Training with multiple datasets

[ellesecotterill]

I was just hoping to clarify how training is performed when multiple training dataset are specified in the config e.g.
cfg.DATASETS.TRAIN = ("train_set_0", "train_set_1", "train_set_2", "train_set_3", "train_set_4", "train_set_5", )

Does the model train first on train_set_0, then train_set_1, etc - or is the data combined and shuffled before training? Is there a way to specify a particular sampling strategy e.g. to enforce that each batch has at least one sample from each dataset?