Skip to content

Commit 9032b29

Browse files
adugnagadugnag
authored
Add files via upload
1 parent 695c916 commit 9032b29

File tree

3 files changed

+757
-0
lines changed

3 files changed

+757
-0
lines changed

‎Train_deSpeckNet_DAG.m

+51
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,51 @@
1+
%author: Adugna Mullissa
2+
3+
%Description: This script runs deSpeckNet training task on single polarization SAR intensity images.
4+
% The input image and the label should be stored in the same folder as this script. The default names are
5+
%Train2 and Test2. This script trains deSpeckNet without TV loss.
6+
% This script is modified from https://github.com/cszn/DnCNN.
7+
8+
clc;
9+
clear all
10+
11+
rng('default')
12+
13+
addpath('utilities');
14+
addpath('../../matlab/matlab');
15+
%-------------------------------------------------------------------------
16+
% Configuration
17+
%-------------------------------------------------------------------------
18+
opts.modelName = 'deSspeckNet'; % model name
19+
opts.learningRate = [logspace(-3,-3,25) logspace(-4,-4,25)];% learning rate
20+
opts.batchSize = 128; %
21+
opts.gpus = 1; %set to [] when using CPU
22+
opts.numSubBatches = 2;
23+
24+
% solver
25+
opts.solver = 'Adam'; % global
26+
opts.derOutputs = {'objective',100, 'objective1',1} ; %Loss weights for Lclean and Lnoisy respectively
27+
28+
opts.backPropDepth = Inf;
29+
%-------------------------------------------------------------------------
30+
% Initialize model
31+
%-------------------------------------------------------------------------
32+
net = deSpeckNet_Init();
33+
%-------------------------------------------------------------------------
34+
% Train
35+
%-------------------------------------------------------------------------
36+
37+
[net, info] = deSpecknet_train_dag(net, ...
38+
'learningRate',opts.learningRate, ...
39+
'derOutputs',opts.derOutputs, ...
40+
'numSubBatches',opts.numSubBatches, ...
41+
'backPropDepth',opts.backPropDepth, ...
42+
'solver',opts.solver, ...
43+
'batchSize', opts.batchSize, ...
44+
'modelname', opts.modelName, ...
45+
'gpus',opts.gpus) ;
46+
47+
48+
49+
50+
51+

