THNETWORK *THLoadNetwork(const char *path)
{
char tmppath[255];
int i, longsize = 8;
THNETWORK *net = calloc(1, sizeof(*net));
sprintf(tmppath, "%s/model.net", path);
net->netobj = malloc(sizeof(*net->netobj));
lasterror = loadtorch(tmppath, net->netobj, longsize);
if(lasterror == ERR_CORRUPTED)
lasterror = loadtorch(tmppath, net->netobj, longsize = 4);
if(lasterror)
{
free(net->netobj);
free(net);
return 0;
}
if(th_debug)
printobject(net->netobj, 0);
net->net = Object2Network(net->netobj);
if(!net->net)
{
lasterror = ERR_WRONGOBJECT;
freeobject(net->netobj);
free(net->netobj);
free(net);
return 0;
}
net->std[0] = net->std[1] = net->std[2] = 1;
net->mean[0] = net->mean[1] = net->mean[2] = 0;
sprintf(tmppath, "%s/stat.t7", path);
net->statobj = malloc(sizeof(*net->statobj));
lasterror = loadtorch(tmppath, net->statobj, longsize);
if(!lasterror)
{
if(net->statobj->type != TYPE_TABLE || net->statobj->table->nelem != 2)
{
lasterror = ERR_WRONGOBJECT;
freenetwork(net->net);
freeobject(net->netobj);
free(net->netobj);
freeobject(net->statobj);
free(net->statobj);
free(net);
return 0;
}
for(i = 0; i < net->statobj->table->nelem; i++)
if(net->statobj->table->records[i].name.type == TYPE_STRING)
{
if(!strcmp(net->statobj->table->records[i].name.string.data, "mean"))
memcpy(net->mean, net->statobj->table->records[i].value.tensor->storage->data, sizeof(net->mean));
else if(!strcmp(net->statobj->table->records[i].name.string.data, "std"))
memcpy(net->std, net->statobj->table->records[i].value.tensor->storage->data, sizeof(net->std));
}
} else {
free(net->statobj);
net->statobj = 0;
}
THUseSpatialConvolutionMM(net, 2);
return net;
}
int main(int argc, char **argv)
{
THNETWORK *net;
float *result;
int i, n = 0, rc, outwidth, outheight, runs = 1, print = 0, alg = 1, nbatch = 1;
const char *modelsdir = 0, *inputfile = 0;
for(i = 1; i < argc; i++)
{
if(argv[i][0] != '-')
continue;
switch(argv[i][1])
{
case 'm':
if(i+1 < argc)
modelsdir = argv[++i];
break;
case 'i':
if(i+1 < argc)
inputfile = argv[++i];
break;
case 'a':
if(i+1 < argc)
alg = atoi(argv[++i]);
break;
case 'p':
print = 1;
break;
case 'r':
if(i+1 < argc)
runs = atoi(argv[++i]);
break;
case 'b':
if(i+1 < argc)
{
nbatch = atoi(argv[++i]);
if(nbatch > 256 || nbatch < 1)
nbatch = 256;
}
break;
}
}
if(!modelsdir || !inputfile)
{
fprintf(stderr, "Syntax: test -m <models directory> -i <input file>\n");
fprintf(stderr, " [-r <number of runs] [-p(rint results)]\n");
fprintf(stderr, " [-a <alg=0:norm,1:MM (default),2:virtMM,3:cuDNN,4:cudNNhalf>]\n");
fprintf(stderr, " [-b <nbatch>]\n");
return -1;
}
if(alg == 4)
{
alg = 3;
THCudaHalfFloat(1);
}
THInit();
net = THLoadNetwork(modelsdir);
if(net)
{
THMakeSpatial(net);
if(alg == 0)
THUseSpatialConvolutionMM(net, 0);
else if(alg == 1 || alg == 2)
THUseSpatialConvolutionMM(net, alg);
else if(alg == 3)
{
THNETWORK *net2 = THCreateCudaNetwork(net);
if(!net2)
THError("CUDA not compiled in");
THFreeNetwork(net);
net = net2;
}
if(strstr(inputfile, ".t7"))
{
struct thobject input_o;
rc = loadtorch(inputfile, &input_o, 8);
if(!rc)
{
THFloatTensor *in = THFloatTensor_newFromObject(&input_o);
// In CuDNN the first one has to do some initializations, so don't count it for timing
if(alg == 3)
THProcessFloat(net, in->storage->data, 1, in->size[2], in->size[1], &result, &outwidth, &outheight);
t = seconds();
for(i = 0; i < runs; i++)
n = THProcessFloat(net, in->storage->data, 1, in->size[2], in->size[1], &result, &outwidth, &outheight);
t = (seconds() - t) / runs;
THFloatTensor_free(in);
freeobject(&input_o);
} else printf("Error loading %s\n", inputfile);
} else {
img_t image;
rc = loadimage(inputfile, &image);
if(!rc)
{
unsigned char *bitmaps[256];
for(i = 0; i < nbatch; i++)
bitmaps[i] = image.bitmap;
// In CuDNN the first one has to do some initializations, so don't count it for timing
if(alg == 3)
THProcessImages(net, &image.bitmap, 1, image.width, image.height, 3*image.width, &result, &outwidth, &outheight, 0);
t = seconds();
for(i = 0; i < runs; i++)
n = THProcessImages(net, bitmaps, nbatch, image.width, image.height, 3*image.width, &result, &outwidth, &outheight, 0);
t = (seconds() - t) / runs;
#ifdef USECUDAHOSTALLOC
cudaFreeHost(image.bitmap);
#else
free(image.bitmap);
#endif
} else printf("Error loading image %s\n", inputfile);
}
if(print)
for(i = 0; i < n; i++)
printf("(%d,%d,%d): %f\n", i/(outwidth*outheight), i % (outwidth*outheight) / outwidth, i % outwidth, result[i]);
printf("1 run processing time: %lf\n", t);
THFreeNetwork(net);
} else printf("The network could not be loaded: %d\n", THLastError());
#ifdef MEMORYDEBUG
debug_memorydump(stderr);
#endif
return 0;
}