Reimplementing DualNet using Avalanche

[zafstojano]

Dear Avalanche community,

I hope you are doing great.

This might be a newbie question, but I need your help. I would like to implement DualNet(https://arxiv.org/abs/2110.00175) using Avalanche, and I am trying to first think through the necessary components that I would need.

Given the open-source implementation of the method (https://github.com/phquang/DualNet), I believe I need to achieve the following:

1. Implement a replay memory, that stores inputs x, targets y, and features t (extracted from the inputs x) at the end of each task.
2. Due to the computation of the loss function, I would need to split the stream of current task examples, and previous task examples. To achieve this, I found this discussion (Implementing new methods which treat incoming and rehearsal data separately #644) very helpful.

My question is related to step 1.

What is the best way to implement a Reservoir to store the triplets (x,y,t), given the requirements of step 2 (separability)?

[vlomonaco]

Hi @zaf-stojano! Thanks for your question. We have an in-depth tutorial on how to handle replay in Avalanche. There are a few data loaders already implemented that may help you in what you are trying to achieve. Let us know if something is still unclear! :)

[zafstojano]

Thanks, I'll have a look :)