E:\caffe-windows-lib\matlab\hdf5creation\predict_label_testdata_for_analysis.m
代码：
clearvars;close all;
if ~exist('ttest_data','var')||~exist('means','var')
test_data = load('ttest_data.mat'); % load ttest_data
test_data = test_data.ttest_data;
means = load('mean_file.mat');
means = means.means;
end
n = length(test_data);
% rand_ind = randperm(n);
rand_ind = 1:n;
ttest_data = test_data(rand_ind(1:1));
ntest = length(ttest_data);
ttest_label = zeros(1,1,2,ntest);
face_size = [100 100]; % height width
% width height channels nsamples
ttest_feature = zeros(face_size(1),face_size(2),3,ntest);
isShow = false;
tic;
for i = 1:ntest
img = imread( ttest_data(i).imgPath);
bbox = ttest_data(i).bbox;
face_img = img(bbox(2):bbox(2)+bbox(4),bbox(1):bbox(1)+bbox(3),:);
res_img = imresize(face_img,face_size);
if isShow
figure;
imshow(face_img);
pause;
end
ttest_label(:,:,1,i) = ttest_data(i).poseId + 2; % 0 1 2 3 4
ttest_label(:,:,2,i) = ttest_data(i).smile ;
ttest_feature(:,:,:,i) = res_img;
end
f_time = toc;
disp(['feature_time :' num2str(f_time)]);
% change rgb to bgr
ttest_feature = ttest_feature(:,:,[3 2 1],:);
% permute width height
ttest_feature = permute(ttest_feature,[2 1 3 4]);
% subtract mean_file
ttest_feature = ttest_feature - means;
ttest_feature = ttest_feature ./ 255;
net_model = 'deploy_1.prototxt';
net_weights = 'lenet_iter_10000.caffemodel';
phase = 'test';
if exist('../+caffe', 'dir')
addpath('..');
else
error('Please run this demo from caffe/matlab/demo');
end
net = caffe.Net(net_model, net_weights, phase);
% input_feature must be a cell
scores = net.forward({ttest_feature});
%% 获取网络中的blob
input_img = net.blobs('data').get_data();
conv1 = net.blobs('conv_all').get_data();
% input_image : w * h * c * n : 为 255 * 255 * 3
% conv1: w * h * c * n :为 96 * 96 *50,
% change w*h*c to h*w*c 即rows * cols * c
input_img = permute(input_img,[2 1 3]);
% change bgr to rgb
input_img = input_img(:,:,[3 2 1]);
ind1 = 6;
ind2 = 2;
conv11 = conv1(:,:,ind1);
conv11_img = permute(conv11,[2,1,3]);
%%
conv12 = conv1(:,:,ind2);
conv12_img = permute(conv12,[2,1,3]);
figure;
subplot(221);
imshow(res_img);
title('原图像');
subplot(222);
imshow(input_img,[]);
title('输入图像(减去均值)');
subplot(223);
imshow(conv11_img,[]);
title('响应map1');
subplot(224);
imshow(conv12_img,[]);
title('响应map2');
%% 权重的可视化
w = net.layers('conv_all').params(1).get_data();
disp(size(w));
b = net.layers('conv_all').params(2).get_data();
disp(size(b));
% change width and height
w = permute(w,[2 1 3 4]);
% change bgr to rgb
w = w(:,:,[3 2 1],:);
w11 = w(:,:,1,ind1);
w12 = w(:,:,2,ind1);
w13 = w(:,:,3,ind1);
figure(2);
% red channel
subplot(221);
imshow(w11,[]);
title('第一个滤波器的R通道');
% green channel
subplot(222);
imshow(w12,[]);
title('第一个滤波器的G通道');
% blue channel
subplot(223);
imshow(w13,[]);
title('第一个滤波器的B通道');
%% 参看是如何计算的
% using: input_img ,w11 w12 w13 ,conv11_img
% figure(4);
% imshow(input_img,[]);
figure(3);
% mean_img
subplot(441);
imshow(input_img(:,:,1),[]);
title({['大小为:' num2str(size(input_img(:,:,1)))];'均值图像的R通道'});
subplot(445);
imshow(input_img(:,:,2),[]);
title('均值图像的G通道');
