int test_step(nrex_search* s, int pos, int level) const
{
if (max == 0)
{
return pos;
}
if ((max >= 1 && level > max) || pos > s->end)
{
return -1;
}
if (!greedy && level > min)
{
int res = pos;
if (next)
{
res = next->test(s, res);
}
if (s->complete)
{
return res;
}
if (res >= 0 && parent->test_parent(s, res) >= 0)
{
return res;
}
}
int res = child->test(s, pos);
if (s->complete)
{
return res;
}
if (res >= 0)
{
int res_step = test_step(s, res, level + 1);
if (res_step >= 0)
{
return res_step;
}
else if (greedy && level >= min)
{
if (next)
{
res = next->test(s, res);
}
if (s->complete)
{
return res;
}
if (res >= 0 && parent->test_parent(s, res) >= 0)
{
return res;
}
}
}
return -1;
}
/* assumes weights have been stored already, and gradient is computed */
float sample(float lr)
{
float e1,e2,e3,e4;
/* move in the direction */
if (myid==0)
test_step(hearing,lr);
parallel_broadcast_weights(hearing);
e1=parallel_f(ps,hearing_examples,0.8);
e2=parallel_f(sp,speaking_examples,0.8);
e3=parallel_f(sem,sem_examples,0.2);
e4=parallel_f(phono,phono_examples,0.2);
return e1+e2+e3+e4;
}
int test(nrex_search* s, int pos) const
{
return test_step(s, pos, 1);
}
int main(int argc,char *argv[])
{
char cmd[255];
int hcount=0,scount=0;
float herr=0,serr=0,dice;
Example *ex;
int do_cg=0;
float epsilon;
int seed=1,first=1;
float e,e0,lr,lrPrev;
int start=1,i,j;
char loadFile[255],fn[255];
setbuf(stdout,NULL);
parallel_init(&argc,&argv);
if (myid==0)
{
announce_version();
system("hostname");
}
printf("pid %d\n",getpid());
loadFile[0]=0;
/* what are the command line arguments? */
for(i=1;i<argc;i++)
{
if (strcmp(argv[i],"-seed")==0)
{
seed=atoi(argv[i+1]);
i++;
}
else if (strcmp(argv[i],"-start")==0)
{
start=atoi(argv[i+1]);
i++;
}
else if (strncmp(argv[i],"-epsilon",5)==0)
{
epsilon=atof(argv[i+1]);
i++;
}
else if (strncmp(argv[i],"-load",5)==0)
{
strcpy(loadFile,argv[i+1]);
i++;
}
else
{
fprintf(stderr,"unknown argument: %s\n",argv[i]);
exit(1);
}
}
default_epsilon=0.05;
mikenet_set_seed(seed);
build_hearing_model(SAMPLES);
/* load in our example set */
phono_examples=load_examples("phono.pat",TICKS);
sem_examples=load_examples("sem.pat",TICKS);
hearing_examples=load_examples("ps.pat",TICKS);
speaking_examples=load_examples("sp.pat",TICKS);
phono_examples->numExamples=500;
sem_examples->numExamples=500;
hearing_examples->numExamples=500;
speaking_examples->numExamples=500;
myid=parallel_proc_id();
#ifdef DO_ONLINE_PRE_TRAIN
if (start==1)
{
for(i=1;i<=10000;i++)
{
dice = mikenet_random();
if (dice <=0.2)
{
ex=get_random_example(phono_examples);
crbp_forward(phono,ex);
crbp_compute_gradients(phono,ex);
crbp_apply_deltas(phono);
}
else if (dice <= 0.5)
{
ex=get_random_example(hearing_examples);
crbp_forward(ps,ex);
crbp_compute_gradients(ps,ex);
herr += compute_error(ps,ex);
crbp_apply_deltas(ps);
}
else if (dice <= 0.7)
{
ex=get_random_example(sem_examples);
crbp_forward(sem,ex);
crbp_compute_gradients(sem,ex);
crbp_apply_deltas(sem);
}
else
{
ex=get_random_example(speaking_examples);
crbp_forward(sp,ex);
crbp_compute_gradients(sp,ex);
serr+=compute_error(sp,ex);
crbp_apply_deltas(sp);
}
if (i % 100 == 0)
{
printf("%d hear %f speak %f\n",i,
herr/hcount,serr/scount);
herr=0.0;
serr=0.0;
hcount=0;
scount=0;
}
}
sprintf(fn,"s%d_online_weights",seed);
save_weights(hearing,fn);
}
#endif
parallel_broadcast_weights(hearing);
do_cg=1;
if (do_cg && myid==0)
printf("USING CG\n");
/* loop for ITER number of times */
for(i=1;i<5;i++)
{
parallel_sync();
store_all_weights(hearing);
e=parallel_g(ps,hearing_examples,0.8) +
parallel_g(sp,speaking_examples,0.8) +
parallel_g(sem,sem_examples,0.2) +
parallel_g(phono,phono_examples,0.2);
if(do_cg)
{
if (first)
init_cg(hearing);
else
cg(hearing);
first=0;
}
e0=e;
if (myid==0)
printf("%d e0: %f\n",i,e);
lr = 0.2/hearing_examples->numExamples;
lrPrev=lr/10;
for(j=1;j<10;j++)
{
e=sample(lr);
if (myid==0)
printf("\t\t%d %f %f\n",j,e,lr);
if (e>e0)
{
if (myid==0)
test_step(hearing,lrPrev);
parallel_broadcast_weights(hearing);
break;
}
e0=e;
lrPrev=lr;
lr *= 1.7;
}
zero_gradients(hearing);
if (i % 5==0 && myid==0)
{
sprintf(fn,"s%d_%d_weights",seed,i);
save_weights(hearing,fn);
}
}
parallel_finish();
return 0;
}
int main() {
test_setup();
while (true) {
test_step();
}
}