predict_3D.py example produce overly large predictions

[fgdfgfthgr-fox]

It's not even finished and is already 5GB. One of my own prediction script produced a over 171GB prediction on 8nm scale mito alone. Compare to the few mb size of the last version, it's certainly too big.

My environment:
cellmap-segmentation-challenge: 0f50f70
cellmap-data: 2026.3.20.1935
The rest you know.

Honestly, the amount of bugs in the included code is way too much, and since the challenge is already over, I'm not sure if I should continue with cell-map anymore...

[rhoadesScholar]

One of my own prediction script produced a over 171GB prediction on 8nm scale mito alone

@fgdfgfthgr-fox does this mean that the script you wrote from the ground up to do prediction produced a 171GB prediction? This actually isn't that absurd, since test crop 984 is 800x800x1000 32nm voxels. Let's do the math:

$$ \begin{aligned} \frac{32\text{ nm}}{8 \text{ nm}} &= 4\text{x more voxels per side in prediction than in GT} \\ (4 × 800)^2 × (4 × 1000) &=4.096×10^{10}\text{ voxels} \\ 4.096×10^{10} × 32\text{ bits} &= 163.84\text{ GB} \end{aligned} $$

[fgdfgfthgr-fox]

... I used existing function in the cellmap-segmentation-challenge. It behaviors identical to yours except being cleaner and works for my model.
And are you sure that's the intended behaviour for your predict_3D.py as well? Pretty sure in the earlier version of the cellmap challange the result files are in the size of MBs, not GBs.