High error rate(0.53) when training a sneakers classfication model using resnet 50,want help

Hello everyone! :grinning:
Recently I’m doing a sneakers classification model using fastai.
I refer to some courses and some others’ code , making my own sneakers classification model as belowed.

I may do something wrong , I use fine_tune and fit_one_cycle+unfreeze alternately,and the error rate is
up to 50% while remaining stable high train loss and valid loss.
The dataset contains pics of adi and nike shoes , which are from the official website.
Sry for my poor english , thanks in advance for your every piece of advice.