Advice for debugging a TorchScript module from C++

C++
1

I have trained a model from python and now I’m trying to load it from a c++ program. Unfortunately it crashes on the line // crash here

#include <torch/script.h>
#include <iostream>
#include <vector>
#include "nets/nets.h"
#include "util/runfiles.h"

int main(int argc, char** argv) {
  std::cout << "Nets example" << std::endl;

  // Custom code that loads the module
  auto runfiles = MakeRunfiles(argv[0]);
  torch::jit::script::Module segnet3 = LoadSegNet3(*runfiles);
  std::cout << "Loaded SegNet3" << std::endl;

  // Make a fake image.
  torch::Tensor input = torch::randn({1, 3, 300, 300});
  std::cout << "Made random input" << std::endl;

  std::vector<torch::jit::IValue> inputs;
  inputs.push_back(input);

  // crash here
  torch::Tensor output = segnet3.forward(inputs).toTensor();
  std::cout << "Output done" << std::endl;
}

The output of the program is

Nets example
Loaded SegNet3
Made random input

and it exits without writing any error message.

What can I do to debug the situation? I am wondering if maybe my input tensor is of the wrong dimensionality. Is there a way to check the dimensions the network is expecting? And/or is there a way to print some debug information about the structure of the network?

2

A segfault should not happen with that program even when the shapes do not match.
Most often segfaults happen when people use from_blob and the memory-owner goes out of scope, but this is not case here.

  • Does this also happen with a small / trivial module? If not, there might be a specific operation that it does not like. Ideally you would be able to isolate it.
  • If it does even for a trivial module, you could compile libtorch / PyTorch in Debug mode and run it with gdb -ex run --args myprogram to get a backtrace (with bt after the segfault).

Best regards

Thomas

3

Thanks tom. It’s not so easy for me to use gdb since I’m on Windows, but I did happen to discover the issue. In my loading code, I had moved the model by calling .to(at::kCUDA) so the tensor needed to be moved too before.

4

Great you figured that out. Is it crashing as in “segfault” or does it nominally print an error message on the console?

Best regards

Thomas

5

There actually is no error message - it exits silently