Add ICaRL to the CL Baselines

[vlomonaco]

Add the ICaRL strategy to the main CL Baselines.

[akshitac8]

@vlomonaco Can I start with this?

[lrzpellegrini]

Hi Akshita, I already wrote and tested the PyTorch code iCaRL (which only needs to be ported to Avalanche).

I had to pause the porting because I'm been working on max priority issues. However, I think I can still handle this!

[akshitac8]

okay @lrzpellegrini

[vlomonaco]

This is gonna be added soon with a PR by Andrea Rosasco.

[Mathieu4141]

Are you still working on that @andrew-r96 ? For a school project we implemented iCaRL as a baseline, and we did it using avalanche. The code is not up to date with the latest version of avalanche, but it shouldn't be too much work to clean it.
Let me know if you're still on it :)

[andrearosasco]

Hi @Mathieu4141 thanks for the interest :)
Yeah I'm still working on it and just opened a PR here #637 if you wanna take a look.

[Mathieu4141]

Great! Yes I'll take a look at your PR 🙂

[qsunyuan]

Hi!

Great work.

Recently I also reproduced iCaRL, but I cannot achieve the original results.

Could u pls share ur CIFAR100 icarl results.

For example, ten tasks which each task has 10 classes.

I cannot achieve the original results, about 49%. How's ur results.

Hope to get ur replay. Best wishes.

[vlomonaco]

Hi @qsunyuan! Results on Cifar100 are reproduced here: https://github.com/ContinualAI/reproducible-continual-learning

[qsunyuan]

THx for ur help, it helps me a lot. Best wishes.