Cunder ⚡ ⚡

Experimental C binding for LibTorch

Quick start

Usage

see playground/main.cpp

#include <cunder/c_libtorch.h>
#include <iostream>

int main()
{
	// create tensor
	float tensor_data[] = {1, 9, 1};
	int tensor_data_shape[] = {3};
	auto cunder_data_tensor = cunder_tensor_from_data_wrap(1, tensor_data_shape, tensor_data, Cunder_DType::Cunder_Float32);
	
	// print the tensor
	printf("Cunder Tensor external data: \n");
	cunder_tensor_print(cunder_data_tensor);
	cunder_tensor_print_attributes(cunder_data_tensor);

	// access tensor data
	auto *tensor_accessor = cunder_tensor_accessor_f32(cunder_data_tensor);
	for (size_t i = 0; i < 3; i++)
		std::cout << tensor_accessor[i] << ',';
	std::cout << std::endl;

	// free the tensor
	cunder_tensor_free(cunder_data_tensor);
	return 0;
}

Structure

NOTE: add LibTorch dlls and include files in external folder OR use CMake find(torch) and add argument DCMAKE_PREFIX_PATH

Roadmap

• torch::version()
• Create tensor from data
  • given shape create zero tensor
  • given some data (pointer) with shape create tensor
    • wrapping the data
    • cloning the data
  • Supported types:
    • bool
    • uint8
    • int8
    • int32 (int)
    • int64 (long)
    • float32
    • float64
  • access tensor data
• Torch script jit model
  • Module struct torch::jit::Module
  • load Module torch::jit::load()
  • call eval() on Module
  • run Module on cpu (call forward() with tensors)
• Add support to external libraries:
  • torch_sparse
  • torch_scatter