VCNN初步探索

总说： 这篇博客主要是讲VCNN框架的一些东西。表示还没怎么看懂。虽然有点眉目了。还是记录一下吧。
https://github.com/Georgezhouzhou/vcnn_double-bladed

Notes on gen_training_data.m

addpath applications/deep_edge_aware_filters/utility/GT_filters/
addpath applications/deep_edge_aware_filters/utility/GT_filters/L0smoothing/
addpath data/

clear;

patch_dim = 64;
num_patches = 1000;
listing = dir('data/deepeaf/BSDS500/*.jpg');
fListing = dir('data/deepeaf/fImgs/*.jpg'); 


for m = 1 : 101
    fprintf('Extracting patch batch: %d / %d\n', m, 101);
    % extract random patches
    samples = zeros(patch_dim, patch_dim, 3, num_patches);
    labels = zeros(size(samples));
    for i = 1 : num_patches / 8  % 1000张共要随机抽取125次图，有重复，每次随机选取左上角的点提取patch
        if (mod(i,100) == 0)
            fprintf('Extracting patch: %d / %d\n', i*8, num_patches);
        end
        r_idx = random('unid', size(listing, 1));
        %由于命名顺序相同，因此图片x对应的index与该图片滤波后对应的index相同
        I = imread(strcat('data/deepeaf/BSDS500/', listing(r_idx).name));
        fI = imread(strcat('data/deepeaf/fImgs/',fListing(r_idx).name));

        orig_img_size = size(I);
        r = random('unid', orig_img_size(1) - patch_dim + 1);
        c = random('unid', orig_img_size(2) - patch_dim + 1);

        % EdgeExtract只进行了竖直方向的提取，进行左右翻转，再到后面4个90度的旋转     
        % 就涵盖了matrix的8种全部形态
        patch = I(r:r+patch_dim-1, c:c+patch_dim-1, :);
        fpatch = fI(r:r+patch_dim-1, c:c+patch_dim-1, :);
        patchHoriFlipped = fliplr(patch);
        fpatch = fliplr(fpatch);  % 同样进行翻转
        idx_list = (i-1)*8+1:(i-1)*8+8;
        for idx = 1:4
            % samples存储的是in，也就是原始图片的梯度
            % labels存储的是vout，也就是滤波后的图片的梯度
            % 8个一组，1~4存储原始图像滤波后的图片的Iy, Ix, -Iy, -Ix
            % 5~8存储着原始/滤波图片经过左对称后的图片的 Iy, Ix, -Iy, -Ix
            % 每间隔4的两张图片左右对称
            patch_rotated = im2double(imrotate(patch, (idx-1)*90));
             patch_filtered = GT_filter(patch_rotated);
                [vin, vout] = EdgeExtract(im2double(patch_rotated), im2double(patch_filtered));
             samples(:,:,:,idx_list(idx)) = vin;
             labels(:,:,:,idx_list(idx)) = vout;            

            patch_rotated = im2double(imrotate(patchHoriFlipped, (idx-1)*90));
            patch_filtered = GT_filter(patch_rotated);  
            [vin, vout] = EdgeExtract(im2double(patch_rotated), im2double(patch_filtered));            
           samples(:,:,:,idx_list(idx+4)) = vin;
             labels(:,:,:,idx_list(idx+4)) = vout;                            
        end
    end
    samples = single(samples);
    labels = single(labels);
    % save it
    filename = strcat('data/deepeaf/certainFilter/train/patches_', num2str(m));
    save(filename, '-v7.3', 'samples', 'labels');
end

Notes on init.m

首先是根据config进行相应的配置

config.GEN_OUTPUT = @gen_output_copy;

.NEW_MEM = @to_gpu (这个就是把数据转换成gpu数据，gpuArray(single(x)) )

.IM2COL = @im2col_gpu

.NONLINEARITY = @relu;

.OUT_ACT = @nonlinearity_nil;

config.COST_FUN = @L2_norm;

.misc下面放置的是杂项，其他的一些配置
.misc.current_layer = 1

初始化权值以及计算一些统计量

第一层

    r = config.weight_range;    
    conv_layer_c = 0;
    pool_layer_c = 0;
    full_layer_c = 0;
    layer_num = length(config.forward_pass_scheme)-1;
    config.layer_num = layer_num;
    config.feature_map_sizes = {};  %这个是三维，第三维表示改成的feturemap的层数或是说“厚度”，即conv_hidden_size
    config.weights = {};
    for idx = 1:layer_num
        if idx == 1
        ....
        if strcmp(config.forward_pass_scheme{idx}, 'conv_v_sr')
         ...
        elseif strcmp(config.forward_pass_scheme{idx}, 'conv_v')
                config.weights{idx} = config.NEW_MEM(randn(config.feature_map_sizes{idx}(3), ...
                                                  config.kernel_size(conv_layer_c, 1)*config.kernel_size(conv_layer_c, 2)*config.chs)*r);
                    if config.normalize_init_weights
                        config.weights{idx} = config.weights{idx} / sqrt(config.kernel_size(conv_layer_c, 1) * config.kernel_size(conv_layer_c, 2) * config.conv_hidden_size(conv_layer_c));
                    end

这里主要是针对不同层的形式，进行设置。
对于第一层来说，conv_v的weights进行初始化，如果进行normalize_init_weights，则进行normalize一下。

其他层

如果是其他层(非第一层），那么

    elseif strcmp(config.forward_pass_scheme{idx}, 'conv_v')
        conv_layer_c = conv_layer_c + 1;
            config.feature_map_sizes{idx} = [config.feature_map_sizes{idx-1}(1)-config.kernel_size(conv_layer_c,1)+1 ...
            config.feature_map_sizes{idx-1}(2)-config