-
Notifications
You must be signed in to change notification settings - Fork 0
/
localSearch2_NLSM.m
214 lines (196 loc) · 9.68 KB
/
localSearch2_NLSM.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
function [bestModel, bestNnet] = localSearch2_NLSM(ds, initModel, initNnet)
rng('default');
algOptions = struct('debug', 0, 'loss', 'logistic', 'regul', 1, 'algo', 'pm');
% dataORG.X = ds.trainData.X;
dataORG.X = [ds.trainData.X; ones(1, size(ds.trainData.X,2))];
dataORG.Y = ds.trainData.Y;
dataORG.T = ds.trainData.T;
% dataORG.X_test = ds.evalData.X;
dataORG.X_test = [ds.evalData.X; ones(1, size(ds.evalData.X,2))];
dataORG.Y_test = ds.evalData.Y;
dataORG.T_test = ds.evalData.T;
[DimX, ~] = size(dataORG.X);
K = max(dataORG.T);
% use random search in case an initial network is not provided
if ~exist('initNet', 'var') || ~exist('initModel', 'var') || isempty(initNet) || isempty(initModel)
bestNnet = struct('srad', 1e20);
bestModel = struct('acc_cv', 0);
for t = 1:150
data = dataORG;
param = randi(3);
if param==1; tu = 0; else; tu = 1; end;
if param==2; tv = 0; else; tv = 1; end;
if strcmpi(ds.name, 'cancer_dataset')
n2 = 1+randi(5); n1 = 1+randi(5);
alpha = 1+2*rand(n1, 1); beta = 1+2*rand(n2, 1);
rhou=rand*rand; rhov=rand*rand; rhow=rand*rand;
elseif strcmpi(ds.name, 'blood')
n2 = 1+randi(5); n1 = 1+randi(5);
alpha = 1+2*rand(n1, 1); beta = 1+2*rand(n2, 1);
rhou=rand*rand; rhov=1; rhow=rand;
elseif strcmpi(ds.name, 'haberman')
n2 = 1+randi(10); n1 = 1+randi(10);
alpha = 1+3*rand(n1, 1); beta = 1+3*rand(n2, 1);
rhou=rand*rand; rhov=1; rhow=rand;
elseif strcmpi(ds.name, 'seeds')
n2 = 1+randi(5); n1 = 1+randi(5);
alpha = 1+3*rand(n1, 1); beta = 1+3*rand(n2, 1);
rhou=rand; rhov=rand; rhow=rand;
elseif strcmpi(ds.name, 'pima')
n2 = 1+randi(10);
n1 = 1+randi(10);
alpha = 1+3*rand(n1, 1);
beta = 1+3*rand(n2, 1);
rhou=rand*rand; rhov=rand*rand; rhow=rand;
else
n2 = 1+randi(5); n1 = 1+randi(5);
alpha = 1+2*rand(n1, 1); beta = 1+2*rand(n2, 1);
rhou=rand*rand; rhov=rand*rand; rhow=rand;
end
ma = max(alpha);
mb = max(beta);
mia = min(alpha);
psi = rhou.^(mia/1e4)*n1^(1-mia/1e4);
thetaW = rhow*norm((rhov*psi).^beta, 1);
thetaV = rhow*norm(beta.*((rhov*psi).^beta), 1);
thetaU = rhow*ma*norm(beta.*((rhov*psi).^beta), 1);
pu = round( 20 + 2*n2*(2*thetaW+1) + 2*(2*thetaV+mb) + 2*(2*thetaU-2+ma+mb) ) + randi(50);
pv = round( 20 + 2*n2*(2*thetaW+1) + 2*(2*thetaV-1+mb) + 2*(2*thetaU+mb) ) + randi(50);
pw = round( 20 + 4*n2*thetaW + 2*(2*thetaV+1) + 2*(2*thetaU+ma) ) + randi(50);
% build network
nnet = struct('nLayers', 4, 'Vec', ones(K, 1), 'AGparam', param);
nnet.layers{1} = struct('name', 'input', 'normFact', 1, 'nUnits', DimX);
nnet.layers{2} = struct('name', 'sparse', 'normFact', 1, 'nUnits', n1, 'normType', tu, 'pNorm', pu, 'rho', rhou, 'alpha', alpha);
