSet/train-labels.idx1-ubyte', 'min_max_scaler', 0);
% images_train = images_train(:, 1:6000);
% labels_train = labels_train(1:6000, :);
[images_Test, labels_Test] = load_MNIST_data('dataSet/t10k-images.idx3-ubyte',...
    'dataSet/t10k-labels.idx1-ubyte', 'min_max_scaler', 0);

%% 设置 SAE训练时 参数
architecture = [784 400 200 10]; % SAE网络的结构
% 设置 AE的预选参数 及 BP的预选参数
preOption_SAE.option_AE.activation  = {'ReLU'};

preOption_SAE.option_AE.is_sparse    = 1;
preOption_SAE.option_AE.sparse_rho   = 0.01;
preOption_SAE.option_AE.sparse_beta  = 0.3;

preOption_SAE.option_AE.is_denoising = 1;
preOption_SAE.option_AE.noise_rate   = 0.15;
% preOption_SAE.option_AE.noise_layer = 'all_layers';
preOption_SAE.option_BP.activation = {'softmax'};
% 得到SAE的预选参数
option_SAE = get_SAE_option(preOption_SAE);

%% 设置 SAE预测时 参数
preOption_BPNN.activation = {'ReLU'; 'ReLU'; 'softmax'};
option_BPNN = get_BPNN_option(preOption_BPNN);

%% 调用 runSAEOnce 运行一次SAE
is_disp_network = 0; % 不展示网络
is_disp_info    = 0; % 展示信息

[opt_theta, accuracy] = run_SAE_once(images_train, labels_train, ...
    images_Test, labels_Test, ... % 数据
    architecture, ...
    option_SAE, option_BPNN, ...
    is_disp_network, is_disp_info);

% 运行30次，求得 均值、标准差， 并计算95%的置信区间
% accuracy = zeros(30, 1);
% for i = 1:30
%     [opt_theta, accuracy(i, 1)] = runSAEOnce(images_Train, labels_Train, ...
%         images_Test, labels_Test, ... % 数据
%         architecture, ...
%         option_SAE, option_BPNN, ...
%         is_disp_network, is_disp_info);
%     
%     disp(['第' num2str(i) '次迭代']);
% end

% mean_accuracy = mean(accuracy);
% std_accuracy  = sqrt(sum((accuracy - mean_accuracy) .^ 2) / (size(accuracy, 1) - 1));
% up_bound      = mean_accuracy + 1.96 * std_accuracy;
% low_bound     = mean_accuracy - 1.96 * std_accuracy;
% disp(['置信度 95% 的情况下，准确率为： ['...
%     num2str(low_bound) ',' num2str(up_bound) ']']);