1.配置caffe環境
[請參考此篇博客：http://blog.csdn.net/ws_20100/article/details/48850449]
本篇介紹如何在caffe環境下，實現"圖像對圖像"的卷積神經網絡的訓練。
2.文件結構
在配置好的caffe文件夾中，進入examples目錄，創建CNN文件夾，並進入文件夾
$ cd caffe-master/examples/
$ mkdir CNN
$ cd CNN/
在CNN文件夾下面創建子文件夾
$ mkdir model snapshot TestPhotos TestLabels TrainPhotos TrainLabels
其中，
model用於以後存儲卷積核矩陣和偏置向量；
snapshot用於存儲訓練中備份的caffe模型，每一段時間存儲一次，防止斷電等一些情況；
TrainPhotos、TrainLabels分別存儲訓練集輸入和監督樣本；
TestPhotos、TestLabels分別存儲測試集輸入和監督樣本，不直接參與到訓練中。
然後，將訓練所用的輸入樣本和監督樣本分別放入到TrainPhotos和TrainLabels中去。注意，樣本文件名無所謂，但是排列次序必須一一對應。同樣，將測試所用的輸入樣本和監督樣本分別放入到TestPhotos和TestLabels中去。
3.產生訓練和測試數據
1.)產生路徑文件
在CNN文件夾下面(以下均是在此文件夾下)創建兩個路徑文件。
$ vim train.txt
輸入內容：
examples/CNN/train.h5
:wq保存文檔。
$ vim test.txt
輸入內容：
examples/CNN/test.h5
:wq保存文檔。
2.)產生訓練數據
$ vim generate_train.m
輸入內容：(把輸入圖像切成11*11的像素塊，監督圖像爲3*3的像素塊(由網絡結構和卷積核大小決定)，步長爲1個像素)
clear;close all;
%% settings
folder_input = 'TrainPhotos';
folder_label = 'TrainLabels';
savepath = 'train.h5';
size_input = 11;
size_label = 3; % size_input - 12
stride = 1;
%% initialization
data = zeros(size_input, size_input, 1, 1);
label = zeros(size_label, size_label, 1, 1);
padding = abs(size_input - size_label) / 2;
count = 0;
%% read data
filepaths_input = dir(fullfile(folder_input,'*.jpg'));
filepaths_label = dir(fullfile(folder_label,'*.jpg'));
if (length(filepaths_input)==length(filepaths_label))
length = length(filepaths_input);
else
error('The Number of Input is NOT equal to the Number of Label.');
end
%% generate data
for i = 1 : length
im_input = imread(fullfile(folder_input,filepaths_input(i).name));
im_input = rgb2ycbcr(im_input);
im_input = im2double(im_input(:, :, 1));
im_label = imread(fullfile(folder_label,filepaths_label(i).name));
im_label = im2double(im_label(:, :, 1));
if size(im_input) == size(im_label)
[hei,wid] = size(im_input);
else
error('The size of input and label are not equal.');
end
for x = 1 : stride : hei-size_input+1
for y = 1 :stride : wid-size_input+1
subim_input = im_input(x : x+size_input-1, y : y+size_input-1);
subim_label = im_label(x+padding : x+padding+size_label-1, y+padding : y+padding+size_label-1);
count = count + 1;
data(:, :, 1, count) = subim_input;
label(:, :, 1, count) = subim_label;
end
end
end
%% randomized the data and label
order = randperm(count);
data = data(:, :, 1, order);
label = label(:, :, 1, order);
%% writing to HDF5
chunksz = 128;
created_flag = false;
totalct = 0;
for batchno = 1:floor(count/chunksz)
last_read=(batchno-1)*chunksz;
batchdata = data(:,:,1,last_read+1:last_read+chunksz);
batchlabs = label(:,:,1,last_read+1:last_read+chunksz);
startloc = struct('dat',[1,1,1,totalct+1], 'lab', [1,1,1,totalct+1]);
curr_dat_sz = store2hdf5(savepath, batchdata, batchlabs, ~created_flag, startloc, chunksz);
created_flag = true;
totalct = curr_dat_sz(end);
end
h5disp(savepath);終端下輸入：
$ matlab -nodesktop -nosplash -logfile generate_train.log
-r generate_train
產生訓練數據train.h5。
3.)產生測試數據
$ vim generate_test.m
generate_test.m只需要將generate_test.m文件開頭改爲：
clear;close all;
%% settings
folder_input = 'TestPhotos';
folder_label = 'TestLabels';
savepath = 'test.h5';
size_input = 11;
size_label = 3;
stride = 30;將最後一段改成：
%% writing to HDF5
chunksz = 2;
created_flag = false;
totalct = 0;
for batchno = 1:floor(count/chunksz)
last_read=(batchno-1)*chunksz;
batchdata = data(:,:,1,last_read+1:last_read+chunksz);
batchlabs = label(:,:,1,last_read+1:last_read+chunksz);
startloc = struct('dat',[1,1,1,totalct+1], 'lab', [1,1,1,totalct+1]);
curr_dat_sz = store2hdf5(savepath, batchdata, batchlabs, ~created_flag, startloc, chunksz);
created_flag = true;
totalct = curr_dat_sz(end);
end
h5disp(savepath);
終端下輸入：
$ matlab -nodesktop -nosplash -logfile generate_test.log -r generate_test
產生測試數據test.h5。僅僅用於判斷訓練到達什麼地步。
4.建立訓練文件
1.)建立solver文件
$ vim CNN_solver.prototxt
此爲運行的配置文件，輸入以下內容：
# The train/test net protocol buffer definition
net: "examples/CNN/CNN_net.prototxt"
