Sample selection in RoI (Faster/Mask R-CNN)

In my dataset there are 3 classes (+background)
I set

'box_positive_fraction':0.75
'box_batch_size_per_image':1024

in the Mask R-CNN constructor arguments. Yet after the sampling (select_training_samples) there are roughly 90% Class 0, 6% Class 1, 2% Class 2 and 3, which is passed on to the loss function. Why is the final breakdown different from the 0.75? Is it how it’s supposed to work?