Opencv -> torch tensor conversion does not work

Dear Community,
I’m experimenting with the C++ API from pytorch. When loading an Image with opencv, I get an error when calling the modules forward function. Using the same logic but initializing as torch::tensor works perfectly. Can someone tell me what I’m doing wrong?

I think the code explains it best. I posted the code everyone can try below:

This is the torch-only version that does work:

int test_torch_model_torch_dummy_data() {
	torch::jit::script::Module module = torch::jit::load("/path/to/traced/model.pt");
	std::cout << "model loaded." << std::endl;

	at::TensorOptions options(at::ScalarType::Float);
	torch::Tensor tensor_image = torch::zeros({ 1, 3, 224, 224 }, options);
	std::cout << tensor_image.sizes() << std::endl;

	std::vector<torch::jit::IValue> inputs;
	inputs.emplace_back(tensor_image);

	std::cout << "input shape torch" << tensor_image.sizes() << std::endl;
	auto ret = module.forward(inputs).toTensor();
	std::cout << "works using torch dummy data.";
	return 0;
}

This is the opencv-function that does not work:

int test_torch_model_cv_dummy_data() {
	torch::jit::script::Module module = torch::jit::load("path/to/traced_model.pt");
	std::cout << "model loaded." << std::endl;

	auto bgr = cv::Mat(cv::Size(224, 224), CV_8UC3);
	std::vector<int64_t> sizes = { 1, 224, 224, 3 };
	at::TensorOptions options(at::ScalarType::Float);
	at::Tensor tensor_image = torch::from_blob(bgr.data, at::IntList(sizes), options);
	auto tensor_image_permuted = tensor_image.permute({ 0, 3, 1, 2 });

	std::cout << "input shape cv data" << tensor_image_permuted.sizes() << std::endl;

	std::vector<torch::jit::IValue> inputs;
	inputs.emplace_back(tensor_image);

	module.forward(inputs).toTensor();

	std::cout << "works using cv dummy data " << std::endl;
}
void main() {
	test_torch_model_torch_dummy_data();
	test_torch_model_cv_dummy_data();
}