How to train VNet using patches of smaller size ?

Hello, I implement VNet using keras to segment cells, and trained this net using 64*64*64 patches. But I found this trained net didn't work on patches of larger size in test. Should I crop the training data and testing data to same size ?  