-
Notifications
You must be signed in to change notification settings - Fork 6
/
Copy pathfcnInitializeModel_3Conv_Dense.m
297 lines (252 loc) · 9.43 KB
/
fcnInitializeModel_3Conv_Dense.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
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
function net = fcnInitializeModel_3Conv_Dense(varargin)
%FCNINITIALIZEMODEL8S Initialize the FCN-8S model from FCN-16S
opts.sourceModelUrl = 'http://www.vlfeat.org/matconvnet/models/imagenet-vgg-verydeep-16.mat' ;
opts.sourceModelPath = 'data/models/imagenet-vgg-verydeep-16.mat' ;
opts = vl_argparse(opts, varargin) ;
% -------------------------------------------------------------------------
% Load & download the source model if needed (VGG VD 16)
% -------------------------------------------------------------------------
if ~exist(opts.sourceModelPath)
fprintf('%s: downloading %s\n', opts.sourceModelUrl) ;
mkdir(fileparts(opts.sourceModelPath)) ;
urlwrite('http://www.vlfeat.org/matconvnet/models/imagenet-vgg-verydeep-16.mat', opts.sourceModelPath) ;
end
net = vl_simplenn_tidy(load(opts.sourceModelPath)) ;
% for convt (deconv) layers, cuDNN seems to be slower?
net.meta.cudnnOpts = {'cudnnworkspacelimit', 512 * 1024^3} ;
%net.meta.cudnnOpts = {'nocudnn'} ;
% -------------------------------------------------------------------------
% Edit the model to create the FCN version
% -------------------------------------------------------------------------
% Add dropout to the fully-connected layers in the source model
drop1 = struct('name', 'dropout1', 'type', 'dropout', 'rate' , 0.5) ;
drop2 = struct('name', 'dropout2', 'type', 'dropout', 'rate' , 0.5) ;
net.layers = [net.layers(1:33) drop1 net.layers(34:35) drop2 net.layers(36:end)] ;
% Convert the model from SimpleNN to DagNN
net = dagnn.DagNN.fromSimpleNN(net, 'canonicalNames', true) ;
% Add more padding to the input layer
%net.layers(1).block.pad = 100 ;
net.layers(5).block.pad = [0 1 0 1] ;
net.layers(10).block.pad = [0 1 0 1] ;
net.layers(17).block.pad = [0 1 0 1] ;
net.layers(24).block.pad = [0 1 0 1] ;
net.layers(31).block.pad = [0 1 0 1] ;
net.layers(32).block.pad = [3 3 3 3] ;
% ^-- we could do [2 3 2 3] but that would not use CuDNN
% Modify the bias learning rate for all layers
for i = 1:numel(net.layers)-1
if (isa(net.layers(i).block, 'dagnn.Conv') && net.layers(i).block.hasBias)
filt = net.getParamIndex(net.layers(i).params{1}) ;
bias = net.getParamIndex(net.layers(i).params{2}) ;
net.params(bias).learningRate = 2 * net.params(filt).learningRate ;
end
end
% Modify the last fully-connected layer to have 21 output classes
% Initialize the new filters to zero
for i = [1 2]
p = net.getParamIndex(net.layers(end-1).params{i}) ;
if i == 1
sz = size(net.params(p).value) ;
sz(end) = 2 ;
else
sz = [2 1] ;
end
net.params(p).value = zeros(sz, 'single') ;
end
net.layers(end-1).block.size = size(...
net.params(net.getParamIndex(net.layers(end-1).params{1})).value) ;
% Remove the last loss layer
net.removeLayer('prob') ;
net.setLayerOutputs('fc8', {'x38'}) ;
%% Add skip layers on top of pool5
net.addLayer('skip5', ...
dagnn.Conv('size', [1 1 512 2], 'pad', 0), ...
'x30', 'x39', {'skip5f','skip5b'});
f = net.getParamIndex('skip5f') ;
net.params(f).value = zeros(1, 1, 512, 2, 'single') ;
net.params(f).learningRate = 0.1 ;
net.params(f).weightDecay = 1 ;
f = net.getParamIndex('skip5b') ;
net.params(f).value = zeros(1, 1, 2, 'single') ;
net.params(f).learningRate = 2 ;
net.params(f).weightDecay = 1 ;
%% Add skip layers on top of pool4
net.addLayer('skip4', ...
dagnn.Conv('size', [1 1 512 2], 'pad', 0), ...
