blow-up in the number of samples in 2D orthoplane predictions

[eminorhan]

I'm trying to generate predictions from a 2D model for all test crops, following the predict_2D.py script under examples (except that crops="test" instead of crops="19"). The predictions for the first crop in test_crop_manifest.csv (crop 557) are completed within a reasonable amount of time, however for the second crop (crop 980), I noticed that the dataset_writer here seems to generate a massive number of samples even for a single axis, for example ~165M samples for the first axis, z (for comparison, for the first test crop, the corresponding number is only 129 samples). This means that it would take an infeasible amount of time to calculate the predictions even for a single axis (even with a large batch size). I'm just trying to understand why this blow-up seems to happen for the second test crop and how I might prevent it.

[eminorhan]

It seems like this behavior is caused by the dataset_writer generating all possible input arrays of given size from the crop volume. I'm not sure if this is ideal though, as there is going to be massive overlap in those arrays (for example, slices that are just 1 pixel shifted versions of each other). Is there a way to override this behavior and generate non-overlapping arrays instead?