-
Notifications
You must be signed in to change notification settings - Fork 3
/
rcn_dag.m
268 lines (244 loc) · 8.25 KB
/
rcn_dag.m
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
function [net, info] = rcn_dag(varargin)
% Recursive Convolutional Network for Image Super-Resolution (CVPR 2016 Submission)
% Authors: Jiwon Kim ([email protected]), Jung Kwon Lee, Jonghyuk Lee
run(fullfile(fileparts(mfilename('fullpath')),...
'snudeep', 'matlab', 'vl_setupnn.m')) ;
%% Download data
if ~exist('data', 'dir'), mkdir('data'); end
if ~exist('data/91', 'dir')
url = 'https://www.dropbox.com/s/sngf409t615mq9c/sr_data_91_291.zip?dl=1';
fprintf('Downloading images (91 and BSDS200) : %s\n', url);
unzip(url, 'data');
fprintf('Images Prepared. Two folders 91 and 291 (91+BSDS200)\n');
end
if ~exist('data/Set5', 'dir')
url = 'https://www.dropbox.com/s/v7ffhhoodo0xstv/Set5.zip?dl=1';
fprintf('Downloading images (Set5) : %s\n', url);
unzip(url, 'data');
fprintf('Images Prepared. Set5\n');
end
%% Set Options
opts.gpus = 1;
model_size = 'medium'; % one of minimal, medium, heavy
opts.test_sf = 3;
opts.resid = 1;
opts.recursive = 1;
opts.deep_supervise = 1;
if strcmp(model_size, 'minimal')
opts.dropout = 0;
opts.depth = 10; % 10 optimal5
opts.filterSize = 64; % 256 for depth 20
opts.augment = false; %data augmentation
opts.continueBest = false;
opts.train.learningRate = 0.1;
elseif strcmp(model_size, 'medium')
opts.dropout = 1;
opts.depth = 15; % 10 optimal5
opts.filterSize = 256; % 256 for depth 20
opts.augment = true; %data augmentation
opts.continueBest = false;
opts.train.learningRate = 0.01;
elseif strcmp(model_size, 'heavy')
opts.dropout = 1;
opts.depth = 15; % 10 optimal5
opts.filterSize = 384; % 256 for depth 20
opts.augment = true; %data augmentation
opts.continueBest = false;
opts.train.learningRate = 0.01;
else
fprintf('Choose model size correctly\n');
return;
end
opts.pad = 0;
opts.momentum = 0.9;
%if opts.dropout, opts.momentum = 0.99; end
opts.useBnorm = false;
exp_name = 'exp';
if opts.useBnorm
exp_name = 'exp_bn';
end
opts = vl_argparse(opts, varargin);
exp_name = sprintf('multi_obj_%s_depth%d', exp_name, opts.depth);
opts.expDir = fullfile('data','exp',exp_name);
opts.dataDir = fullfile('data', '91');
opts.imdbPath = fullfile(opts.expDir, 'imdb.mat');
opts = vl_argparse(opts, varargin);
if opts.depth <= 20
opts.train.batchSize = 64;
elseif opts.depth <= 30
if opts.filterSize <= 128
opts.train.batchSize = 32; % 18
else
opts.train.batchSize = 18;
end
else
if opts.filterSize <= 128
opts.train.batchSize = 8; % 4
else
opts.train.batchSize = 3;
end
end;
%rep = 20*3;
%if opts.dropout, rep = rep * 5; opts.train.learningRate = [0.1*ones(1,rep) 0.01*ones(1,rep) 0.001*ones(1,rep) 0.0001*ones(1,rep)]; end %*0.99 .^ (0:500);
% 0.1 %[0.1*ones(1,rep) 0.01*ones(1,rep) 0.001*ones(1,rep) 0.0001*ones(1,rep)];%*0.99 .^ (0:500);
%opts.train.numEpochs = numel(opts.train.learningRate);
opts.train.continue = 0;
%opts.train.gradRange = 1e-3 / opts.depth;
if opts.depth <= 10
opts.train.gradRange = 1e-4; % 1e-4
elseif opts.depth <= 15
opts.train.gradRange = 1e-4;%0.001*1e-4;
elseif opts.depth <= 20
opts.train.gradRange = 1e-5;%0.001*1e-4;
else
opts.train.gradRange = 1e-6;
end
opts.train.useBnorm = opts.useBnorm;
opts.train.sync = true;
opts.train.expDir = opts.expDir;
opts.train.gpus = opts.gpus;
opts.train.testPath = fullfile('data', 'Set5', 'baby_GT.bmp');
opts.train.dropout = opts.dropout;
opts.train.recursive = opts.recursive;
opts.train.momentum = opts.momentum;
opts.train.rep = 5;
if opts.augment, opts.train.rep = 3; end
% --------------------------------------------------------------------
% Prepare data
% --------------------------------------------------------------------
if exist(opts.imdbPath, 'file') & 0
imdb = load(opts.imdbPath) ;
else
imdb = getRcnImdb(opts.dataDir, opts.test_sf, opts.depth, opts.pad, opts.resid, opts.augment);
mkdir(opts.expDir) ;
%save(opts.imdbPath, '-struct', 'imdb') ;
end
% --------------------------------------------------------------------
% Initialization
% Look for a network for initialization
% If both depth and filter sizses exactly match, it's the best.
