...

[sbamopoulos]

Summary

Avoid loading the full new_labeling.npy array into memory on every Dask tile during the distributed relabeling step.

Problem

After stitching, distributed_eval saves a large new_labeling vector to disk and relabels the unstitched segmentation with:

dask.array.map_blocks(
    lambda block: np.load(new_labeling_path)[block],
    ...
)

np.load without mmap_mode copies the entire .npy file into each worker process for every block. On whole-slide runs this array can be multi-GB, so relabeling multiplies RAM use by the number of concurrent workers and can cause OOM or severe swapping.

Using mmap_mode='r' memory-maps the file so each block only reads the pages it needs.