FSRCNN网络资源学习（Fast Region-based Convolutional Network）
FSRCNN（Fast Region-based Convolutional Network）
参考:
https://arxiv.org/pdf/1608.00367v1.pdf Accelerating the Super-Resolution Convolutional Neural Networks (ECCV 2016)
1. FSRCNN放在 Caffe/examples/下
2. 运行generate_train.m和generate_test.m
生成hdf5文件：
generate_train.m
clear; close all;
%% settings
folder = 'Train/General-100';%
savepath = 'train.h5';%
size_input = 11;% There are 4 pixels padding. Paper presents 7
size_label = 19;% (11-4) *3 - (3-1)
scale = 3;
stride = 4;
%% 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;
%% generate data
filepaths = dir(fullfile(folder,'*.bmp'));
for i = 1 : length(filepaths)
image = imread(fullfile(folder, filepaths(i).name));
image = rgb2ycbcr(image);
image = im2double(image(:, :, 1));
im_label = modcrop(image, scale);
im_input = imresize(im_label, 1/scale, 'bicubic');
[hei,wid] = size(im_input);
for x = 1 : stride : hei - size_input + 1
for y = 1 : stride : wid - size_input + 1
locx = scale * (x + floor((size_input - 1)/2)) - floor((size_label + scale)/2 - 1);
locy = scale * (y + floor((size_input - 1)/2)) - floor((size_label + scale)/2 - 1);
subim_input = im_input(x : size_input + x - 1, y : size_input + y - 1);
subim_label = im_label(locx : size_label + locx - 1, locy : size_label + locy - 1);
count = count + 1;
data(:, :, 1, count) = subim_input;
label(:, :, 1, count) = subim_label;
end
end
end
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);
3.
运行./build/tools/caffe train –solver examples/FSRCNN/FSRCNN_solver.prototxt
问题：
I1208 11:11:17.523556 2658 layer_factory.hpp:77] Creating layer data
I1208 11:11:17.524196 2658 net.cpp:100] Creating Layer data
I1208 11:11:17.524271 2658 net.cpp:408] data -> data
I1208 11:11:17.524309 2658 net.cpp:408] data -> label
I1208 11:11:17.524338 2658 hdf5_data_layer.cpp:79] Loading list of HDF5 filenames from: examples/FSRCNN/train.txt
I1208 11:11:17.524843 2658 hdf5_data_layer.cpp:93] Number of HDF5 files: 1
HDF5-DIAG: Error detected in HDF5 (1.8.15-patch1) thread 140052527355712:
#000: ../../../src/H5F.c line 604 in H5Fopen(): unable to open file
major: File accessibilty
minor: Unable to open file
#001: ../../../src/H5Fint.c line 990 in H5F_open(): unable to open file: time = Thu Dec 8 11:11:17 2016
, name = 'examples/FSRCNN/train.h5', tent_flags = 0
major: File accessibilty
minor: Unable to open file
#002: ../../../src/H5FD.c line 993 in H5FD_open(): open failed
major: Virtual File Layer
minor: Unable to initialize object
#003: ../../../src/H5FDsec2.c line 343 in H5FD_sec2_open(): unable to open file: name = 'examples/FSRCNN/train.h5', errno = 2, error message = 'No such file or directory', flags = 0, o_flags = 0
major: File accessibilty
minor: Unable to open file
F1208 11:11:17.579452 2658 hdf5_data_layer.cpp:31] Failed opening HDF5 file: examples/FSRCNN/train.h5
*** Check failure stack trace: ***
@ 0x7f60848805cd google::LogMessage::Fail()
@ 0x7f6084882433 google::LogMessage::SendToLog()
@ 0x7f608488015b google::LogMessage::Flush()
@ 0x7f6084882e1e google::LogMessageFatal::~LogMessageFatal()
@ 0x7f6084cac2e3 caffe::HDF5DataLayer<>::LoadHDF5FileData()
@ 0x7f6084ca7b27 caffe::HDF5DataLayer<>::LayerSetUp()
@ 0x7f6084db90af caffe::Net<>::Init()
@ 0x7f6084dba931 caffe::Net<>::Net()
@ 0x7f6084d978ea caffe::Solver<>::InitTrainNet()
@ 0x7f6084d98c67 caffe::Solver<>::Init()
@ 0x7f6084d9900a caffe::Solver<>::Solver()
@ 0x7f6084c18fd3 caffe::Creator_SGDSolver<>()
@ 0x40a569 train()
@ 0x407060 main
@ 0x7f60837f0ac0 __libc_start_main
@ 0x407889 _start
@ (nil) (unknown)
Aborted (core dumped)
解决：
添加：
train.h5 test.h5 （Matlab2016a生成）
训练中：
~/caffe$ ./build/tools/caffe train --solver examples/FSRCNN/FSRCNN_solver.prototxt
I1208 19:20:04.535562 3428 caffe.cpp:210] Use CPU.
