[hydra-configs-torchvision] Configs for MNIST datasets

[romesco]

Source: https://github.com/pytorch/vision/blob/master/torchvision/datasets/mnist.py

Using configen, create a subset of the torchvision datasets for MNIST. Pair with tests.

These should be along the same lines as @tkornuta-nvidia prototype in NeMo:
https://github.com/NVIDIA/NeMo/blob/main-vis-res/nemo/collections/cv/datasets/configs.py

[romesco]

@omry The PR for this will also give us a chance to think about how to logically structure managing 'two projects' - both torch and torchvision.

[romesco]

Another question is whether torchvision should become a dependency. It will need to be at least for tests.

[omry]

We need to decide if we want to have a different config distribution (pip pacakge) or if we are good with putting all subprojects in the same artifact.

Is torchvesion released with it's own versioning or is it tied to PyTorch's version?

[tkornuta-nvidia]

please note there is more than torchvision, i.e. torchaudio (that we are using in NeMo) and torchtext (that I don't really have experience)...

[tkornuta-nvidia]

We need to decide if we want to have a different config distribution (pip pacakge) or if we are good with putting all subprojects in the same artifact.

Is torchvesion released with it's own versioning or is it tied to PyTorch's version?

Yap, all those projects have their own versioning.

My proposition is to do it as optional package, with optional dependencies

so one can install

• torch-configs = torch-configs[torch] -> (torch) [default]
• torch-configs[all] -> torch + torchvision + torchtext + torchnlp
• or any subset e.g. torch-configs[torch,torchvision]

here's how we did it in NeMo when it comes to requirements:
https://github.com/NVIDIA/NeMo/blob/90775abc3a413baa0a6c8e68beb6160c7ebd108d/setup.py#L90
(still, I think we can make it better when it comes to actual "installation", as far as I know NeMo tries to install all collections disregarding the user choice during installation)

[omry]

Okay, so if we do go with different pip packages per target library we support we end up with:
For each package:

1. What Hydra version (or versions?) a package is supporting.
2. What version (or versions?) of the library the package is supporting.

I added some text to the google doc about versioning, let's discuss there.