nnet.layers{3} = struct('name', 'sparse', 'normFact', 1, 'nUnits', n2, 'normType', tv, 'pNorm', pv, 'rho', rhov, 'alpha', beta);
nnet.layers{4} = struct('name', 'full', 'normFact', 1, 'nUnits', K, 'normType', 1, 'pNorm', pw, 'rho', rhow);
for l = 2:nnet.nLayers-1
mask = ones(nnet.layers{l}.nUnits, nnet.layers{l-1}.nUnits);
check = 0;
while(~check)
for k = 1:size(mask, 2)
check2 = 0;
while(~check2)
mask(:, k) = double(sprand(size(mask,1), 1, 0.95)>0);
if(sum(mask(:,k))>0), check2 = 1; end
end
end
if(min(sum(mask))>0), if(min(sum(mask'))>0); check=1; end, end
end
nnet.layers{l}.mask = mask;
end
nnet.rhox = 1;
puprime = nnet.layers{2}.pNorm/(nnet.layers{2}.pNorm-1);
rhox = max(sum(abs(data.X).^puprime) .^ (1/puprime));
data.X = nnet.rhox * data.X ./ rhox;
data.X_test = nnet.rhox * data.X_test ./ rhox;
acc_cv = 0;
for fold = 1:max(ds.cvid)
trainID = ds.cvid~=fold;
testID = ds.cvid==fold;
data_cv.X = data.X(:, trainID);
data_cv.Y = data.Y(:, trainID);
data_cv.T = data.T(trainID);
data_cv.X_test = data.X(:, testID);
data_cv.Y_test = data.Y(:, testID);
data_cv.T_test = data.T(testID);
model_cv = train_NLSM(nnet, data_cv, algOptions);
acc_cv = acc_cv + model_cv.trainAcc(end);
end
acc_cv = acc_cv/max(ds.cvid);
model = train_NLSM(nnet, data, algOptions);
model.acc_cv = acc_cv;
[nnet.srad, ~] = Compute_srad_2(nnet.layers{4}.pNorm, nnet.layers{3}.pNorm, nnet.layers{2}.pNorm, nnet.layers{4}.rho, nnet.layers{3}.rho, nnet.layers{2}.rho, nnet.rhox, nnet.layers{nnet.nLayers}.nUnits, nnet.layers{2}.alpha, nnet.layers{3}.alpha, nnet.AGparam);
if bestNnet.srad < 1
if nnet.srad < 1 && bestModel.acc_cv < model.acc_cv
bestNnet = nnet; bestModel = model;
end
elseif nnet.srad < 1 || bestModel.acc_cv < model.acc_cv || ((bestModel.acc_cv == model.acc_cv && nnet.srad<bestNnet.srad))
bestNnet = nnet; bestModel = model;
end
disp(['randomSearch2 ',num2str(t), ': trainAcc ',num2str(model.trainAcc(end)), ', testAcc ', num2str(bestModel.testAcc(end)), ', srad ', num2str(nnet.srad), ', bestSrad ', num2str(bestNnet.srad)]);
end
else
bestModel = initModel; bestNnet = initNnet;
end
% build data for optimal model found by random search
data = dataORG;
puprime = bestNnet.layers{2}.pNorm/(bestNnet.layers{2}.pNorm-1);
rhox = max(sum(abs(data.X).^puprime) .^ (1/puprime));
data.X = bestNnet.rhox * data.X ./ rhox;
data.X_test = bestNnet.rhox * data.X_test ./ rhox;
% local search
max_runs = 5; noimprove = max_runs;
while noimprove > 0
% check alpha
for t = 1:100
nnet = bestNnet;
for l = 2:nnet.nLayers-1
nnet.layers{l}.alpha = max(1.1, nnet.layers{l}.alpha+rand*rand*rand*randn(size(nnet.layers{l}.alpha)) );
end
acc_cv = 0;
for fold = 1:max(ds.cvid)
trainID = ds.cvid~=fold;
testID = ds.cvid==fold;
data_cv.X = data.X(:, trainID);
data_cv.Y = data.Y(:, trainID);
