int THProcessYUYV(THNETWORK *network, unsigned char *image, int width, int height, float **results, int *outwidth, int *outheight)
{
THFloatTensor *out;
THFloatStorage *st;
#ifdef CUDNN
if(network->net->cuda)
THError("This function is not supported with CUDNN");
#endif
#ifdef OPENCL
if(network->net->cuda)
THError("This function is not supported with OpenCL");
#endif
st = THFloatStorage_new(width * height * 3);
yuyv2fRGB(image, st->data, width*height, width, width, height, network->mean, network->std);
THFloatTensor *t = THFloatTensor_new();
t->storage = st;
t->nDimension = 3;
t->size[0] = 3;
t->size[1] = height;
t->size[2] = width;
t->stride[0] = width * height;
t->stride[1] = width;
t->stride[2] = 1;
out = forward(network->net, t);
THFloatTensor_free(t);
*results = out->storage->data;
if(out->nDimension >= 3)
{
*outwidth = out->size[out->nDimension - 1];
*outheight = out->size[out->nDimension - 2];
} else *outwidth = *outheight = 1;
return THFloatTensor_nElement(out);
}
int THProcessFloat(THNETWORK *network, float *data, int batchsize, int width, int height, float **result, int *outwidth, int *outheight)
{
int b, c, i;
THFloatTensor *t = THFloatTensor_new();
THFloatTensor *out;
t->nDimension = 4;
t->size[0] = batchsize;
t->size[1] = 3;
t->size[2] = height;
t->size[3] = width;
t->stride[0] = 3 * width * height;
t->stride[1] = width * height;
t->stride[2] = width;
t->stride[3] = 1;
t->storage = THFloatStorage_newwithbuffer((float *)data);
#pragma omp parallel for private(b, c, i)
for(b = 0; b < batchsize; b++)
for(c = 0; c < 3; c++)
for(i = 0; i < width*height; i++)
data[b * t->stride[0] + c * t->stride[1] + i] =
(data[b * t->stride[0] + c * t->stride[1] + i] - network->mean[c]) / network->std[c];
#ifdef CUDNN
if(network->net->cuda)
{
THFloatTensor *t2 = THCudaTensor_newFromFloatTensor(t);
out = forward(network->net, t2);
THFloatTensor_free(t2);
if(network->out)
THFloatTensor_free(network->out);
network->out = THFloatTensor_newFromCudaTensor(out);
out = network->out;
} else
#endif
#ifdef OPENCL
if(network->net->opencl)
{
THFloatTensor *t2 = THOpenCLTensor_newFromImageTensor(t);
out = forward(network->net, t2);
THFloatTensor_free(t2);
if(network->out)
THFloatTensor_free(network->out);
network->out = THFloatTensor_newFromOpenCLImageTensor(out);
out = network->out;
} else
#endif
out = forward(network->net, t);
THFloatTensor_free(t);
*result = out->storage->data;
if(out->nDimension >= 3)
{
*outwidth = out->size[out->nDimension - 1];
*outheight = out->size[out->nDimension - 2];
} else *outwidth = *outheight = 1;
return THFloatTensor_nElement(out);
}
THFloatTensor *THFloatTensor_newWithStorage1d(THFloatStorage *storage, long storageOffset, long size0, long stride0)
{
THFloatTensor *t = THFloatTensor_new();
t->nDimension = 1;
t->size[0] = size0;
t->stride[0] = stride0 == -1 ? 1 : stride0;
t->storage = storage;
t->storageOffset = storageOffset;
THAtomicIncrement(&t->storage->nref);
return t;
}
void THMakeSpatial(THNETWORK *network)
{
int i, size = 231, nInputPlane = 3;
for(i = 0; i < network->net->nelem; i++)
{
if(network->net->modules[i].type == MT_View || network->net->modules[i].type == MT_Reshape)
{
THFloatTensor_free(network->net->modules[i].output);
