如何从 libtorch 中的 Tensor 取回图像？

Question

I am trying to get the image back from the tensor I created earlier and visualize it, However, the resulting image is distorted/garbage it seems. This is how I convert a CV_8UC3 into the corresponding at::Tensor :

at::Tensor tensor_image = torch::from_blob(img.data, { img.rows, img.cols, 3 }, at::kByte);

and this is how I convert it back to the image:

auto ToCvImage(at::Tensor tensor)
{
    int width = tensor.sizes()[0];
    int height = tensor.sizes()[1];
    try
    {
        cv::Mat output_mat(cv::Size{ height, width }, CV_8UC3, tensor.data_ptr<int>());
        return output_mat.clone();
    }
    catch (const c10::Error& e)
    {
        std::cout << "an error has occured : " << e.msg() << std::endl;
    }
    return cv::Mat(height, width, CV_8UC3);
}

This is how the original image looks like:

and this is what I get after conversion:

Now if I use at::kInt instead of kByte during the creation of the tensor:

at::Tensor tensor_image = torch::from_blob(img.data, { img.rows, img.cols, 3 }, at::kByte);

I no longer get the distorted image, however, the network output will be off which means something has gone wrong in the input!

What's the issue here and how should I be going about this?

Answer 1

When the tensor was created using a c10::kByte for the cast we need to use uchar and not char or uint , etc. so in order to get this fixed I only had to use uchar instead of int :

auto ToCvImage(at::Tensor tensor)
{
    int width = tensor.sizes()[0];
    int height = tensor.sizes()[1];
    try
    {
        cv::Mat output_mat(cv::Size{ height, width }, CV_8UC3, tensor.data_ptr<uchar>());
        return output_mat.clone();
    }
    catch (const c10::Error& e)
    {
        std::cout << "an error has occured : " << e.msg() << std::endl;
    }
    return cv::Mat(height, width, CV_8UC3);
}

Side note: In case you created your Tensor with any other type, make sure to use Tensor::totype() effectively and convert to the proper type before hand. That is before I feed this tensor to my network, eg I convert it to KFloat and then carry on! its an obvious point that may very well be neglected and cost you hours of debugging!