Add MiniImagenet to the Benchmarks

[vlomonaco]

We should add MiniImagenet to the benchmarks. It's been used here: https://arxiv.org/abs/1902.10486

[lrzpellegrini]

This may be a good starting point: https://github.com/yaoyao-liu/mini-imagenet-tools They also have a link to a drive with all the processed images!

However, there doesn't seem to be an official static and long-term reliable download URL...

@vlomonaco what do you think? Considering that we don't have an official download URl we may just use the base ImageNet dataset and then subset+transform its patterns.

[vlomonaco]

I think at this stage whatever is faster to implement it's better, we will fix it afterwards in case it breaks! :-P

[vlomonaco]

the Dataset has been added in #291 ... we should create a "classic" benchmark based on the setting proposed in this paper: https://arxiv.org/abs/1902.10486

[vlomonaco]

@Mattdl can you add the "classic" benchmark based on your latest paper maybe? :)

[Mattdl]

We actually used the preprocessed MiniImagenet data in an AvalancheDataset, but as @lrzpellegrini mentioned, there is no official download URL for this preprocessed data.

[vlomonaco]

Yes, but I was asking about the "classic" benchmark, not the dataset itself. Since you have more experience with it maybe you can propose an high-level method to easily generate that benchmark.

[Mattdl]

I'll first take a look to see if I can be of help there (: I'll keep you posted here.

[Mattdl]

Due to prioritizing some other open issues and NeurIPS coming soon, I probably won't have time until the NeurIPS deadline.. Feel free to assign this one to another enthusiast (: