-
Notifications
You must be signed in to change notification settings - Fork 0
/
pitch_lstm_traindev.m
309 lines (293 loc) · 8.2 KB
/
pitch_lstm_traindev.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
298
299
300
301
302
303
304
305
306
307
308
309
%lstm trainer along with dev
itermax=10;
tot_length=1033; %1000;
%initialize params
%bsize=1;
hparams;
clip=0;
lr_0=-1; %learning rate
t_max=100;
d=100; %hidden size
alp=1; %sigmoid alpha
issoftmax=0;
islogloss=0;
nx=106;
ny=48;
load names
load names1
names=[names;names1];
%initialize network
W=randn(4*d,nx+d)*(1/d);
Wc=W(1:d,1:nx);
Wi=W(d+1:2*d,1:nx);
Wf=W(2*d+1:3*d,1:nx);
Wo=W(3*d+1:4*d,1:nx);
Uc=W(1:d,nx+1:nx+d);
Ui=W(d+1:2*d,nx+1:nx+d);
Uf=W(2*d+1:3*d,nx+1:nx+d);
Uo=W(3*d+1:4*d,nx+1:nx+d);
Why=randn(ny,d)*(1/d);
mW=zeros(size(W));
mWhy=zeros(size(Why));
trains=1:tot_length; % 1:2:1000 for train-dev
ltrain=length(trains);
devs=1:10:1000;
ldev=length(devs);
%se=zeros(itermax*length(trains),1);
se_train=zeros(ltrain,1);
se_dev=zeros(ldev,1);
se_large=zeros(itermax,2);
perm=trains(randperm(ltrain));
%perm=ones(size(perm)); %for instance only
en=0;
for iter=1:itermax
lr=lr_0*min(2/iter,iter/2);
perm=trains(randperm(ltrain));
en=0;
for dataid=1:ltrain
%lr=lr_0*min(1400/en,en/1400);
%initialize input
load(['D:\atlas\pitch\mirres\',names{perm(dataid)},'.mat']);
load(['D:\atlas\pitch\mirpv\',names{perm(dataid)},'.mat']);
sres=size(allres);
spv=size(pv);
% nx=sres(1);
% ny=spv(1);
lres=sres(2);
lpv=spv(2);
if (lres>lpv)
allres=allres(:,1:lpv);
lres=lpv;
end
if (lres<lpv)
pv=pv(:,1:lres);
lpv=lres;
end
%tmax=max(sx(2),sy(2))+1;
%x=[zeros(nx,1),x];
%yt=[zeros(ny,1),yt];
bsize=ceil(lres/t_max);
sn=0;
en=en+1;
for b=1:bsize
%initialize input here
p=(b-1)*t_max+1;
q=min(b*t_max,lres);
tmax=q-p+2;
x=[zeros(nx,1),allres(:,p:q)];
yt=[zeros(ny,1),pv(:,p:q)];
% x, yt, tmax
h=randn(d,1);
c=zeros(d,1);
%initialize recorder
A=zeros(d,tmax);
I=zeros(d,tmax);
F=zeros(d,tmax);
O=zeros(d,tmax);
C=zeros(d,tmax);
H=zeros(d,tmax);
H(:,1)=h;
Y=zeros(ny,tmax);
E=zeros(1,tmax);
% Forward
for t=2:tmax
%forward
a=tanh(Wc*x(:,t)+Uc*h);
i=sigmf(Wi*x(:,t)+Ui*h,[alp,0]);
f=sigmf(Wf*x(:,t)+Uf*h,[alp,0]);
o=sigmf(Wo*x(:,t)+Uo*h,[alp,0]);
c=i.*a+f.*c;
h=o.*tanh(c);
y=Why*h;
if (issoftmax==1)
y=y-max(y);
y=exp(y);
y=y/sum(y);
end
if (islogloss==1)
e=-sum(yt(:,t).*log(y));
else
e=sum((y-yt(:,t)).^2);
end
%record
A(:,t)=a;
I(:,t)=i;
F(:,t)=f;
O(:,t)=o;
C(:,t)=c;
H(:,t)=h;
Y(:,t)=y;
E(t)=e;
se_train(en)=se_train(en)+e;
sn=sn+1;
end
% Backword
dc1=zeros(size(h));
dh1=zeros(size(h));
c1=C(:,tmax);
h1=H(:,tmax);
for t=tmax:-1:2
%extract record
a=A(:,t);
i=I(:,t);
f=F(:,t);
o=O(:,t);
c=c1;
h=h1;
y=Y(:,t);
c1=C(:,t-1);
h1=H(:,t-1);
J=[x(:,t);h1];
%backward part 1
