training dataset

[backseason]

training dataset

[backseason]

Let N_s, N_e, and N_k represent the number of training samples of the saliency, edge and skeleton datasets, respectively.
During each iteration, we random sample an index i within the range of [0, max([N_s, N_e, N_k])).
Then the samples of indexes i % N_s, i % N_e, and i % N_k are used to select the training samples for the saliency, edge, and skeleton tasks with respect to the corresponding datasets.
It's just a work-around for random sampling from datasets of different shapes.

[clelouch]

Many thanks for your kind reply!