[MRG] Adding support for on-the-fly transforms

[hubertjb]

This PR adds support for on-the-fly transforms in BaseDataset and WindowsDataset. It also adds a similar API to modify the underlying transforms in BaseConcatDataset.

It is based on the way transforms are implemented in torchvision: https://pytorch.org/tutorials/beginner/data_loading_tutorial.html

[hubertjb]

It was brought to my attention that the discussion on #177 leaned toward using the collate_fn argument of DataLoader for implementing transforms as it allows batch-level augmentations.

With this in mind, do we want to also include a simpler Dataset-based implementation as in torch-vision and torch-audio? That's what @robintibor suggested here: #177 (comment)

@cedricrommel

[robintibor]

Sorry @sbbrandt I meant to ask here haha:
You also had something in this direction? How do you feel about this implementation? Is it compatible with what you had?

[robintibor]

But based on extensive previous discussion (thanks for linking @hubertjb ) I would also say this way could be merged already. and then batch level we can add separately at a later point.

[sbbrandt]

Sorry @sbbrandt I meant to ask here haha:
You also had something in this direction? How do you feel about this implementation? Is it compatible with what you had?

@robintibor Yes, makes more sense ;)

This implementation is also nice and works for many DA methods. E.g. mixup would need another implementation, but can still be implemented via a custom DataLoader, as well as betch level manipulations.

[robintibor]

Can you rebase this on current master to see if CIs run @hubertjb ?

For example in my setup steps would be:

git checkout master
git pull upstream master
git checkout on-the-fly-transforms
git rebase master
git push origin on-the-fly-transforms --force

[robintibor]

Just would like to know answer, then we can merge

[robintibor]

Ok I applied the changes then, also because I felt curious how Github UI works for this hihi @hubertjb

[robintibor]

Ok great, Github Actions succeeded, so merging! :) Thanks for the work