if (islogloss==1)
dy=-yt(:,t)./y;
else
dy=y-yt(:,t);
end
if (issoftmax==1)
Ja=diag(y)-y*y';
dy=Ja*dy;
end
dWhy=dy*h';
dh=dh1+Why'*dy;
do=dh.*tanh(c);
dc=dc1+dh.*o.*(1-(tanh(c)).^2);
di=dc.*a;
da=dc.*i;
df=dc.*c1;
dc1=dc.*f;
%backward part 2
da2=da.*(1-(tanh(a)).^2);
di2=di.*i.*(1-i);
df2=df.*f.*(1-f);
do2=do.*o.*(1-o);
dz=[da2;di2;df2;do2];
dJ=W'*dz;
dh1=dJ(nx+1:nx+d);
dW=dz*J';
%clip
if (clip>0)
dW=dW.*(abs(dW)<clip)+clip*(dW>=clip)-clip*(dW<=-clip);
dWhy=dWhy.*(abs(dWhy)<clip)+clip*(dWhy>=clip)-clip*(dWhy<=-clip);
end
%record
mW=mW+dW;
mWhy=mWhy+dWhy;
end
end
%update
W=W+lr*mW/sn;
Why=Why+lr*mWhy/sn;
mW=zeros(size(W));
mWhy=zeros(size(Why));
Wc=W(1:d,1:nx);
Wi=W(d+1:2*d,1:nx);
Wf=W(2*d+1:3*d,1:nx);
Wo=W(3*d+1:4*d,1:nx);
Uc=W(1:d,nx+1:nx+d);
Ui=W(d+1:2*d,nx+1:nx+d);
Uf=W(2*d+1:3*d,nx+1:nx+d);
Uo=W(3*d+1:4*d,nx+1:nx+d);
se_train(en)=se_train(en)/sn;
clc
iter
dataid
end
se_large(iter,1)=mean(se_train);
save(['savs/pitch_lstm_',num2str(iter),'.mat'],'Wc','Wi','Wf','Wo',...
'Uc','Ui','Uf','Uo','Why','issoftmax','islogloss','alp','se_large');
% DDDDD EEEEEE VV VV
% DD DD EE VV VV
% DD DD EEEEEE VV VV
% DD DD EE VVVV
% DDDDD EEEEEE VV
perm=devs(randperm(ldev));
en=0;
for dataid=1:ldev
%initialize input
load(['D:\atlas\pitch\mirres\',names{perm(dataid)},'.mat']);
load(['D:\atlas\pitch\mirpv\',names{perm(dataid)},'.mat']);
sres=size(allres);
spv=size(pv);
% nx=sres(1);
% ny=spv(1);
lres=sres(2);
lpv=spv(2);
if (lres>lpv)
allres=allres(:,1:lpv);
lres=lpv;
end
if (lres<lpv)
pv=pv(:,1:lres);
lpv=lres;
end
%tmax=max(sx(2),sy(2))+1;
%x=[zeros(nx,1),x];
%yt=[zeros(ny,1),yt];
sn=0;
en=en+1;
%initialize input here
tmax=lres+1;
x=[zeros(nx,1),allres];
yt=[zeros(ny,1),pv];
% x, yt, tmax
h=randn(d,1);
c=zeros(d,1);
%initialize recorder
A=zeros(d,tmax);
I=zeros(d,tmax);
F=zeros(d,tmax);
O=zeros(d,tmax);
C=zeros(d,tmax);
H=zeros(d,tmax);
H(:,1)=h;
Y=zeros(ny,tmax);
E=zeros(1,tmax);
% Forward
for t=2:tmax
%forward
a=tanh(Wc*x(:,t)+Uc*h);
i=sigmf(Wi*x(:,t)+Ui*h,[alp,0]);
f=sigmf(Wf*x(:,t)+Uf*h,[alp,0]);
o=sigmf(Wo*x(:,t)+Uo*h,[alp,0]);
c=i.*a+f.*c;
h=o.*tanh(c);
y=Why*h;
if (issoftmax==1)
y=y-max(y);
y=exp(y);
y=y/sum(y);
end
if (islogloss==1)
e=-sum(yt(:,t).*log(y));
else
e=sum((y-yt(:,t)).^2);
end
%record
A(:,t)=a;
I(:,t)=i;
F(:,t)=f;
O(:,t)=o;
C(:,t)=c;
H(:,t)=h;
Y(:,t)=y;
E(t)=e;
se_dev(en)=se_dev(en)+e;
sn=sn+1;
end
% No Backword
% No update
se_dev(en)=se_dev(en)/sn;
clc
iter
ltrain+dataid
end
se_large(iter,2)=mean(se_dev);
plot(1:iter,log(se_large(1:iter,:)));
xlabel('songs trained');
ylabel('log error');
legend('train','dev','Location','NorthEast');
drawnow;
end
save('pitch_lstm.mat','Wc','Wi','Wf','Wo','Uc','Ui','Uf','Uo','Why',...
'issoftmax','islogloss','alp','se_large');