Remove need for registration mechanisms?

[dustinvtran]

Whether it be tfds' (e.g., replace tfds.builder('cifar10', ...) with tfds.image_classification.cifar10(...)) or our own (e.g., replace ub.models.get('resnet_50', ...) with ub.models.ResNet50(...) and even remove the registration mechanisms themselves under, say, datasets.py and models.py).

Registration mechanisms appear quite a bit in ML libraries. Their main purpose is to be easily varied under commandline/config flags where you can just change a different string. IMO modern tooling like ml_collection's config dict obviates the need for string arguments as you can execute arbitrarily complicated code as part of that config (it's all Python). In contrast, passing 'func(some_arg1=5, some_arg2=3)' leads to two major issues: (1) parsing indirection (2) unclear dependencies where you may need to import a module in order to register a custom dataset but that module is never directly used.

Curious what folks' thoughts are on this!

context: @BlackHC @y0ast @zi-w from chatting