This repository was archived by the owner on Mar 15, 2024. It is now read-only.

Description
Hi, thank you for the great repo!
I was wondering what were the commands to launch the training to reproduce the results of the CAIT models.
For now I am mostely interested in the XS and XXS models with 224 resolution.
Is changing drop path and layer scale initialization sufficient?
Also dropout should be at zero?
for now I have the following, after adding the following import: import cait_models in main.py:
python main.py \
--data-path Imagenet \
--model cait_XXS24_224 \
--batch-size 1024 \
--lr 0.001 \
--warmup-lr 0.00001 \
--drop-path 0.05
Thank you!
Elias