Training config for EfficientNet v2s

[bonlime]

Hi,
I've noticed that you've released weights for EfficientNet v2s with score pretty close to the original one. I'm impressed and good job! Could you please share the training config used? I'm trying to reproduce their results using your codebase but haven't been successful yet. Hope your config could save me couple of weeks of trial-and-error.

[rwightman]

@bonlime it's a bit of a mess ;) I don't have adaptive resolution / hparams in my default training loop so I did it manually, started at 224x224 train w/ a RA + mixup EfficientNet v1 inspired train sched w/ 700 epoch targ (w/ a few extra timm specific addons like randerase), then around 300 epochs I resumed at 288x288 w/ increases in dropout, drop path, RE/RA strength. And then eval at approx 1.3x res.

[orilifs]

@rwightman Thank you for a detailed answer and a wonderful repo in general.

Speaking personally, I think I could learn a lot by seeing the code/script you used to execute the above procedure (or incase it was done manually, evan sharing the args for the train.py script would be awesome).

Much apprecitated anyhow :)

[bonlime]

@orilifs
I ended up writing my own Data Manager to allow changing augmentation during one training run. You could get some inspiration here

[orilifs]

@bonlime
Many thanks much appreciated.

[rwightman]

also hparams for v2_rw_m model were posted a while back here https://gist.github.com/rwightman/e69d5f456047c16773a77182cea68c3c ... I used something similar for the v2_rw_t but less augreg, no resolution switching was done for those two models...