-
Notifications
You must be signed in to change notification settings - Fork 142
/
Copy pathbatch_cell_lstm.m
executable file
·241 lines (201 loc) · 8.65 KB
/
batch_cell_lstm.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
function [inerr,dw,inLL,right ]=batch_cell_lstm(W,dataall,maskall )
% gradient threshold for bptt need done ,inLL,right
global mzeros convert in_size gate_size out_size share_size numMmcell usegpu...
share_size2 in ingate cellstate cells outgate globalData globalMask ...
node_outgateInit cellinInit node_cellbiasInit delta_outInit cellstatusInit signum
if nargin<=1
data=globalData;
mask=globalMask;
else
data=dataall;
mask=maskall;
end
if usegpu
fun=@sigmoidnGpu;
delta_fun=@delta_sigmoidnGpu;
activation=@activationGpu;
deactivation=@deactivationGpu;
else
fun=@sigmoidnCpu;
delta_fun=@delta_sigmoidnCpu;
activation=@activationCpu;
deactivation=@deactivationCpu;
end
actNum1=convert(1);
actNum2=2*actNum1;
numsamples=size(data,1) ;
dw= zeros(1 ,2*share_size2 +share_size2*numMmcell + (gate_size*numMmcell+1)*out_size);
[ Wingate,Wcell , Woutgate, Wout]=unpack(W);
[dwingate,dwcell,dwoutgate,dwout] = unpack2(dw);
dwingate=convert(dwingate);
dwcell=convert(dwcell);
dwoutgate=convert(dwoutgate);
dwout=convert(dwout);
right=zeros(numsamples,1,'int32');
inLL = 0;
inerr = zeros( numsamples,out_size);
timespan = size( data , 3 );
node_tmpindex=cell(1,timespan);
node_tmpindex2=cell(1,timespan);
node_tmpindex3=cell(1,timespan);
node_outgate= node_outgateInit;
cellin=cellinInit;
node_cellbias=node_cellbiasInit;
delta_out=delta_outInit;
cellstatus= cellstatusInit;
node_ingate = node_outgate;
node_cell = cellin;
Y_cellout = cellin;
delta_outgate=node_outgate;
errorstate=cellin;
delta_cellin=cellin;
delta_ingate=node_outgate;
Wingate_in=convert( Wingate(in,:) );
Wingate_ingate= convert( Wingate(ingate,:) );
Wingate_cellstate=convert( Wingate(cellstate,:) );
Wingate_cell=convert( Wingate(cells,:) );
Wingate_outgate=convert( Wingate(outgate,:) );
% to cellin
Wcell_in=convert( Wcell(in,:) ) ;
Wcell_ingate=convert( Wcell(ingate,:) );
Wcell_cell=convert( Wcell(cells,:) );
Wcell_outgate=convert( Wcell(outgate,:) );
% to outgate
Woutgate_in=convert( Woutgate(in,:) );
Woutgate_ingate=convert( Woutgate(ingate,:) );
Woutgate_cellstate=convert( Woutgate(cellstate,:) );
Woutgate_cell = convert( Woutgate(cells,:) );
Woutgate_outgate=convert( Woutgate(outgate,:) );
t=2;
while t<=timespan
output = data(:,1:out_size,t ) ;
if ~isempty(mask)
masko=mask( : , 1 , t ) ;
elseif isempty(mask)
masko=1;
end
% forward pass
node_in = data(:,1:in_size,t-1);
node_cellbias{t } = data(:,in_size+1,t-1);
% to ingate
tmpp=node_in*Wingate_in + node_outgate{t-1}*Wingate_outgate +...
cellstatus{t-1} * Wingate_cellstate + node_cell{t-1}*Wingate_cell + node_ingate{t-1} * Wingate_ingate;
node_ingate{t}=fun( tmpp ,signum);
tmpp=node_in*Wcell_in + node_ingate{t-1} * Wcell_ingate +...
+ node_cell{t-1}*Wcell_cell +node_outgate{t-1} * Wcell_outgate ;
cellin{t} = activation( tmpp,actNum2 ) ;
cellstatus{t} = cellstatus{t-1} + cellin{t } .* repmat(node_ingate{t},1,numMmcell ) ;
Y_cellout{t}= activation( cellstatus{t } ,actNum1 ) ;
tmpp=node_in*Woutgate_in + node_outgate{t-1} *Woutgate_outgate +...
cellstatus{t} * Woutgate_cellstate + node_cell{t-1}*Woutgate_cell + node_ingate{t-1} * Woutgate_ingate;
node_outgate{t } = fun(tmpp ,signum);
node_cell{t } = repmat(node_outgate{t},1,numMmcell ) .* Y_cellout{t } ;
node_tmpindex{t } = [ node_in node_ingate{t-1} cellstatus{t-1} node_cell{t-1} node_outgate{t-1} ];
node_tmpindex2{t} = [ node_in node_ingate{t-1} cellstatus{t} node_cell{t-1} node_outgate{t-1} ];
node_tmpindex3{t } = [ node_in node_ingate{t-1} node_cell{t-1} node_outgate{t-1} ];
if any(masko)
node_out = fun( [ node_cell{t} node_cellbias{t} ] * Wout ,signum ) ;
delta_out{t} = masko .* ( - output + node_out ) .* delta_fun(node_out,signum);
inerr = inerr + ( masko .* (output - node_out )).^2 ;
% right = right + rightfun(masko,output,node_out) ;
end
t=t+1;
end
t=t-1;
% just cell
tmp_outgate=repmat(node_outgate{t},1,numMmcell);
tmp_ingate=repmat(node_ingate{t},1,numMmcell);
errorcell = delta_out{t} * Wout(1:end-1,:)';
% just cell to outgate
delta_outgate{t} = squeezing(delta_fun( tmp_outgate,signum) .* errorcell .* Y_cellout{t}) ;
% peephole
errorstate{t} = errorcell .* tmp_outgate .* deactivation(Y_cellout{t},actNum1)+...
