Open b93901190 opened 4 years ago
As my understanding for training procedure, RGB input patches are randomly cropped from training images. Then, it seems that not every pixels in training images are being processed in one epoch. Is it for some reason?
This is also my question, is it possible we miss some pixels, or how we can make sure that overlapping has effects on results or not?
As my understanding for training procedure, RGB input patches are randomly cropped from training images. Then, it seems that not every pixels in training images are being processed in one epoch. Is it for some reason?