void SpatialConvolution::init(TorchData& input) {
if (input.type() != TorchDataType::TENSOR_DATA) {
throw std::runtime_error("SpatialConvolution::init() - "
"FloatTensor expected!");
}
Tensor<float>& in = (Tensor<float>&)input;
if (in.dim()[2] != feats_in_) {
throw std::runtime_error("SpatialConvolution::init() - ERROR: "
"incorrect number of input features!");
}
if (output != NULL) {
Int3 out_dim(in.dim());
out_dim[0] = out_dim[0] - filt_width_ + 1;
out_dim[1] = out_dim[1] - filt_height_ + 1;
out_dim[2] = feats_out_;
if (!Int3::equal(out_dim, ((Tensor<float>*)output)->dim())) {
// Input dimension has changed!
SAFE_DELETE(output);
}
}
if (output == NULL) {
Int3 out_dim(in.dim());
out_dim[0] = out_dim[0] - filt_width_ + 1;
out_dim[1] = out_dim[1] - filt_height_ + 1;
out_dim[2] = feats_out_;
output = new Tensor<float>(out_dim);
//cl_context->getOptimalLocalWorkgroupSizes(deviceid,
// ((Tensor<float>*)output)->dim(), local_worgroup_size);
}
}
void Parallel::forwardProp(TorchData& input) {
if (input.type() != TorchDataType::TABLE_DATA) {
throw std::runtime_error("Parallel::forwardProp() - "
"Table expected!");
}
Table& in = (Table&)input;
if (in.tableSize() != network_->size()) {
throw std::runtime_error("Parallel::forwardProp() - ERROR: "
"Table size does not match number of parallel stages!");
}
for (uint32_t i = 0; i < network_->size(); i++) {
(*network_)[i]->forwardProp(*in(i));
}
initOutput(); // Init output just copies the pointers from the output
// of all the parallel stages and fills up a table with them
}
void SpatialSubtractiveNormalization::init(TorchData& input) {
if (input.type() != TorchDataType::TENSOR_DATA) {
throw std::runtime_error("SpatialSubtractiveNormalization::init() - "
"FloatTensor expected!");
}
Tensor<float>& in = (Tensor<float>&)input;
if (in.dim() != 3) {
throw std::runtime_error("SpatialDivisiveNormalization::init() - "
"3D input is expected!");
}
if (output != NULL) {
if (!in.isSameSizeAs(*(Tensor<float>*)output)) {
// Input dimension has changed!
cleanup();
}
}
if (output == NULL) {
output = new Tensor<float>(in.dim(), in.size());
mean_pass1_ = new Tensor<float>(in.dim(), in.size());
mean_pass2_ = new Tensor<float>(in.dim(), in.size());
}
if (mean_coef_ == NULL) {
uint32_t mean_coeff_size[2];
mean_coeff_size[0] = TO_TENSOR_PTR(output)->size()[0];
mean_coeff_size[1] = TO_TENSOR_PTR(output)->size()[1];
mean_coef_ = new Tensor<float>(2, mean_coeff_size);
float* mean_coef_cpu = new float[mean_coef_->nelems()];
float* kernel_cpu = new float[kernel_->nelems()];
kernel_->getData(kernel_cpu);
bool onedim_kernel = kernel_->dim() == 1;
// Filter an image of all 1 values to create the normalization constants
// See norm_test.lua for proof that this works as well as:
// https://github.com/andresy/torch/blob/master/extra/nn/SpatialSubtractiveNormalization.lua
int32_t n_feats = TO_TENSOR_PTR(output)->size()[2];
int32_t height = TO_TENSOR_PTR(output)->size()[1];
int32_t width = TO_TENSOR_PTR(output)->size()[0];
if (onedim_kernel) {
// 1D case - The filter is seperable, but we'll just do the dumb 2D
// version since we only do this once on startup. --> O(n * m)
uint32_t kernel_size = kernel_->size()[0];
int32_t filt_rad = (kernel_size - 1) / 2;
for (int32_t v = 0; v < height; v++) {
for (int32_t u = 0; u < width; u++) {
float tmp = 0.0f;
for (int32_t v_filt = -filt_rad; v_filt <= filt_rad; v_filt++) {
for (int32_t u_filt = -filt_rad; u_filt <= filt_rad; u_filt++) {
int32_t u_in = u + u_filt;
int32_t v_in = v + v_filt;
if (u_in >= 0 && u_in < width && v_in >= 0 && v_in < height) {
// Pixel is inside --> We'll effectively clamp zeros elsewhere.