'x23', 'x40', {'skip4f','skip4b'});
f = net.getParamIndex('skip4f') ;
net.params(f).value = zeros(1, 1, 512, 2, 'single') ;
net.params(f).learningRate = 0.1 ;
net.params(f).weightDecay = 1 ;
f = net.getParamIndex('skip4b') ;
net.params(f).value = zeros(1, 1, 2, 'single') ;
net.params(f).learningRate = 2 ;
net.params(f).weightDecay = 1 ;
%% Add skip layers on top of pool3
net.addLayer('skip3', ...
dagnn.Conv('size', [1 1 256 2]), ...
'x16', 'x41', {'skip3f','skip3b'});
f = net.getParamIndex('skip3f') ;
net.params(f).value = zeros(1, 1, 256, 2, 'single') ;
net.params(f).learningRate = 0.01 ;
net.params(f).weightDecay = 1 ;
f = net.getParamIndex('skip3b') ;
net.params(f).value = zeros(1, 1, 2, 'single') ;
net.params(f).learningRate = 2 ;
net.params(f).weightDecay = 1 ;
%%%%%%%%%C8럽얩생%%%%%%
filters_8_f = single(bilinear_u(64, 1, 2)) ;
net.addLayer('deconv_c8_full', ...
dagnn.ConvTranspose(...
'size', size(filters_8_f), ...
'upsample', 32, ...
'crop', [16 16 16 16], ...
'hasBias', false, ...
'opts', net.meta.cudnnOpts), ...
'x38', 'x42', 'deconv_c8_full') ;
f = net.getParamIndex('deconv_c8_full') ;
net.params(f).value = filters_8_f ;
net.params(f).learningRate = 0 ;
net.params(f).weightDecay = 1 ;
filters_8_5 = single(bilinear_u(4, 2, 2)) ;
net.addLayer('deconv_c8_dc5', ...
dagnn.ConvTranspose(...
'size', size(filters_8_5), ...
'upsample', 2, ...
'crop', 1, ...
'numGroups', 2, ...
'hasBias', false, ...
'opts', net.meta.cudnnOpts), ...
'x38', 'x43', 'deconv_c8_dc5') ;
f = net.getParamIndex('deconv_c8_dc5') ;
net.params(f).value = filters_8_5 ;
net.params(f).learningRate = 0 ;
net.params(f).weightDecay = 1 ;
filters_8_4 = single(bilinear_u(8, 2, 2)) ;
net.addLayer('deconv_c8_dc4', ...
dagnn.ConvTranspose(...
'size', size(filters_8_4), ...
'upsample', 4, ...
'crop', 2, ...
'numGroups', 2, ...
'hasBias', false, ...
'opts', net.meta.cudnnOpts), ...
'x38', 'x44', 'deconv_c8_dc4') ;
f = net.getParamIndex('deconv_c8_dc4') ;
net.params(f).value = filters_8_4 ;
net.params(f).learningRate = 0 ;
net.params(f).weightDecay = 1 ;
filters_8_3 = single(bilinear_u(16, 2, 2)) ;
net.addLayer('deconv_c8_dc3', ...
dagnn.ConvTranspose(...
'size', size(filters_8_3), ...
'upsample', 8, ...
'crop', 4, ...
'numGroups', 2, ...
'hasBias', false, ...
'opts', net.meta.cudnnOpts), ...
'x38', 'x45', 'deconv_c8_dc3') ;
f = net.getParamIndex('deconv_c8_dc3') ;
net.params(f).value = filters_8_3 ;
net.params(f).learningRate = 0 ;
net.params(f).weightDecay = 1 ;
%%%%%%%%%C5럽얩생%%%%%%
filters_5_4 = single(bilinear_u(4, 2, 2)) ;
net.addLayer('deconv_c5_dc4', ...
dagnn.ConvTranspose(...
'size', size(filters_5_4), ...
'upsample', 2, ...
'crop', 1, ...
'numGroups', 2, ...
'hasBias', false, ...
'opts', net.meta.cudnnOpts), ...
