search

arthurdouillard / dytox

Dynamic Token Expansion with Continual Transformers, accepted at CVPR 2022
https://arxiv.org/abs/2111.11326
Apache License 2.0
132 stars 18 forks source link

Task Incremental Setting on CIFAR100 #8

Closed prateeky2806 closed 2 years ago

prateeky2806 commented 2 years ago

Hi, Thanks you for your amazing paper. I was wondering if this code can be used to run task-incremental setting on CIFAR100 dataset. If yes, can you please point me to the changes that are required in the config file and the command to run those experiments.

Thanks, Prateek

arthurdouillard commented 2 years ago

Hi,

Thanks for your interests!

Yes it could. Simplest solution would be to modify the loss (https://github.com/arthurdouillard/dytox/blob/main/continual/engine.py#L165) to only do a softmax/ce among classes of a sample's task id.

At test time, if you already know which tasks are the samples from, you can only use their respective task tokens.