% Otherwise, search a network with the closest depth and the same number
% of filters.
% --------------------------------------------------------------------
close_depth = 0;
if opts.continueBest%% && ~strcmp(model_size, 'minimal') % *** TODO remove 0
for i=1:100
bestPath = sprintf('best/best_D%d_F%d.mat', i, opts.filterSize);
if exist(bestPath, 'file') && abs(i - opts.depth) < abs(close_depth - opts.depth)
load(bestPath);
best_net = net;
close_depth = i;
end
end
end
[net, opts.train.derOutputs] = rcn_init_dag(opts);
net.initParams();
if close_depth > 0
for i=1:numel(net.params)
net.params(i).value = best_net.params(i).value;
end
%opts.train.learningRate = 0.01;
end
% --------------------------------------------------------------------
% Train
% --------------------------------------------------------------------
info = rcn_train_dag(net, imdb, @getBatch, ...
opts.depth, opts.filterSize, opts.train, ...
'val', find(imdb.images.set == 3)) ;
% --------------------------------------------------------------------
function inputs = getBatch(opts, imdb, batch)
% --------------------------------------------------------------------
if numel(opts.gpus) > 0
inputs = {'input', gpuArray(imdb.images.data(:,:,:,batch)), ...
'label', gpuArray(imdb.images.labels(:,:,:,batch))} ;
else
inputs = {'input', imdb.images.data(:,:,:,batch), ...
'label', imdb.images.labels(:,:,:,batch)} ;
end
function imdb = getRcnImdb(dataDir, test_sf, depth, pad, diff, augment)
f_lst = dir(fullfile(dataDir, '*.*'));
ps = 2*depth+1; % patch size
stride = 21;%ps;%31;%ps - 2*pad;
nPatches = 0;
for sf = test_sf-1:test_sf
for f_iter = 1:numel(f_lst)
f_info = f_lst(f_iter);
if f_info.isdir, continue; end
im = imread(fullfile(dataDir, f_info.name));
im = rgb2ycbcr(im);
im = im(:,:,1);
imhigh = modcrop(im, sf);
for j = 1:stride:size(imhigh,1)-ps+1
for k = 1:stride:size(imhigh,2)-ps+1
if augment
nPatches = nPatches +7;
else
nPatches = nPatches +1;
end
end
end
end
end
imsuba = zeros(nPatches, ps-pad*2, ps-pad*2);
imsublowa = zeros(nPatches, ps, ps);
ind = 0;
val_ind = 0;
for sf = test_sf-1:test_sf
for f_iter = 1:numel(f_lst)
f_info = f_lst(f_iter);
if strcmp(f_info.name, 'baby_GT.bmp'), val_ind = ind; end
for aug=0:6 % 0, flip:1~3, rot:4~6
if ~augment && aug>0, continue; end
if f_info.isdir, continue; end
im = imread(fullfile(dataDir, f_info.name));
im = rgb2ycbcr(im);
im = im(:,:,1);
switch aug
case 1
im = im(:,end:-1:1);
case 2
im = im(end:-1:1,:);
case 3
im = im(end:-1:1, end:-1:1);
case 4
im = rot90(im, 1);
case 5
im = rot90(im, 2);
case 6
im = rot90(im, 3);
otherwise
end
imhigh = modcrop(im, sf);
imhigh = single(imhigh)/255;
imlow = imresize(imhigh, 1/sf, 'bicubic');
imlow = imresize(imlow, size(imhigh), 'bicubic');
imlow = max(16.0/255, min(235.0/255, imlow));
if diff, imhigh=imhigh-imlow; end;
for j = 1:stride:size(imhigh,1)-ps+1
for k = 1:stride:size(imhigh,2)-ps+1
imsub = imhigh(j+pad:j+ps-1-pad,k+pad:k+ps-1-pad);
imsublow = imlow(j:j+ps-1,k:k+ps-1);
ind = ind + 1;
imsuba(ind,:,:,:)=imsub;
imsublowa(ind,:,:,:)=imsublow;
end
end
end
end
end
fprintf('Total Subimages %d\n', ind);
imsuba = imsuba(1:ind,:,:);
imsublowa = imsublowa(1:ind,:,:);
s = randperm(ind); %shuffle
image_set = [ones(1, val_ind) 3*ones(1, ind-val_ind)];
image_set = image_set(1,s);
imsuba = imsuba(s,:,:);
imsublowa = imsublowa(s,:,:);
imsuba = reshape(imsuba, ind, ps-2*pad, ps-2*pad, 1);
imsublowa = reshape(imsublowa, ind, ps, ps, 1);
imsuba = permute(imsuba, [2 3 4 1]);
imsublowa = permute(imsublowa, [2 3 4 1]);
imdb.images.data = single(imsublowa);
imdb.images.labels = single(imsuba);
imdb.images.set = image_set;
imdb.meta.sets = {'train', 'val', 'test'} ;