tmp +=
(kernel_cpu[v_filt + filt_rad] * kernel_cpu[u_filt + filt_rad]);
}
}
}
mean_coef_cpu[v * width + u] = tmp / n_feats;
}
}
} else {
// 2D case
int32_t kernel_size_u = kernel_->size()[0];
int32_t kernel_size_v = kernel_->size()[1];
int32_t filt_rad_u = (kernel_size_u - 1) / 2;
int32_t filt_rad_v = (kernel_size_v - 1) / 2;
for (int32_t v = 0; v < height; v++) {
for (int32_t u = 0; u < width; u++) {
float tmp = 0.0f;
for (int32_t v_filt = -filt_rad_v; v_filt <= filt_rad_v; v_filt++) {
for (int32_t u_filt = -filt_rad_u; u_filt <= filt_rad_u; u_filt++) {
int32_t u_in = u + u_filt;
int32_t v_in = v + v_filt;
if (u_in >= 0 && u_in < width && v_in >= 0 && v_in < height) {
// Pixel is inside --> We'll effectively clamp zeros elsewhere.
tmp +=
kernel_cpu[(v_filt + filt_rad_v) * kernel_size_u + (u_filt + filt_rad_u)];
}
}
}
mean_coef_cpu[v * width + u] = tmp / n_feats;
}
}
}
mean_coef_->setData(mean_coef_cpu);
delete[] mean_coef_cpu;
delete[] kernel_cpu;
}
if (mean_ == NULL) {
uint32_t mean_coeff_size[2];
mean_coeff_size[0] = TO_TENSOR_PTR(output)->size()[0];
mean_coeff_size[1] = TO_TENSOR_PTR(output)->size()[1];
mean_ = new Tensor<float>(2, mean_coeff_size);
}
}
void JoinTable::init(TorchData& input) {
if (input.type() != TorchDataType::TABLE_DATA) {
throw std::runtime_error("JoinTable::forwardProp() - "
"Table expected!");
}
Table& in = (Table&)input;
if (in.tableSize() == 0) {
throw std::runtime_error("JoinTable::forwardProp() - "
"Empty input Table!");
}
// Check that it is a table of FloatTensors
for (uint32_t i = 0; i < in.tableSize(); i++) {
if (in(i)->type() != TENSOR_DATA) {
throw std::runtime_error("JoinTable::forwardProp() - "
"Table of float tensors expected!");
}
}
uint32_t dim = TO_TENSOR_PTR(in(0))->dim();
if (dim <= dimension_) {
throw std::runtime_error("JoinTable::forwardProp() - "
"Input is smaller than join dimension!");
}
uint32_t jdim = dim - dimension_ - 1; // dimension_=0 is the top dim
// Make sure the dimensions OTHER than the join dimension are all the same
for (uint32_t d = 0; d < dim; d++) {
if (d != jdim) {
for (uint32_t j = 1; j < in.tableSize(); j++) {
if (TO_TENSOR_PTR(in(j))->size()[d] != TO_TENSOR_PTR(in(0))->size()[d]) {
throw std::runtime_error("JoinTable::forwardProp() - "
"Size mismatch!");
}
}
if (output != NULL && TO_TENSOR_PTR(output)->size()[d] !=
TO_TENSOR_PTR(in(0))->size()[d]) {
SAFE_DELETE(output);
}
}
}
uint32_t nelems_jdim = 0;
for (uint32_t j = 1; j < in.tableSize(); j++) {
nelems_jdim += TO_TENSOR_PTR(in(j))->size()[jdim];
}
if (output != NULL &&
TO_TENSOR_PTR(output)->size()[jdim] != nelems_jdim) {
SAFE_DELETE(output);
}
if (output == NULL) {
uint32_t* size = new uint32_t[dim];
memcpy(size, TO_TENSOR_PTR(in(0))->size(), sizeof(size[0]) * dim);
size[dimension_] = nelems_jdim;
output = new Tensor<float>(dim, size);
SAFE_DELETE_ARR(size);
}
}