memmove(network->net->modules+i, network->net->modules+i+1, sizeof(*network->net->modules) * (network->net->nelem - i - 1));
network->net->nelem--;
i--;
} else if(network->net->modules[i].type == MT_Linear)
{
THFloatTensor_free(network->net->modules[i].Linear.addBuffer);
network->net->modules[i].updateOutput = nn_SpatialConvolutionMM_updateOutput;
#ifndef USEBLAS
network->net->modules[i].type = MT_SpatialConvolutionVirtMM;
#else
network->net->modules[i].type = MT_SpatialConvolutionMM;
#endif
struct SpatialConvolution *c = &network->net->modules[i].SpatialConvolution;
c->finput = THFloatTensor_new();
c->padW = c->padH = 0;
c->dW = c->dH = 1;
c->kW = c->kH = size;
c->nInputPlane = nInputPlane;
nInputPlane = c->nOutputPlane = c->weight->size[0];
size = (size + 2*c->padW - c->kW) / c->dW + 1;
} else if(network->net->modules[i].type == MT_SpatialConvolution ||
network->net->modules[i].type == MT_SpatialConvolutionMM ||
network->net->modules[i].type == MT_SpatialConvolutionVirtMM)
{
struct SpatialConvolution *c = &network->net->modules[i].SpatialConvolution;
size = (size + 2*c->padW - c->kW) / c->dW + 1;
nInputPlane = network->net->modules[i].SpatialConvolution.nOutputPlane;
} else if(network->net->modules[i].type == MT_SpatialMaxPooling)
{
struct SpatialMaxPooling *c = &network->net->modules[i].SpatialMaxPooling;
if(c->ceil_mode)
size = (long)(ceil((float)(size - c->kH + 2*c->padH) / c->dH)) + 1;
else size = (long)(floor((float)(size - c->kH + 2*c->padH) / c->dH)) + 1;
} else if(network->net->modules[i].type == MT_SpatialZeroPadding)
{
struct SpatialZeroPadding *c = &network->net->modules[i].SpatialZeroPadding;
size += c->pad_l + c->pad_r;
}
}
}
THFloatTensor *THFloatTensor_newWithStorage3d(THFloatStorage *storage, long storageOffset, long size0, long stride0, long size1, long stride1, long size2, long stride2)
{
THFloatTensor *t = THFloatTensor_new();
t->nDimension = 3;
t->size[0] = size0;
t->size[1] = size1;
t->size[2] = size2;
t->stride[0] = stride0 == -1 ? size1 * size2 : stride0;
t->stride[1] = stride1 == -1 ? size2 : stride1;
t->stride[2] = stride2 == -1 ? 1 : stride2;
t->storage = storage;
t->storageOffset = storageOffset;
THAtomicIncrement(&t->storage->nref);
return t;
}
THFloatTensor *forward(struct network *net, THFloatTensor *in)
{
int i;
double t = 0, convtot = 0, convflops = 0;
#ifdef OPENCL
if(net->opencl == 1)
OpenCL_Build(net, in);
#endif
for(i = 0; i < net->nelem; i++)
{
if(th_profile)
t = th_seconds();
in = net->modules[i].updateOutput(&net->modules[i], in);
// You can remove these lines if you don't have problems with memory
// These lines free intermediate results
if(i > 0)
{
THFloatTensor_free(net->modules[i-1].output);
net->modules[i-1].output = THFloatTensor_new();
}
if(th_profile)
{
#ifdef OPENCL
if(net->opencl)
clFinish(cl_queue);
#endif
t = th_seconds() - t;
if(net->modules[i].type == MT_SpatialConvolutionMM ||
net->modules[i].type == MT_SpatialConvolutionVirtMM ||
net->modules[i].type == MT_SpatialConvolution)
{
double flops = 2.0 * THFloatTensor_nElement(in) * net->modules[i].SpatialConvolution.nInputPlane *
net->modules[i].SpatialConvolution.kW * net->modules[i].SpatialConvolution.kH;
printf("%f seconds for module %d, %f Gflops/s\n", t, i+1, flops * 1e-9 / t);
convtot += t;
convflops += flops;
