The training results shown for fast.ai Jan 2020 all use the following params:
--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
It is trained with this version of fastai2's train_imagenette example. These params are changed as appropriate for each run:
--woof --size --epochs
The 80 and 200 epoch runs use mixup. 5 and 20 do not use mixup. All training was done on a single Titan RTX GPU. All results are from a single run; submissions using average of 5 runs would be much appreciated!
Full results can be analyzed in this spreadsheet.