I1208 19:20:04.536381 3428 solver.cpp:48] Initializing solver from parameters:
test_iter: 231
test_interval: 5000
base_lr: 0.001
display: 100
max_iter: 10000000
lr_policy: "fixed"
momentum: 0.9
weight_decay: 0
snapshot: 100000
snapshot_prefix: "examples/FSRCNN/N-56-12-4"
solver_mode: CPU
net: "examples/FSRCNN/FSRCNN_net.prototxt"
train_state {
level: 0
stage: ""
}
I1208 19:20:04.536700 3428 solver.cpp:91] Creating training net from net file: examples/FSRCNN/FSRCNN_net.prototxt
I1208 19:20:04.537152 3428 net.cpp:322] The NetState phase (0) differed from the phase (1) specified by a rule in layer data
I1208 19:20:04.537380 3428 net.cpp:58] Initializing net from parameters:
name: "SR_test"
state {
phase: TRAIN
level: 0
stage: ""
}
layer {
name: "data"
type: "HDF5Data"
top: "data"
top: "label"
include {
phase: TRAIN
}
hdf5_data_param {
source: "examples/FSRCNN/train.txt"
batch_size: 128
}
}
layer {
name: "conv1"
type: "Convolution"
bottom: "data"
top: "conv1"
param {
lr_mult: 1
}
param {
lr_mult: 0.1
}
convolution_param {
num_output: 56
pad: 0
kernel_size: 5
stride: 1
weight_filler {
type: "gaussian"
std: 0.0378
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "relu1"
type: "PReLU"
bottom: "conv1"
top: "conv1"
prelu_param {
channel_shared: true
}
}
layer {
name: "conv2"
type: "Convolution"
bottom: "conv1"
top: "conv2"
param {
lr_mult: 1
}
param {
lr_mult: 0.1
}
convolution_param {
num_output: 12
pad: 0
kernel_size: 1
group: 1
stride: 1
weight_filler {
type: "gaussian"
std: 0.3536
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "relu2"
type: "PReLU"
bottom: "conv2"
top: "conv2"
prelu_param {
channel_shared: true
}
}
layer {
name: "conv22"
type: "Convolution"
bottom: "conv2"
top: "conv22"
param {
lr_mult: 1
}
param {
lr_mult: 0.1
}
convolution_param {
num_output: 12
pad: 1
kernel_size: 3
group: 1
stride: 1
weight_filler {
type: "gaussian"
std: 0.1179
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "relu22"
type: "PReLU"
bottom: "conv22"
top: "conv22"
prelu_param {
channel_shared: true
}
}
layer {
name: "conv23"
type: "Convolution"
bottom: "conv22"
top: "conv23"
param {
lr_mult: 1
}
param {
lr_mult: 0.1
}
convolution_param {
num_output: 12
pad: 1
kernel_size: 3
group: 1
stride: 1
weight_filler {
type: "gaussian"
std: 0.1179
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "relu23"
type: "PReLU"
bottom: "conv23"
top: "conv23"
prelu_param {
channel_shared: true
}
}
layer {
name: "conv24"
type: "Convolution"
bottom: "conv23"
top: "conv24"
param {
lr_mult: 1
}
param {
lr_mult: 0.1
}
convolution_param {
num_output: 12
pad: 1
kernel_size: 3
group: 1
stride: 1
weight_filler {
type: "gaussian"
std: 0.1179
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "relu24"
