Too much gap between train and validation accuracy

Hi, I’m doing Transfer Learning on a senet50 pretrained model, but I got weird results on the accuracy.
After one epoch my train accuracy is about 70% while validation on 25%. I read something about BatcNorm so I tried to retrain, setting to eval, rise the momentum, add batch accumulation but I got the same results.
What can I do? The code looks fine.
Train set contains 211588 images while validation 65060