'x39', 'x46', 'deconv_c5_dc4') ;
f = net.getParamIndex('deconv_c5_dc4') ;
net.params(f).value = filters_5_4 ;
net.params(f).learningRate = 0 ;
net.params(f).weightDecay = 1 ;
filters_5_3 = single(bilinear_u(8, 2, 2)) ;
net.addLayer('deconv_c5_dc3', ...
dagnn.ConvTranspose(...
'size', size(filters_5_3), ...
'upsample', 4, ...
'crop', 2, ...
'numGroups', 2, ...
'hasBias', false, ...
'opts', net.meta.cudnnOpts), ...
'x39', 'x47', 'deconv_c5_dc3') ;
f = net.getParamIndex('deconv_c5_dc3') ;
net.params(f).value = filters_5_3 ;
net.params(f).learningRate = 0 ;
net.params(f).weightDecay = 1 ;
net.addLayer('sum1', dagnn.Sum(), {'x43', 'x39'}, 'x48') ;
filters_5_f = single(bilinear_u(32, 1, 2)) ;
net.addLayer('deconv_c5_full', ...
dagnn.ConvTranspose(...
'size', size(filters_5_f), ...
'upsample', 16, ...
'crop', [8 8 8 8], ...
'hasBias', false, ...
'opts', net.meta.cudnnOpts), ...
'x48', 'x49', 'deconv_c5_full') ;
f = net.getParamIndex('deconv_c5_full') ;
net.params(f).value = filters_5_f ;
net.params(f).learningRate = 0 ;
net.params(f).weightDecay = 1 ;
%%%%%%%%%C4럽얩생%%%%%%
filters_4_3 = single(bilinear_u(4, 2, 2)) ;
net.addLayer('deconv_c4_dc3', ...
dagnn.ConvTranspose(...
'size', size(filters_4_3), ...
'upsample', 2, ...
'crop', 1, ...
'numGroups', 2, ...
'hasBias', false, ...
'opts', net.meta.cudnnOpts), ...
'x40', 'x50', 'deconv_c4_dc3') ;
f = net.getParamIndex('deconv_c4_dc3') ;
net.params(f).value = filters_4_3 ;
net.params(f).learningRate = 0 ;
net.params(f).weightDecay = 1 ;
net.addLayer('sum2', dagnn.Sum(), {'x46', 'x44','x40'}, 'x51') ;
filters_4_f = single(bilinear_u(16, 1, 2)) ;
net.addLayer('deconv_c4_full', ...
dagnn.ConvTranspose(...
'size', size(filters_4_f), ...
'upsample', 8, ...
'crop', [4 4 4 4], ...
'hasBias', false, ...
'opts', net.meta.cudnnOpts), ...
'x51', 'x52', 'deconv_c4_full') ;
f = net.getParamIndex('deconv_c4_full') ;
net.params(f).value = filters_4_f ;
net.params(f).learningRate = 0 ;
net.params(f).weightDecay = 1 ;
%%%%%%%%%C3얩생휨북%%%%%%
net.addLayer('sum3', dagnn.Sum(), {'x50','x47','x45','x41'}, 'x53') ;
filters_3_f = single(bilinear_u(8, 1, 2)) ;
net.addLayer('deconv_c3_full', ...
dagnn.ConvTranspose(...
'size', size(filters_3_f), ...
'upsample', 4, ...
'crop', [2 2 2 2], ...
'hasBias', false, ...
'opts', net.meta.cudnnOpts), ...
'x53', 'x54', 'deconv_c3_full') ;
f = net.getParamIndex('deconv_c3_full') ;
net.params(f).value = filters_3_f ;
net.params(f).learningRate = 0 ;
net.params(f).weightDecay = 1 ;
%%%%%%%%%C3-C8얩생휨북%%%%%%
net.addLayer('sumdeconv', dagnn.Sum(), {'x54','x52','x49','x42'}, 'prediction') ;
net.addLayer('objective', ...
SegmentationLoss('loss', 'softmaxlog'), ...
{'prediction', 'label'}, 'objective') ;
% Add accuracy layer
net.addLayer('accuracy', ...
SegmentationAccuracy(), ...
{'prediction', 'label'}, 'accuracy') ;
% Make the output of the bilinear interpolator is not discared for
% visualization purposes
net.vars(net.getVarIndex('prediction')).precious = 1 ;
end