type: "PReLU"
bottom: "conv24"
top: "conv24"
prelu_param {
channel_shared: true
}
}
layer {
name: "conv25"
type: "Convolution"
bottom: "conv24"
top: "conv25"
param {
lr_mult: 1
}
param {
lr_mult: 0.1
}
convolution_param {
num_output: 12
pad: 1
kernel_size: 3
group: 1
stride: 1
weight_filler {
type: "gaussian"
std: 0.1179
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "relu25"
type: "PReLU"
bottom: "conv25"
top: "conv25"
prelu_param {
channel_shared: true
}
}
layer {
name: "conv26"
type: "Convolution"
bottom: "conv25"
top: "conv26"
param {
lr_mult: 1
}
param {
lr_mult: 0.1
}
convolution_param {
num_output: 56
pad: 0
kernel_size: 1
group: 1
stride: 1
weight_filler {
type: "gaussian"
std: 0.189
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "relu26"
type: "PReLU"
bottom: "conv26"
top: "conv26"
prelu_param {
channel_shared: true
}
}
layer {
name: "conv3"
type: "Deconvolution"
bottom: "conv26"
top: "conv3"
param {
lr_mult: 0.1
}
param {
lr_mult: 0.1
}
convolution_param {
num_output: 1
pad: 4
kernel_size: 9
stride: 3
weight_filler {
type: "gaussian"
std: 0.001
}
bias_filler {
type: "constant"
value: 0
}
}
}
layer {
name: "loss"
type: "EuclideanLoss"
bottom: "conv3"
bottom: "label"
top: "loss"
}
I1208 19:20:04.539083 3428 layer_factory.hpp:77] Creating layer data
I1208 19:20:04.539145 3428 net.cpp:100] Creating Layer data
I1208 19:20:04.539166 3428 net.cpp:408] data -> data
I1208 19:20:04.539196 3428 net.cpp:408] data -> label
I1208 19:20:04.539242 3428 hdf5_data_layer.cpp:79] Loading list of HDF5 filenames from: examples/FSRCNN/train.txt
I1208 19:20:04.539290 3428 hdf5_data_layer.cpp:93] Number of HDF5 files: 1
I1208 19:20:04.540756 3428 hdf5.cpp:32] Datatype class: H5T_FLOAT
I1208 19:20:04.676849 3428 net.cpp:150] Setting up data
I1208 19:20:04.676926 3428 net.cpp:157] Top shape: 128 1 11 11 (15488)
I1208 19:20:04.676942 3428 net.cpp:157] Top shape: 128 1 19 19 (46208)
I1208 19:20:04.676959 3428 net.cpp:165] Memory required for data: 246784
I1208 19:20:04.676972 3428 layer_factory.hpp:77] Creating layer conv1
I1208 19:20:04.677001 3428 net.cpp:100] Creating Layer conv1
I1208 19:20:04.677034 3428 net.cpp:434] conv1 <- data
I1208 19:20:04.677062 3428 net.cpp:408] conv1 -> conv1
I1208 19:20:04.677525 3428 net.cpp:150] Setting up conv1
I1208 19:20:04.677563 3428 net.cpp:157] Top shape: 128 56 7 7 (351232)
I1208 19:20:04.677573 3428 net.cpp:165] Memory required for data: 1651712
I1208 19:20:04.677599 3428 layer_factory.hpp:77] Creating layer relu1
I1208 19:20:04.677630 3428 net.cpp:100] Creating Layer relu1
I1208 19:20:04.677644 3428 net.cpp:434] relu1 <- conv1
I1208 19:20:04.677654 3428 net.cpp:395] relu1 -> conv1 (in-place)
