we should allow users to predict on overlapping patches. this could improve outputs.
to do this, we can get a running sum of predictions in each region and also maintain a count of how many times each region is used. then we can divide the sum by the count to get the mean softmax probabilities for the regions.
we should allow users to predict on overlapping patches. this could improve outputs.
to do this, we can get a running sum of predictions in each region and also maintain a count of how many times each region is used. then we can divide the sum by the count to get the mean softmax probabilities for the regions.