fastremap.unique returns an incorrect counts if applied to a large image

[ivonindima]

The function fastremap.unique returns an incorrect counts if applied to a large image. This happens on Windows if the number of voxels in any phase exceeds the maximum allowable value for np.uint32 (4_294_967_295).
For me the problem was solved by replacing uint32 with uint64 in unique_via_array.

import fastremap

N = 2500
img = np.random.randint(2, size=(N, N,N), dtype=np.uint8)
vals, counts = fastremap.unique(img, return_counts=True)
print(vals, counts/img.size) # Expected [0 1] [0.5 0.5]

[yojeep]

I have the same question

[william-silversmith]

For some reason, the obvious way to make this work with cython templates didn't work. I'm hesitant to simply just bump np.uint32 to np.uint64 because it will inflate memory usage significantly in some cases. Will think about this.