I1208 19:20:04.677721 3428 net.cpp:150] Setting up relu1
I1208 19:20:04.677753 3428 net.cpp:157] Top shape: 128 56 7 7 (351232)
I1208 19:20:04.677762 3428 net.cpp:165] Memory required for data: 3056640
I1208 19:20:04.677786 3428 layer_factory.hpp:77] Creating layer conv2
I1208 19:20:04.677816 3428 net.cpp:100] Creating Layer conv2
I1208 19:20:04.677826 3428 net.cpp:434] conv2 <- conv1
I1208 19:20:04.677839 3428 net.cpp:408] conv2 -> conv2
I1208 19:20:04.677873 3428 net.cpp:150] Setting up conv2
I1208 19:20:04.677903 3428 net.cpp:157] Top shape: 128 12 7 7 (75264)
I1208 19:20:04.677914 3428 net.cpp:165] Memory required for data: 3357696
I1208 19:20:04.677932 3428 layer_factory.hpp:77] Creating layer relu2
I1208 19:20:04.677942 3428 net.cpp:100] Creating Layer relu2
I1208 19:20:04.677953 3428 net.cpp:434] relu2 <- conv2
I1208 19:20:04.677961 3428 net.cpp:395] relu2 -> conv2 (in-place)
I1208 19:20:04.677976 3428 net.cpp:150] Setting up relu2
I1208 19:20:04.678004 3428 net.cpp:157] Top shape: 128 12 7 7 (75264)
I1208 19:20:04.678015 3428 net.cpp:165] Memory required for data: 3658752
I1208 19:20:04.678025 3428 layer_factory.hpp:77] Creating layer conv22
I1208 19:20:04.678041 3428 net.cpp:100] Creating Layer conv22
I1208 19:20:04.678050 3428 net.cpp:434] conv22 <- conv2
I1208 19:20:04.678061 3428 net.cpp:408] conv22 -> conv22
I1208 19:20:04.678098 3428 net.cpp:150] Setting up conv22
I1208 19:20:04.678128 3428 net.cpp:157] Top shape: 128 12 7 7 (75264)
I1208 19:20:04.678139 3428 net.cpp:165] Memory required for data: 3959808
I1208 19:20:04.678153 3428 layer_factory.hpp:77] Creating layer relu22
I1208 19:20:04.678163 3428 net.cpp:100] Creating Layer relu22
I1208 19:20:04.678171 3428 net.cpp:434] relu22 <- conv22
I1208 19:20:04.678184 3428 net.cpp:395] relu22 -> conv22 (in-place)
I1208 19:20:04.678198 3428 net.cpp:150] Setting up relu22
I1208 19:20:04.678400 3428 net.cpp:157] Top shape: 128 12 7 7 (75264)
data_aug.m
clear;
%% To do data augmentation
folder = 'Train/General-100';
savepath = 'Train/General-100-aug/';
filepaths = dir(fullfile(folder,'*.bmp'));
for i = 1 : length(filepaths)
filename = filepaths(i).name;
[add, im_name, type] = fileparts(filepaths(i).name);
image = imread(fullfile(folder, filename));
for angle = 0: 90 :270
im_rot = rot90(image, angle);
imwrite(im_rot, [savepath im_name, '_rot' num2str(angle) '.bmp']);
for scale = 0.6 : 0.1 :0.9
im_down = imresize(im_rot, scale, 'bicubic');
imwrite(im_down, [savepath im_name, '_rot' num2str(angle) '_s' num2str(scale*10) '.bmp']);
end
end
end
下载代码：
http://download.csdn.net/detail/forest_world/9705982
参考资料：
http://blog.csdn.net/zuolunqiang/article/details/52411673 super-resolution技术日记——FSRCNN