‎deSpeckNet_Init.m

+171
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,171 @@
1+
function net = deSpeckNet_Init()
2+
3+
% Create DAGNN object
4+
net = dagnn.DagNN();
5+
6+
% conv + relu
7+
blockNum = 1;
8+
inVar = 'input';
9+
channel= 1; % grayscale image
10+
dims = [3,3,channel,64];
11+
pad = [1,1];
12+
dilate = [1,1];
13+
stride = [1,1];
14+
lr = [1,1];
15+
%FCN clean
16+
[net, inVar1, blockNum] = addConv(net, blockNum, inVar, dims, pad,dilate, stride, lr);
17+
[net, inVar1, blockNum] = addReLU(net, blockNum, inVar1);
18+
19+
for i = 1:15
20+
% conv + bn + relu
21+
dims0 = [3,3,64,64];
22+
[net, inVar1, blockNum] = addConv(net, blockNum, inVar1, dims0, pad,dilate, stride, lr);
23+
n_ch = dims0(4);
24+
[net, inVar1, blockNum] = addBnorm(net, blockNum, inVar1, n_ch);
25+
[net, inVar1, blockNum] = addReLU(net, blockNum, inVar1);
26+
end
27+
28+
% conv
29+
dims1 = [3,3,64,channel];
30+
[net, inVar5, blockNum] = addConv(net, blockNum, inVar1, dims1, pad,dilate, stride, lr);
31+
32+
%__________________________________________________________________________
33+
%FCN noise
34+
35+
[net, inVar8, blockNum] = addConv(net, blockNum, inVar, dims, pad,dilate, stride, lr);
36+
[net, inVar8, blockNum] = addReLU(net, blockNum, inVar8);
37+
38+
for i = 1:15
39+
% conv + bn + relu
40+
[net, inVar8, blockNum] = addConv(net, blockNum, inVar8, dims0, pad,dilate, stride, lr);
41+
n_ch = dims0(4);
42+
[net, inVar8, blockNum] = addBnorm(net, blockNum, inVar8, n_ch);
43+
[net, inVar8, blockNum] = addReLU(net, blockNum, inVar8);
44+
end
45+
46+
% conv
47+
[net, inVar13, blockNum] = addConv(net, blockNum, inVar8, dims1, pad,dilate, stride, lr);
48+
49+
50+
% % % Multiply and reconstruct noisy image
51+
inVarr = {inVar13,inVar5};
52+
[net, inVar30, blockNum] = addMultiply(net, blockNum, inVarr);
53+
%
54+
55+
outputName = 'prediction';
56+
net.renameVar(inVar5,outputName)
57+
58+
%__________________________________________________________________________
59+
60+
61+
% loss clean
62+
net.addLayer('loss', dagnn.Loss('loss','L2'), {'prediction','label'}, {'objective'},{});
63+
net.vars(net.getVarIndex('prediction')).precious = 1;
64+
65+
66+
outputName1 = 'prediction1'; %Final noisy image reconstruction
67+
net.renameVar(inVar30,outputName1)
68+
69+
% loss noisy
70+
net.addLayer('loss1', dagnn.Loss('loss','L2'), {'prediction1','input'}, {'objective1'},{});
71+
net.vars(net.getVarIndex('prediction1')).precious = 1;
72+
73+
end
74+
75+
76+
% Add a multiply layer
77+
function [net, inVar, blockNum] = addMultiply(net, blockNum, inVar)
78+
79+
outVar = sprintf('mult%d', blockNum);
80+
layerCur = sprintf('mult%d', blockNum);
81+
82+
block = dagnn.Multiply();
83+
net.addLayer(layerCur, block, inVar, {outVar},{});
84+
85+
inVar = outVar;
86+
blockNum = blockNum + 1;
87+
end
88+
89+
90+
% Add a relu layer
91+
function [net, inVar, blockNum] = addReLU(net, blockNum, inVar)
92+
93+
outVar = sprintf('relu%d', blockNum);
94+
layerCur = sprintf('relu%d', blockNum);
95+
96+
block = dagnn.ReLU('leak',0);
97+
net.addLayer(layerCur, block, {inVar}, {outVar},{});
98+
99+
inVar = outVar;
100+
blockNum = blockNum + 1;
101+
end
102+
103+
104+
% Add a bnorm layer
105+
function [net, inVar, blockNum] = addBnorm(net, blockNum, inVar, n_ch)
106+
107+
trainMethod = 'adam';
108+
outVar = sprintf('bnorm%d', blockNum);
109+
layerCur = sprintf('bnorm%d', blockNum);
110+
111+
params={[layerCur '_g'], [layerCur '_b'], [layerCur '_m']};
112+
net.addLayer(layerCur, dagnn.BatchNorm('numChannels', n_ch), {inVar}, {outVar},params) ;
113+
114+
pidx = net.getParamIndex({[layerCur '_g'], [layerCur '_b'], [layerCur '_m']});
115+
b_min = 0.025;
116+
net.params(pidx(1)).value = clipping(sqrt(2/(9*n_ch))*randn(n_ch,1,'single'),b_min);
117+
net.params(pidx(1)).learningRate= 1;
118+
net.params(pidx(1)).weightDecay = 0;
119+
net.params(pidx(1)).trainMethod = trainMethod;
120+
121+
net.params(pidx(2)).value = zeros(n_ch, 1, 'single');
122+
net.params(pidx(2)).learningRate= 1;
123+
net.params(pidx(2)).weightDecay = 0;
124+
net.params(pidx(2)).trainMethod = trainMethod;
125+
126+
net.params(pidx(3)).value = [zeros(n_ch,1,'single'), 0.01*ones(n_ch,1,'single')];
127+
net.params(pidx(3)).learningRate= 1;
128+
net.params(pidx(3)).weightDecay = 0;
129+
net.params(pidx(3)).trainMethod = 'average';
130+
131+
inVar = outVar;
132+
blockNum = blockNum + 1;
133+
end
134+
135+
136+
% add a Conv layer
137+
function [net, inVar, blockNum] = addConv(net, blockNum, inVar, dims, pad, dilate, stride, lr)
138+
opts.cudnnWorkspaceLimit = 1024*1024*1024*2; % 2GB
139+
convOpts = {'CudnnWorkspaceLimit', opts.cudnnWorkspaceLimit} ;
140+
trainMethod = 'adam';
141+
142+
outVar = sprintf('conv%d', blockNum);
143+
layerCur = sprintf('conv%d', blockNum);
144+
145+
convBlock = dagnn.Conv('size', dims, 'pad', pad, 'dilate', dilate, 'stride', stride, ...
146+
'hasBias', true, 'opts', convOpts);
147+
148+
net.addLayer(layerCur, convBlock, {inVar}, {outVar},{[layerCur '_f'], [layerCur '_b']});
149+
150+
f = net.getParamIndex([layerCur '_f']) ;
151+
sc = sqrt(2/(dims(1)*dims(2)*max(dims(3), dims(4)))) ; %improved Xavier
152+
net.params(f).value = sc*randn(dims, 'single') ;
153+
net.params(f).learningRate = lr(1);
154+
net.params(f).weightDecay = 1;
155+
net.params(f).trainMethod = trainMethod;
156+
157+
f = net.getParamIndex([layerCur '_b']) ;
158+
net.params(f).value = zeros(dims(4), 1, 'single');
159+
net.params(f).learningRate = lr(2);
160+
net.params(f).weightDecay = 1;
161+
net.params(f).trainMethod = trainMethod;
162+
163+
inVar = outVar;
164+
blockNum = blockNum + 1;
165+
end
166+
167+
168+
function A = clipping(A,b)
169+
A(A>=0&A<b) = b;
170+
A(A<0&A>-b) = -b;
171+
end

0 commit comments

Comments
 (0)