test_iter: 556
# Carry out testing every 500 training iterations.
test_interval: 500
# The base learning rate, momentum and the weight decay of the network.
base_lr: 0.0001
momentum: 0.9
weight_decay: 0
# The learning rate policy
lr_policy: "fixed"
# Display every 100 iterations
display: 100
# The maximum number of iterations
max_iter: 15000000
# snapshot intermediate results
snapshot: 500
snapshot_prefix: "examples/CNN/snapshot/CNN"
# solver mode: CPU or GPU
solver_mode: GPU:wq保存退出。
2.)建立net文件
$ vim CNN_net.prototxt
此爲網絡結構配置文件，可以配置網絡層數，節點數，卷積核等參數。輸入以下內容：
name: "CNN"
layer {
name: "data"
type: "HDF5Data"
top: "data"
top: "label"
hdf5_data_param {
source: "examples/CNN/train.txt"
batch_size: 128
}
include: { phase: TRAIN }
}
layer {
name: "data"
type: "HDF5Data"
top: "data"
top: "label"
hdf5_data_param {
source: "examples/CNN/test.txt"
batch_size: 2
}
include: { phase: TEST }
}
layer {
name: "conv1"
type: "Convolution"
bottom: "data"
top: "conv1"
param {
lr_mult: 1
}
param {
lr_mult: 0.1
}
convolution_param {
num_output: 128
kernel_size: 5
stride: 1
pad: 0
weight_filler {
type: "gaussian"
std: 0.001
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "relu1"
type: "ReLU"
bottom: "conv1"
top: "conv1"
}
layer {
name: "conv2"
type: "Convolution"
bottom: "conv1"
top: "conv2"
param {
lr_mult: 1
}
param {
lr_mult: 0.1
}
convolution_param {
num_output: 64
kernel_size: 3
stride: 1
pad: 0
weight_filler {
type: "gaussian"
std: 0.001
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "relu2"
type: "ReLU"
bottom: "conv2"
top: "conv2"
}
layer {
name: "conv3"
type: "Convolution"
bottom: "conv2"
top: "conv3"
param {
lr_mult: 0.1
}
param {
lr_mult: 0.1
}
convolution_param {
num_output: 1
kernel_size: 3
stride: 1
pad: 0
weight_filler {
type: "gaussian"
std: 0.001
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "loss"
type: "EuclideanLoss"
bottom: "conv3"
bottom: "label"
top: "loss"
}:wq保存退出。
5.CNN訓練
$ vim train.sh
輸入以下shell：
#!/bin/bash
cd ../../
./build/tools/caffe train --solver examples/CNN/CNN_solver.prototxt 2>&1 | tee examples/CNN/CNN.log
增加運行權限：
$ chmod +x train.sh
運行腳本文件：
$ ./train.sh
時間可能會運行幾天，也可以提前退出(Ctrl+C)，因爲在snapshot中有中間備份存儲。
6.保存濾波器
1.)創建mat文件
$ cp CNN_net.prototxt CNN_mat.prototxt
將CNN_mat.prototxt文件開頭兩個layer段改爲：
name: "CNN"
input: "data"
input_dim: 1
input_dim: 1
input_dim: 11
input_dim: 11
input: "label"
input_dim: 1
input_dim: 1
input_dim: 3
input_dim: 3:wq保存即可。
2.)創建M文件
$ vim saveFilters.m
輸入以下內容：(第7行可以更改需要轉換的caffemodel文件名)
caffe.reset_all();
clear; close all;
%% settings
%folder = 'examples/CNN/';
folder = './';
model = [folder 'CNN_mat.prototxt'];
weights = [folder 'snapshot/CNN_iter_550000.caffemodel'];
savepath = [folder 'model/x.mat'];
layers = 3;
%% load model using mat_caffe
net = caffe.Net(model,weights,'test');
%% reshap parameters
weights_conv = cell(layers,1);
for idx = 1 : layers
conv_filters = net.layers(['conv' num2str(idx)]).params(1).get_data();
[~,fsize,channel,fnum] = size(conv_filters);
if channel == 1
weights = double(ones(fsize^2, fnum));
else
weights = double(ones(channel, fsize^2, fnum));
end
for i = 1 : channel
for j = 1 : fnum
temp = conv_filters(:,:,i,j);
if channel == 1
weights(:,j) = temp(:);
else
weights(i,:,j) = temp(:);
end
end
end
weights_conv{idx} = weights;
end
%% save parameters
weights_conv1 = weights_conv{1};
weights_conv2 = weights_conv{2};
weights_conv3 = weights_conv{3};
biases_conv1 = double(net.layers('conv1').params(2).get_data());
biases_conv2 = double(net.layers('conv2').params(2).get_data());
biases_conv3 = double(net.layers('conv3').params(2).get_data());
save(savepath,'weights_conv1','biases_conv1','weights_conv2','biases_conv2','weights_conv3','biases_conv3');:wq保存。
3.)運行M文件
$
matlab -nodesktop -nosplash -logfile saveFilters.log -r
saveFilters
此時，在model中會生成x.mat文件。
7.CNN重構
已經知道了x.mat文件中，有三層卷積層的卷積核矩陣weights_conv*和偏置向量biases_conv*。
編寫一個demo_net.m文件，使用這些參數構建卷積網絡結構，對輸入圖像(矩陣)進行處理，即可得到結果。
不同應用有不同的源碼，這裏省略該文件源碼。
本文可能敘述不全面，如有錯誤，歡迎指正！
Enjoy~~
