Hi and thanks for interest!

I think it is the same logic here: https://medium.com/@arthurflor23/hi-20ff375d7153 but using MNIST database as custom_dataset, you know.

Then you can make a function that partition the mnist dataset at 80/10/10% (train/valid/test), use `--transform` to create a HDF5 file and then use the`--source=mnist` to train and predict.