The baseline was trained with the following argumeents:
--lr 8e-3 --sqrmom 0.99 --mom 0.95 --eps 1e-6 --bs 64 --opt ranger --sa 1 --fp16 1 --arch xse_resnext50 --pool MaxPool
These params are changed as appropriate for each run:
--woof --pct_noise --size --epochs
It was trained with this version of train_noisy_imagenette.py. Training jobs were run with this version of the run_leaderboard.sh script. Training was done on 4 Titan RTX GPUs, each GPU running a different level of noise. All results are from a single run, with a seed set to 999 (set_seed(999, reproducible=True)).