-
Notifications
You must be signed in to change notification settings - Fork 142
/
Main.m
executable file
·190 lines (153 loc) · 6.5 KB
/
Main.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
function []=Main(result,speed,options,problem,sgd)
global mones mzeros convert in_size gate_size numMmcell globalData globalMask usegpu...
out_size share_size share_size2 in ingate cellstate cells outgate...
node_outgateInit cellinInit node_cellbiasInit delta_outInit cellstatusInit signum
signum = 3.5;
if problem.batchsize>problem.numsamples
problem.batchsize=problem.numsamples;
end
[mones,mzeros,convert,usegpu]=gputype(speed.usegpu);
[data,mask,test,masktest ]=feval(['gen' problem.name],problem);
problem.numsamples=size(data,1);
[in_size,gate_size,out_size,share_size,share_size2,numMmcell,W,in,ingate,cellstate,cells,outgate,...
node_outgateInit,cellinInit,node_cellbiasInit,delta_outInit,cellstatusInit ]=netInit(problem);
if problem.continualPredict
Bpfunc=@batch_equal_nomask_lstm;
else
Bpfunc=@batch_cell_lstm;
end
if speed.gradientchecking
for i=1:fix(problem.numsamples/problem.batchsize)
bashdata{i} = data( (i-1)*problem.batchsize + 1:i*problem.batchsize , : , : );
maskdata{i}= mask( (i-1)*problem.batchsize + 1:i*problem.batchsize , : , : );
end
globalData=bashdata{1};
globalMask=maskdata{1};
[ err,dw,inLL,right ]=Bpfunc(W);
[numgrad]=computeNumericalGradient(Bpfunc,W);
disp( [numgrad dw])
diff = norm(numgrad-dw)/norm(numgrad+dw);
disp(diff);
fprintf('Norm of the difference between numerical and analytical gradient (should be < 1e-9)\n\n');
return
end
if speed.usecluster
serveroptions.slavecount=speed.numcluster;
serveroptions.slavedir='/usr/local/anew-lstm';
server=Server(serveroptions);
markers = floor(linspace(1, size(data,2), server.slaveCount+1));
markers(end) = size(data,2);
markers(1) = 0;
trainStarts = cell(1,server.slaveCount);
trainEnds =cell(1,server.slaveCount);
for i=1:server.slaveCount
trainStarts{i} = markers(i)+1;
trainEnds{i} = markers(i+1);
fprintf('i = %d, trainStart = %d, trainEnd = %d\n', i, trainStarts{i}, trainEnds{i});
end
options.Method = 'lbfgs';
for i=1:fix(problem.numsamples/problem.batchsize/server.slaveCount )
bashdata{i} = data( (i-1)*problem.batchsize*server.slaveCount + 1:...
i*problem.batchsize*server.slaveCount , : , : );
maskdata{i}= mask( (i-1)*problem.batchsize*server.slaveCount + 1:...
i*problem.batchsize*server.slaveCount , : , : );
end
clear data
record=1;
timer=tic;
for epoch =1:1
for i=1:fix(problem.numsamples/problem.batchsize/server.slaveCount )
[Winit]= server.rpc('clientLoadDataMinibatchNomask_ref',...
trainStarts, trainEnds, problem.batchsize ,...
problem,speed,bashdata{i},maskdata{i},{1:server.slaveCount});
if i==1&&epoch==1
W=Winit{1};
end
for repeat = 1:32
[W, cost] = minFunc( @server_batch_cell_lstm,W,options,server);
if toc(timer)>1
timearray{1}( record )= toc(timer)
[errorarray{1}(record),~,~,rightarray{1}(record) ]=testmodel(W,test,masktest ,problem);
timer=tic;
record=record+1;
end
end
end
end
figure
for method = 1:1
plot(cumsum(timearray{method}), rightarray{method} , result.linspec{method}); hold on
end
legend('server ' );
xlabel('time(seconds)');
ylabel('test obj');
saveas(gcf,[ pwd '/picture/' num2str(problem.numsamples) '_' num2str(problem.Ttest) problem.name '.fig']);
saveas(gcf,[ pwd '/picture/' num2str(problem.numsamples) '_' num2str(problem.Ttest) problem.name '.eps'],'epsc');
eps2pdf([pwd '/picture/' num2str(problem.numsamples) '_' num2str(problem.Ttest) problem.name '.eps']);
save('server.mat','timearray','errorarray');
elseif ~speed.usecluster
for i=1:fix(problem.numsamples/problem.batchsize)
bashdata{i} = data( (i-1)*problem.batchsize + 1:i*problem.batchsize , : , : );
maskdata{i}= mask( (i-1)*problem.batchsize + 1:i*problem.batchsize , : , : );
end
Winit=W;
timer=tic;
record=1;
for i=1:fix(problem.numsamples/problem.batchsize)
for inrepeat=1:20
globalData = bashdata{i};
globalMask= maskdata{i};
options.Method = 'lbfgs';
[W, cost] = minFunc( Bpfunc,W,options);
if toc(timer)>2
timearray{1}( record )= toc(timer);
[errorarray{1}(record),~,~,rightarray{1}(record) ]=testmodel(W,test,masktest ,problem);
disp([ 'lbfgs error ' num2str(errorarray{1}(record) )] )
record=record+1;
timer=tic;
end
end
end
W=Winit;
momentum=sgd.momentum;
alpha=sgd.alpha;
oldGradient=0;
record =1;
timer=tic;
for repeat =1 :5
for inner =1:50
for i=1:fix(problem.numsamples/problem.batchsize)
globalData = bashdata{i};
globalMask= maskdata{i};
[~,dw,~,~]=Bpfunc(W);
oldGradient = alpha*dw + momentum* oldGradient;
W= W - alpha* oldGradient;
end
timearray{2}(record )= toc(timer);
[errorarray{2}(record),~,~, rightarray{2}(record) ]=testmodel(W,test,masktest ,problem);
record=record+1;
timer=tic;
end
disp([ 'sgd error ' num2str(errorarray{2}(record-1) )] )
end
figure
for method = 1:2
plot(cumsum(timearray{method}), rightarray{method} , result.linespec{method}); hold on
end
legend('lbfgs','sgd');
xlabel('time(seconds)');
ylabel('test obj');
figure
for method = 1:2
plot(cumsum(timearray{method}),errorarray{method} , result.linespec{method}); hold on
end
legend('lbfgs','sgd');
xlabel('time(seconds)');
ylabel('test e');
if result.savepic
saveas(gcf,[ pwd '/picture/' num2str(problem.numsamples) '_' num2str(problem.Ttest) problem.name '.fig']);
saveas(gcf,[ pwd '/picture/' num2str(problem.numsamples) '_' num2str(problem.Ttest) problem.name '.eps'],'epsc');
% eps2pdf([pwd '/picture/' num2str(problem.numsamples) '_' num2str(problem.Ttest) problem.name '.eps']);
end
end
end