search

arthurdouillard / dytox

Dynamic Token Expansion with Continual Transformers, accepted at CVPR 2022
https://arxiv.org/abs/2111.11326
Apache License 2.0
134 stars 17 forks source link

Confusing about the configurations #20

Closed haoranD closed 1 year ago

haoranD commented 1 year ago

Dear Authors,

Many thanks for such amazing work and I am very interested in your work.

Recently, It tried to run the code on two GPU via:

bash train.sh 0,1 \ --options options/data/cifar100_2-2.yaml options/data/cifar100_order1.yaml options/model/cifar_dytox.yaml \ --name dytox \ --data-path MY_PATH_TO_DATASET \ --output-basedir PATH_TO_SAVE_CHECKPOINTS \ --memory-size 1000

I am happy that I can successfully get the results you reported, which is 64.82 exactly.

WechatIMG10

HOWEVER, I am still a bit confused, how can I get the result as around 70.20 reported in the paper?

Looking forward to your reply, and I would be truly grateful if you could help with this. Thank you.

Best, Haoran

arthurdouillard commented 1 year ago

Those results were obtained with distributed memory while training with many GPUs (more than 2, thus giving more rehearsal than the white line with only 2 GPUs).

Those gray results were the original results in the paper. They make sense as we limit the rehearsal per-process, however for a fair evaluation, I've gray them out, as you should compare with the other newer results instead.