Great, enjoy the rest of the course 
Hi, I’m also facing the same situation as Yesheng_Huang. I too opted to use the one hot encoding under train_val_generators function.
I’m clear about what you have suggested; to user sparse_categorical_crossentropy as the loss function. But isn’t it the case that we are required to set #units in the final Dense layer to 25 or 26 for above to work? is it correct to do so when the sign language seems to have only 24 signs?
Replying this thread because I am having this same issue and I’ve tried all the solutions presented here and nothing seems to be working @m.abidat , I’d appreciate an assistance from you.