How to manage memory buffer with Avalanche?

[bit0123]

Dear Avalanche team,
Thanks for making this excellent prototyping tool for CL experiments.
I would appreciate if you please address my following questions:

1. How to split available Avalanche dataset into train and validation? Does standard approach for splitting pytorch Dataset work here?
2. How to manage memory buffer for keeping subset of samples from old task and marge that to new dataset for task at time t?

Thanks for your reply.
Cheers!

[AntonioCarta]

Hello @bit0123,

1. you can't do it right now but we are going to solve it soon, probably next week. Follow issue Access the targets field through Subset #421 for updates on this.
2. We have avalanche.training.plugins.ReplayPlugin and related avalanche.training.strategies.Replay. You can use them directly or if you want to define your own custom replay you can start from their implementation.

[bit0123]

Hi @AntonioCarta , Thanks for prompt response.

[vlomonaco]

@bit0123 for 1. yes they would work but you won't be able to use the Avalanche benchmarks generators on them till #421 is solved. Can we mark this question as solved? :-)

[bit0123]

@vlomonaco Sure, Please. Thanks!