Reason for using random region only in 'mini-batch'

[ildoonet]

If we use random regions of data from full dataset, Can it improve the performance further?

Have you experimented this?

(currently, random regions are sampled from 'mini-batch')

[hellbell]

@ildoonet
It is a good question. Unfortunately, we didn't experiment that.
Currently, we consider the efficiency for loading image data, so the random region is selected only in the 'mini-batch'.
It might improve the performance by using outside the 'mini-batch' data, but I guess there is no big difference if we use sufficiently large mini-batch size (we use 256).
I will be very happy if you test and verify this. :)
Thanks!