training and validation acc is low on custom data

Hi! Thanks for the nice work. I have tried it on a custom data with 5 cell type classes. It works fine with the defalt parameters in train.py. However, the training and validation acc is low (aboud 0.16 and 0.12) on another custom data with about 20 celltypes. Could you please give some suggestions what issue is here? Looking forword for your reply.