delta_outgate{t} * Woutgate_cellstate' ;
delta_cellin{t} =tmp_ingate .* errorstate{t}.* deactivation(cellin{t},actNum2);
delta_ingate{t} = squeezing(cellin{t} .* errorstate{t} .* delta_fun(tmp_ingate,signum));
for t = timespan-1:-1:2
% just cell
tmp_outgate=repmat(node_outgate{t},1,numMmcell);
tmp_ingate=repmat(node_ingate{t},1,numMmcell);
errorcell = delta_out{t } * Wout(1:end-1,:)' + delta_outgate{t+1} * Woutgate_cell' +...
delta_cellin{t+1} * Wcell_cell' + delta_ingate{t+1} * Wingate_cell';
% just cell to outgate
delta_outgate{t} = delta_fun( node_outgate{t},signum ) .* ( squeezing(errorcell .* Y_cellout{t}) +...
delta_ingate{t+1} * Wingate_outgate'+ delta_cellin{t+1} * Wcell_outgate'+...
delta_outgate{t+1} *Woutgate_outgate') ;
% peephole
errorstate{t} = errorcell .* tmp_outgate .* deactivation( Y_cellout{t} ,actNum1 ) + ...
errorstate{t+1} + delta_ingate{t+1}* Wingate_cellstate' +...
delta_outgate{t} * Woutgate_cellstate' ;
delta_cellin{t} = tmp_ingate .* deactivation( cellin{t } ,actNum2) .* errorstate{t} ;
delta_ingate{t} = delta_fun( node_ingate{t} ,signum).* ( squeezing(cellin{t} .* errorstate{t}) +...
+ delta_ingate{t+1} * Wingate_ingate' +delta_cellin{t+1} * Wcell_ingate' + ...
delta_outgate{t+1} * Woutgate_ingate' );
end
for t = 2:timespan
dwout = dwout + [node_cell{t} node_cellbias{t} ]' * delta_out{t} ;
dwoutgate =dwoutgate + node_tmpindex2{t}' * delta_outgate{t} ;
dwcell=dwcell+ node_tmpindex3{t}'*delta_cellin{t};
dwingate = dwingate + node_tmpindex{t}' * delta_ingate{t} ;
end
inerr = gather( 1/2* sum( inerr) /numsamples) ;% 1/2* for gradient checking
%right=gather(sum(right));
dw=pack2(dwingate,dwcell ,dwoutgate,dwout);
dw = gather(dw / numsamples );
function [Wingate,Wcell ,Woutgate,Wout] = unpack(W)
Wingate= reshape( W( 1:share_size2 ) , share_size ,gate_size );
Wcell = reshape( W( share_size2 +1 : share_size2 *(1+numMmcell) ), share_size, numMmcell*gate_size);
Woutgate = reshape( W( (1+numMmcell) * share_size2 +1 : (2+numMmcell)* share_size2 ), share_size , gate_size);
Wout = reshape( W( (2+numMmcell) * share_size2 +1 : end ) ,gate_size*numMmcell+1 , out_size);
end
function [Wingate,Wcell ,Woutgate,Wout] = unpack2(W)
Wingate= reshape( W( 1:share_size2 ) , share_size ,gate_size );
Wcell = reshape( W( share_size2 +1 : share_size2 *(1+numMmcell) ), share_size, numMmcell*gate_size);
Wcell = Wcell([1:in_size+gate_size (in_size+gate_size+gate_size*numMmcell+1):end],:) ;
Woutgate = reshape( W( (1+numMmcell) * share_size2 +1 : (2+numMmcell)* share_size2 ), share_size , gate_size);
Wout = reshape( W( (2+numMmcell) * share_size2 +1 : end ) ,gate_size*numMmcell+1 , out_size);
end
function [W]=pack2(Wingate,Wcell , Woutgate,Wout)
tmp=mzeros(share_size,gate_size*numMmcell);
tmp([1:in_size+gate_size (in_size+gate_size+gate_size*numMmcell+1):end],:)=Wcell;
Wcell=tmp;
W = [Wingate(:);Wcell(:) ;Woutgate(:);Wout(:)];
end
function y=sigmoidnGpu(x,num)
y=arrayfun(@(x,num)1./(1+exp(- num*x)),x,num);
end
function y=delta_sigmoidnGpu(x,num)
y=arrayfun(@(x,num)num*x.*(1-x),x,num);
end
function y=sigmoidnCpu(x,num)
y=1./(1+exp(- num*x));
end
function y=delta_sigmoidnCpu(x,num)
y= num*x.*(1-x);
end
function y = activationCpu(x,num)
y=num*2./(1+exp(-x))-num;
end
function y = deactivationCpu(x,num)
y=0.5/num*(num+x).*(num-x);
end
function y = activationGpu(x,num)
y= arrayfun(@(x,num)num*2./(1+exp(-x))-num,x,num);
end
function y = deactivationGpu(x,num)
y=arrayfun(@(x,num)0.5/num*(num+x).*(num-x),x,num);
end
function y = squeezing(x)
y=zeros(size(x,1),gate_size);
for i = 1:numMmcell
y = y + x(:,1+(i-1)*gate_size: i*gate_size);
end
% x=mat2cell(x,size(x,1),gate_size*ones(1,numMmcell));
%
% for i=2:numMmcell
% x{1}=x{1}+x{i};
% end
% y=x{1};
% y= cell2mat(cellfun(@plus ,x, 'UniformOutput',false));
end
end