Installing dependencies on the fly via the source code might not be best practice

[henriksod]

I noticed that dependencies are automatically installed via TestRequirements. I am not sure this is the best idea.

It compromises the users own package installation without their control which might not be desired.

I can see why something like this is needed though. We want to be able to dynamically set dependencies like onnxruntime and onnxruntime-gpu, as well as not install dependencies that might not be needed by the user (e.g. nvidia-tensorrt).

It seems like there are ongoing discussions to support something like this in setuptools here.

I would instead propose to give a LOG error message to the user when they want to use a feature like ReIDDetectMultiBackend with a non-pytorch weights file.

Example:

Running ReIDDetectMultiBackend with the specified ONNX weights

"<path>.onnx"

requires onnxruntime or onnxruntime-gpu to be installed!

Install with one of the following commands:
For CPU:

pip install onnx onnxruntime

For GPU:

pip install onnx onnxruntime-gpu

[mikel-brostrom]

The non-crucial packages are installed that way. The experience should be as smooth as possible. Giving an error if it is automatable is not an option 😄

[henriksod]

What if I use my global python env and have packages dependent on specific versions, that were to be auto updated running these? IMO being able to install packages dynamically is a design flaw with python.

At least there should be a flag and/or environment variable to toggle automatic installations. Just a thought :-)

[henriksod]

I had removed it in #986, but I will revert it then.

It would be easy to fix if it wasn't for the onnxruntime vs onnxruntime-gpu, but can't we have an environment variable or a program argument to activate auto install? It won't of course be possible in the code under boxmot/...

[mikel-brostrom]

The idea of envs is that you use one for a specific purpose. Working with a global env is always a bad idea. But I get what you mean. I think this solves the requirements customization based on environment needs: https://python-poetry.org/. Never used it though but seems like a good option.

[mikel-brostrom]

But let's discuss this further 😄 before implementing