data_cv.T = data.T(trainID);
data_cv.X_test = data.X(:, testID);
data_cv.Y_test = data.Y(:, testID);
data_cv.T_test = data.T(testID);
model_cv = train_NLSM(nnet, data_cv, algOptions);
acc_cv = acc_cv + model_cv.trainAcc(end);
end
acc_cv = acc_cv/max(ds.cvid);
model = train_NLSM(nnet, data, algOptions);
model.acc_cv = acc_cv;
[nnet.srad, ~] = Compute_srad_2(nnet.layers{4}.pNorm, nnet.layers{3}.pNorm, nnet.layers{2}.pNorm, nnet.layers{4}.rho, nnet.layers{3}.rho, nnet.layers{2}.rho, nnet.rhox, nnet.layers{nnet.nLayers}.nUnits, nnet.layers{2}.alpha, nnet.layers{3}.alpha, nnet.AGparam);
if bestNnet.srad < 1
if nnet.srad < 1 && bestModel.acc_cv < model.acc_cv
bestNnet = nnet; bestModel = model; noimprove = max_runs;
end
elseif bestModel.acc_cv < model.acc_cv || ((bestModel.acc_cv == model.acc_cv && nnet.srad<bestNnet.srad))
bestNnet = nnet; bestModel = model; noimprove = max_runs;
elseif nnet.srad < 1
bestNnet = nnet; bestModel = model;
end
disp(['localSearch2 ',num2str(t), ': trainAcc ',num2str(model.trainAcc(end)), ', testAcc ', num2str(bestModel.testAcc(end)), ', srad: ', num2str(nnet.srad), ', bestSrad ', num2str(bestNnet.srad)]);
end
% check rho
[f1, f2] = ndgrid([0.9,0.95,0.975,1.025,1.05,1.1]);
f1 = f1(:);
f2 = f2(:);
% f1 = [0.8,0.9,0.95,0.975,1.025,1.05,1.1];
for t = 1:numel(f1)
nnet = bestNnet;
nnet.layers{2}.rho = bestNnet.layers{2}.rho * f1(t);
nnet.layers{3}.rho = bestNnet.layers{3}.rho * f2(t);
acc_cv = 0;
for fold = 1:max(ds.cvid)
trainID = ds.cvid~=fold;
testID = ds.cvid==fold;
data_cv.X = data.X(:, trainID);
data_cv.Y = data.Y(:, trainID);
data_cv.T = data.T(trainID);
data_cv.X_test = data.X(:, testID);
data_cv.Y_test = data.Y(:, testID);
data_cv.T_test = data.T(testID);
model_cv = train_NLSM(nnet, data_cv, algOptions);
acc_cv = acc_cv + model_cv.trainAcc(end);
end
acc_cv = acc_cv/max(ds.cvid);
model = train_NLSM(nnet, data, algOptions);
model.acc_cv = acc_cv;
[nnet.srad, ~] = Compute_srad_2(nnet.layers{4}.pNorm, nnet.layers{3}.pNorm, nnet.layers{2}.pNorm, nnet.layers{4}.rho, nnet.layers{3}.rho, nnet.layers{2}.rho, nnet.rhox, nnet.layers{nnet.nLayers}.nUnits, nnet.layers{2}.alpha, nnet.layers{3}.alpha, nnet.AGparam);
if bestNnet.srad < 1
if nnet.srad < 1 && bestModel.acc_cv < model.acc_cv
bestNnet = nnet; bestModel = model; noimprove = max_runs;
end
elseif bestModel.acc_cv < model.acc_cv || ((bestModel.acc_cv == model.acc_cv && nnet.srad<bestNnet.srad))
bestNnet = nnet; bestModel = model; noimprove = max_runs;
elseif nnet.srad < 1
bestNnet = nnet; bestModel = model;
end
disp(['localSearch2 ',num2str(t), ': trainAcc ',num2str(model.trainAcc(end)), ', testAcc ', num2str(bestModel.testAcc(end)), ', srad: ', num2str(nnet.srad), ', bestSrad ', num2str(bestNnet.srad)]);
end
noimprove = noimprove - 1;
end