Data Loaders should have the same API about avalanche HOT 7 CLOSED

That's more or less how I usually implement my dataloader for CL, so I guess it should be fine. I thought about downloading dataset in our code or inheriting the feature from PyTorch datasets when possible, but in the end I believe reimplementing the download component would give us more robustness against future updates.

from avalanche.

I guess we should leave this issue opened at least until Avalance 0.1 is ready. Since we are still adding new datasets, I don't think it would be safe to start modifying the existing ones.

from avalanche.

We can start moving all the benchmarks following the API proposed by @lrzpellegrini in the PR: #64
See the mnist example and especially the train_using_dataset function.

from avalanche.

Warning: this may break the preprocessing steps of most of the strategies already implemented!

from avalanche.

Ok, I will keep it in a separate branch and avoiding PR in the immediate future. We can separately test it before merging into master.

from avalanche.

I've thought a bit about this. I think that to have a general API we need to make all the benchmarks subclass of the Pytorch Dataset class, so that we can use the Pytorch data_loader class as the iterator.

Then, for every task/batch this iterator should return another Pytorch Dataset class so that we can use a data loader to iterate over it as well. What do you think?

from avalanche.

@lrzpellegrini any news on this?

from avalanche.