subplot(449);
imshow(input_img(:,:,3),[]);
title('均值图像的B通道');
% filter
subplot(442);
imshow(w11,[]);
title({['大小为:' num2str(size(w11))];'滤波器的R通道'});
subplot(446);
imshow(w12,[]);
title('滤波器的G通道');
subplot(4,4,10);
imshow(w13,[]);
title('滤波器的B通道');
% map1
subplot(447);
imshow(conv11_img,[]);
title({['大小为:' num2str(size(conv11_img))];'响应map'});
%% check 计算过程:
% using: input_img ,w11 w12 w13 ,conv11_img
% 因为matlab卷积进行了旋转180度。
w11 = rot90(w11,2);
w12 = rot90(w12,2);
w13 = rot90(w13,2);
mapr = conv2(input_img(:,:,1),w11,'valid');%valid 只取有效部分
mapg = conv2(input_img(:,:,2),w12,'valid');
mapb = conv2(input_img(:,:,3),w13,'valid');
map = mapr + mapg + mapb;
map = map + b(ind1);
%因为代码中带了个relu层
map(find(map<0)) = 0;
subplot(448)
imshow(map,[]);
title({['大小为:' num2str(size(map))];'计算的map'});
运行结果：
图中，左边三张图像是输入单个图像的三个通道（已经减去均值的图像）,从左第2栏是三个通道对应的学习的滤波器权重(额外的一个偏置，不能画出来），第三栏是Caffe给出的卷积层+relu的结果（因为我们卷积层后面跟着一个in place relu操作），最后一栏是模拟caffe操作形成的同样的结果。
在caffe中的卷积层的训练参数为如下，我们没有设置卷积层的group参数，默认为1,即将所有通道看成一组。因此操作为：
conv_oper= w11 * img_r + w12 * img_g+w13 * img__b + bias
img_r,img_g,img_b是img的三个通道图像,bias是偏置。
layer {
name: "conv_all"
type: "Convolution"
bottom: "data"
top: "conv_all"
convolution_param {
num_output: 50
kernel_size: 5
stride: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
}
}
}
layer {
name: "relu_all"
type: "ReLU"
bottom: "conv_all"
top: "conv_all"
}实验代码2：
clearvars;close all;
if ~exist('ttest_data','var')||~exist('means','var')
test_data = load('ttest_data.mat'); % load ttest_data
test_data = test_data.ttest_data;
means = load('mean_file.mat');
means = means.means;
end
n = length(test_data);
% rand_ind = randperm(n);
rand_ind = 1:n;
ttest_data = test_data(rand_ind(1:1));
ntest = length(ttest_data);
ttest_label = zeros(1,1,2,ntest);
face_size = [100 100]; % height width
% width height channels nsamples
ttest_feature = zeros(face_size(1),face_size(2),3,ntest);
isShow = false;
tic;
for i = 1:ntest
img = imread( ttest_data(i).imgPath);
bbox = ttest_data(i).bbox;
face_img = img(bbox(2):bbox(2)+bbox(4),bbox(1):bbox(1)+bbox(3),:);
res_img = imresize(face_img,face_size);
if isShow
figure;
imshow(face_img);
pause;
end
ttest_label(:,:,1,i) = ttest_data(i).poseId + 2; % 0 1 2 3 4
ttest_label(:,:,2,i) = ttest_data(i).smile ;
ttest_feature(:,:,:,i) = res_img;
end
f_time = toc;
disp(['feature_time :' num2str(f_time)]);
% change rgb to bgr
ttest_feature = ttest_feature(:,:,[3 2 1],:);
% permute width height
ttest_feature = permute(ttest_feature,[2 1 3 4]);
% subtract mean_file
ttest_feature = ttest_feature - means;
ttest_feature = ttest_feature ./ 255;
net_model = 'deploy.prototxt';
net_weights = 'lenet_iter_10000.caffemodel';
phase = 'test';
if exist('../+caffe', 'dir')
addpath('..');
else
error('Please run this demo from caffe/matlab/demo');
end
net = caffe.Net(net_model, net_weights, phase);
% input_feature must be a cell
scores = net.forward({ttest_feature});
%% 获取网络中的blob
input_img = net.blobs('data').get_data();
conv1 = net.blobs('conv_all').get_data();
% input_image : w * h * c * n : 为 255 * 255 * 3
% conv1: w * h * c * n :为 96 * 96 *60,
