i already do one cyclic schedular and the same transforms but still fastai much better in loss results

do you recommend other tricks?

I noticed also that besides the one cycle schedular, fastai also uses other tricks such as discriminative learning rates. This is very obvious when using the fine_tune method. Perhpas that contributes to fastai getting better results.