WIP CUDA 11.7 support

[mightyCelu]

This branch contains changes to make Pedestron work with CUDA 11.7.

IMPORTANT: This is not mergeable, as it breaks compilation for lower versions of CUDA and PyTorch, and it is not a trivially usable version, but it works for me, so I figured I'd share it here for reference & potential improvement by others.

I was trying to get Pedestron running and after struggling with older CUDA versions and compatibility to Ampere GPUs I figured I'd give porting the Code to the newer version instead a go, which in the end worked.

Deployment

Environment

Run on Debian 11 with CUDA 11.7 and a RTX3060Ti.

I created a virutalenv for the dependencies.

package list

``` ⇒ pip3 freeze addict==2.4.0 certifi==2022.9.24 charset-normalizer==2.1.1 contourpy==1.0.5 cycler==0.11.0 Cython==0.29.32 fonttools==4.37.3 idna==3.4 kiwisolver==1.4.4 matplotlib==3.6.0 mmcv==0.2.10 mmdet==0.6.0+746cdce mmengine==0.1.0 numpy==1.23.3 opencv-python==4.6.0.66 packaging==21.3 Pillow==9.2.0 pycocotools==2.0.5 pyparsing==3.0.9 python-dateutil==2.8.2 PyYAML==6.0 requests==2.28.1 scipy==1.9.1 six==1.16.0 termcolor==2.0.1 terminaltables==3.1.10 torch==1.13.0.dev20220929+cu117 torchvision==0.14.0.dev20220929+cu117 typing-extensions==4.3.0 urllib3==1.26.12 yapf==0.32.0 ```

torch and torchvision had to be installed via pip install --pre torch torchvision -f https://download.pytorch.org/whl/nightly/cu117/torch_nightly.html, I read that the torch+cu117 syntax should work to, but it didn't for me.

Building

Activate the venv and run:

python3 setup.py install

As a workaround for pip3 install . which didn't work, as the build environment was different from the venv.

Execution

Run the demo as described in the project readme.

Hints for integration

This is a very rough changeset, little more than a proof of concept. If someone were to try and refine it into proper support for more versions of CUDA & PyTorch, here are some hints for how to do that.

I applied three groups of changes to make Pedestron work with my CUDA & PyTorch versions:

• Avoid circular imports: I moved file-level imports in Python to the respective functions.
• Replace AT_CHECK macro with TORCH_CHECK as described in #131
• Replace THC function family

along with some changes to the project setup.

The circular dependency issue may well be a non-issue due to the tools/demo.py import of the current working directory over the installed one which (I think) I removed later. The changes to CUDA & C++ code should be sanitized by providing headers which define appropriate functions/macros depending on the detected CUDA/PyTorch versions. Moreover, someone with actual understanding should validate that there are no sideeffects introduced by the changes. Lastly, I started changing the project description in pyproject.toml and setup.py to be more self-contained, currently these files don't reflect reality when it comes to supported versions, these should be explicitly encoded (e.g., there seems to be a PyTorch version beyond which the current code-base doesn't work, also I constrained mmcv to 0.2.10 because that worked for be, probably there's more versions that will work), s.t. the package can be more easily deloyed.

[hasanirtiza]

Appreciate the effort. I think for the moment it is best that, we either create a section in installation file, with your instructions or simply provide a link to this PR.

[mightyCelu]

Sure, I opened this PR for sharing/documentation only in case it helps anyone.