} else printf("%f seconds for module %d\n", t, i+1);
}
if(th_debug > 1)
printf("%d) %d %d %ld %ld %ld %ld\n", i+1, net->modules[i].type, in->nDimension, in->size[0], in->size[1], in->size[2], in->size[3]);
}
if(th_profile)
printf("%f seconds for convolutions %f Gflops/s\n", convtot, convflops * 1e-9 / convtot);
return in;
}
int THProcessImages(THNETWORK *network, unsigned char **images, int batchsize, int width, int height, int stride, float **results, int *outwidth, int *outheight, int bgr)
{
int i;
THFloatTensor *out, *t = 0;
THFloatStorage *st;
#ifdef CUDNN
if(network->net->cuda)
{
#ifdef HAVEFP16
if(floattype == CUDNN_DATA_HALF)
{
st = THCudaStorage_new(batchsize * (width * height * 3));
for(i = 0; i < batchsize; i++)
cuda_rgb2half((unsigned short *)st->data + i * (width * height * 3), images[i], width, height, stride, network->mean, network->std, bgr);
} else
#endif
{
st = THCudaStorage_new(batchsize * width * height * 3);
for(i = 0; i < batchsize; i++)
cuda_rgb2float(st->data + i * width * height * 3, images[i], width, height, stride, network->mean, network->std, bgr);
}
} else
#endif
#ifdef OPENCL
if(network->net->opencl)
t = OpenCL_LoadImage(images[0], width, height, stride, network->mean, network->std, bgr);
else
#endif
{
st = THFloatStorage_new(batchsize * width * height * 3);
if(bgr)
#pragma omp parallel for if(batchsize>1) private(i)
for(i = 0; i < batchsize; i++)
bgr2float(st->data + i * width * height * 3, images[i], width, height, stride, network->mean, network->std);
else
#pragma omp parallel for if(batchsize>1) private(i)
for(i = 0; i < batchsize; i++)
rgb2float(st->data + i * width * height * 3, images[i], width, height, stride, network->mean, network->std);
}
if(!t)
{
t = THFloatTensor_new();
t->storage = st;
if(batchsize == 1)
{
t->nDimension = 3;
t->size[0] = 3;
t->size[1] = height;
t->size[2] = width;
t->stride[0] = width * height;
t->stride[1] = width;
t->stride[2] = 1;
} else {
t->nDimension = 4;
t->size[0] = batchsize;
t->size[1] = 3;
t->size[2] = height;
t->size[3] = width;
t->stride[0] = 3 * width * height;
t->stride[1] = width * height;
t->stride[2] = width;
t->stride[3] = 1;
}
}
#ifdef CUDNN
if(network->net->cuda)
{
out = forward(network->net, t);
if(network->out)
THFloatTensor_free(network->out);
#ifdef HAVEFP16
if(floattype == CUDNN_DATA_HALF)
network->out = THFloatTensor_newFromHalfCudaTensor(out);
else
#endif
network->out = THFloatTensor_newFromCudaTensor(out);
out = network->out;
} else
#endif
#ifdef OPENCL
if(network->net->opencl)
{
out = forward(network->net, t);
if(network->out)
THFloatTensor_free(network->out);
#ifdef HAVEFP16
if(cl_datasize == 2)
network->out = THFloatTensor_newFromHalfOpenCLImageTensor(out);
else
#endif
network->out = THFloatTensor_newFromOpenCLImageTensor(out);
out = network->out;
} else
#endif
out = forward(network->net, t);
THFloatTensor_free(t);
*results = out->storage->data;
if(out->nDimension >= 3)
{
*outwidth = out->size[out->nDimension - 1];
*outheight = out->size[out->nDimension - 2];
} else *outwidth = *outheight = 1;
return THFloatTensor_nElement(out);
}
THFloatTensor *THFloatTensor_newWithTensor(THFloatTensor *tensor)
{
THFloatTensor *self = THFloatTensor_new();
THFloatTensor_set(self, tensor);
return self;
}
