I have been using Fastai to treat computer vision problems for 3 years and I still believe that Fastai is a powerful library to climb the LB.

As always in any CV competition, I used the baseline: fastai+some heavy data augmentation using fastai Vision augmentation + efficientnet-b0 --> efficientnet-b5

My 3rd place solution was based on efficientnet-b4 model (score 0.9789963593689572). I forget to select my best two submissions. Sadly the system selected automatically the 3rd place solution. The second best submission was an Ensembling (Averaging) of efficientnet-b4 and efficientnet-b2 and it give me the same score as efficientnet-b4 on the public LB but on the private LB it give me a score: 0.9799698614426501 which can lead me to the 2nd place.