How to extract output of torch model in c++?

Question

I have got trained keras model and converted it using mmdnn. Then I try use it in c++ code:

#include <iostream>

#include <opencv2/core/core.hpp>
#include <opencv2/highgui/highgui.hpp>

#include <torch.h>

int main()
{
    cv::Mat image;
    image= cv::imread("test_img.png", cv::IMREAD_GRAYSCALE);   // Read the file

try
{
    torch::jit::script::Module module;
    module = torch::jit::load("my_model.pth");

    torch::IntArrayRef input_dim = std::vector<int64_t>({ 1, 2, 256, 256});

    cv::Mat input_img;
    image.convertTo(input_img, CV_32FC3, 1 / 255.0);
    torch::Tensor x = torch::from_blob(input_img.data, { 1, 2, 256, 256 }, torch::kFloat);
    torch::NoGradGuard no_grad;

    auto output = module.forward({ x });

    float* data = static_cast<float*>(output.toTensor().data_ptr());

    cv::Mat output_img = cv::Mat(256, 256, CV_32FC3, data);
    cv::imwrite("output_img.png", output_img);
}
catch (std::exception &ex)
{
    std::cout << "exception! " << ex.what() << std::endl;
}

    return 0;
}

This code throws an exception:

exception! isTensor() INTERNAL ASSERT FAILED at E:\20B\pytorch\pytorch\aten\src\ATen/core/ivalue_inl.h:112, please report a bug to PyTorch. Expected Tensor but got Tuple (toTensor at E:\20B\pytorch\pytorch\aten\src\ATen/core/ivalue_inl.h:112) (no backtrace available)

This was thrown in line float* data = static_cast<float*>(output.toTensor().data_ptr()); when the function toTensor() was called. If I use toTuple() instead of toTensor() then the result doesn't have the function data_ptr(), but I need this for extracting data (and putting it into opencv image).

How to extract image from the model output?

Answer 1

In this case the answer of model is tuple of 2 images. We can extract them by such way:

torch::Tensor t0 = output.toTuple()->elements()[0].toTensor();
torch::Tensor t1 = output.toTuple()->elements()[1].toTensor();

Variables t0 and t1 contain tensors with output of model.