How did you train on multiple datasets.

Very nice work, thanks for your efforts with open source.
I notice that you are combining multiple datasets to train for one large dataset.
Engineering wise, how did you combine these datasets? For example the image file obviously has multiple addresses, do you put them in the same folder under
Looking forward to your reply:)