C++ TorchScript example

[Mika412]

So I've added a C++ TorchScript example. Unfortunately, I wasn't able to get it running with CUDA, only the CPU.

Steps to run:

1. wget https://download.pytorch.org/libtorch/nightly/cpu/libtorch-shared-with-deps-latest.zip
2. unzip libtorch-shared-with-deps-latest.zip
3. cd cpp_test
4. mkdir build && cd build
5. cmake -DCMAKE_PREFIX_PATH=/path/to/libtorch ..
6. cmake --build . --config Release
7. ./test_app ../../scripts/scripted_controller.pt

[NicolayP]

Will have a look at it. Do you have time metrics associated with it? :)

[Mika412]

Was just getting them.

C++ CPU:

Python GPU:

[NicolayP]

So we have a small improvement over onnx, this is still with k=20/h=5 right?

[Mika412]

The results are for k=100/h=10

[NicolayP]

After 3h27min i get this error. Seems very blurry to me, any clue?

[Mika412]

Ohh, damn 3h.... I don't think it's anything with the code per se. Most likely it's due to the Pytorch version. I'm using the latest nightly, 2.1.0.dev20230709+cu121.
But did it work with smaller params?

[NicolayP]

Okay will try to upgrade pytorch, I haven't tried compiling with smaller parameters, going to do that now :)

[NicolayP]

Hey can you update the requirments.txt file that you use? The current one points to torch 1.2.1. Would like to try with the same environment setup as yours

[Mika412]

Sorry about the delay. Too many meetings... I've updated the requirements.txt, but it may have some extra unnecessary dependencies, as I'm using this env for both projects.

[NicolayP]

And out of curiosity what version of pyhon and what version of linux are you running? I can't seem to find pytorch 2.1.x_dev. And I'm not sure if I just need to update python to a newer version or install pytorch from source

[Mika412]

I'm on Ubuntu 22.04, using Python 3.9.17 in a conda environment. You can just install the latest nightly pytorch