search

vlomonaco / ar1-pytorch

AR1* with Latent Replay, implemented in PyTorch
https://arxiv.org/abs/1912.01100
Other
31 stars 9 forks source link

Reproducing results of NI and NC scenarios #5

Open fujenchu opened 3 years ago

fujenchu commented 3 years ago

Hi @vlomonaco,

Thanks for the great algorithm and codebase! I was about the reproduce the results of NICv2_391 by following the readme instruction.

I'd like to also run experiments of NI and NC scenarios as shown in this paper. I attempted to modify line 87 directly in ar1star_lat_replay.py but the results (~30 to 40%) I got was worse than NICv2_391 (which I expected to be a harder case).

Could you kindly show us where else needs to be adjusted (lr or other hyperparameters) in order to make it work properly? I think I am missing something. Thanks in advance!

Fu-Jen

vlomonaco commented 3 years ago

Dear @fujenchu,

In the paper you mentioned AR1 was not present and was only later proposed in [1, 2]. In these papers you'll find some tests on NC you can reproduce using the original codebases here: https://github.com/lrzpellegrini

This implementation in Pytorch instead will be discontinued soon in favor of the Avalanche implementation of AR1. Through Avalanche and this repo we plan to provide soon all the means to have good results on all the CORe50 scenarios.

Hope this helps!