% change w*h*c to h*w*c 即rows * cols * c
input_img = permute(input_img,[2 1 3]);
% change bgr to rgb
input_img = input_img(:,:,[3 2 1]);
ind1 = 6;
% change width to height
conv1 = permute(conv1,[2,1,3,4]);
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% point1
conv1b = conv1(:,:,1); % blue
conv1g = conv1(:,:,21); % green
conv1r = conv1(:,:,41); % red
figure;
subplot(121);
imshow(res_img);
title('原图像');
subplot(122);
imshow(input_img,[]);
title('输入图像(减去均值)');
%% 权重的可视化
w = net.layers('conv_all').params(1).get_data();% 5 * 5 * 1 * 60
disp(size(w));
b = net.layers('conv_all').params(2).get_data(); % 60 * 1
disp(size(b));
% change width and height
w = permute(w,[2 1 3 4]);
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% point 2
w1b = w(:,:,1,1);
w1g = w(:,:,1,21);
w1r = w(:,:,1,41);
figure(2);
% red channel
subplot(221);
imshow(w1b,[]);
title('第一个滤波器的b通道');
% green channel
subplot(222);
imshow(w1g,[]);
title('第一个滤波器的G通道');
% blue channel
subplot(223);
imshow(w1r,[]);
title('第一个滤波器的r通道');
%% 参看是如何计算的
% using: input_img ,w11 w12 w13 ,conv11_img
% figure(4);
% imshow(input_img,[]);
figure(3);
% mean_img
subplot(441);
imshow(input_img(:,:,1),[]);
title({['大小为:' num2str(size(input_img(:,:,1)))];'均值图像的R通道'});
subplot(445);
imshow(input_img(:,:,2),[]);
title('均值图像的G通道');
subplot(449);
imshow(input_img(:,:,3),[]);
title('均值图像的B通道');
% filter
subplot(442);
imshow(w1r,[]);
title({['大小为:' num2str(size(w1b))];'滤波器的r通道：41'});
subplot(446);
imshow(w1g,[]);
title('滤波器的G通道:21');
subplot(4,4,10);
imshow(w1b,[]);
title('滤波器的b通道:1');
% map1
subplot(443);
imshow(conv1r,[]);
title({['大小为:' num2str(size(conv1b))];'conv1r'});
subplot(447);
imshow(conv1g,[]);
title('conv1g');
subplot(4,4,11);
imshow(conv1b,[]);
title('conv1b');
%% check 计算过程:
% using: input_img ,w11 w12 w13 ,conv11_img
% 因为matlab卷积进行了旋转180度。
w1b = rot90(w1b,2);
w1g = rot90(w1g,2);
w1r = rot90(w1r,2);
mapr = conv2(input_img(:,:,1),w1r,'valid');%valid 只取有效部分
mapg = conv2(input_img(:,:,2),w1g,'valid');
mapb = conv2(input_img(:,:,3),w1b,'valid');
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% point 3
mapb = mapb + b(1);
mapg = mapg + b(21);
mapr = mapr + b(41);
%因为代码中带了个relu层
mapb(find(mapb<0)) = 0;
mapg(find(mapg<0)) = 0;
mapr(find(mapr<0)) = 0;
subplot(444)
imshow(mapr,[]);
title({['大小为:' num2str(size(mapb))];'计算的mapr'});
subplot(448)
imshow(mapg,[]);
title('计算的mapg');
subplot(4,4,12)
imshow(mapb,[]);
title('计算的mapb');
%%% 我们可以检查数据: mapb == conv1b? mapr == conv1r? mapg == conv1g?
实验中，我们将第一个conv1层的group设置为3，输入数据为三通道的彩色图像(采用测试时跟踪数据的流向）.
部分deploy.prototxt协议如下：
# 数据输入
layer
{
name:"data"
type:"Input"
top:"data"
input_param{shape:{dim:1 dim:3 dim:100 dim:100}}
}
# 第一个卷积层
layer {
name: "conv_all"
type: "Convolution"
bottom: "data"
top: "conv_all"
convolution_param {
num_output: 60
kernel_size: 5
stride: 1
group: 3
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
}
}
}结果图像：
执行过程如下图：
说明，对于单图像的预测，首先将输入图像减去均值图像，然后分离出三个通道b,g,r，然后首先将b通道图像与20个滤波器卷积，然后将g通道与20个滤波器卷积,r亦如此。所以为了演示三个通道的卷积过程，会采用下标1,21,41的访问形式。
4297
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
