Update Qhull

[ximion]

Hi!
This was created in an attempt to fix #235, which seems to be a kind of stack corruption issue. Unfortunately I could not track down the root cause, but updating Qhull and guarding the message functions appears to make the problem go away for normal operation (even not guarding the log functions makes it much less likely). Unfortunately address sanitizer is still unhappy with the result, so this will not resolve the root cause.

It would probably need an extensive session in Valgrind (a major pain given that TensorFlow is in the mix and we allocate huge chunks of memory) to find out what is going on.
In the meanwhile, I take the less crashes.

Updating Qhull is probably a good idea anyway, since the shipped version was quite old.
Thanks for creating Stardist and considering this PR!

[maweigert]

Wow, thanks a lot for digging that deep! The whole qhull part is indeed a bit opaque and I remember that it was rather tedious to ensure its using the qhull thread-safe varieties throughout (as needed for OpenMP).

[ximion]

That makes sense! - Unfortunately we're probably not done with this, I just had Stardist process my large image (instead of just a smaller chunk of it), and it crashed again, after working flawlessly for 30 test runs before 😒
If I knew the code better, it would probably be easier to know where to look for a real cause of this issue.

(Still, updating Qhull would be useful and I think this PR is still fine - the generated results look sane to me)

[maweigert]

Ok, too bad!

You mentioned that tensorflow is a pain during debugging: one could avoid calling tensorflow during the NMS/qhull section by splitting the prediction and NMS call into two separate steps:

1. get predictions and save them somewhere

prob, dist, points = model.predict_sparse(img,...)

2. Do NMS/qhull part:

model._instances_from_prediction(img.shape, prob, dist, points, ...) 

cf.

stardist/stardist/models/model3d.py

Line 563 in 3ae059c

def _instances_from_prediction(self, img_shape, prob, dist, points=None, prob_class=None, prob_thresh=None, nms_thresh=None, overlap_label=None, return_labels=True, scale=None, **nms_kwargs):

[ximion]

Due to #249 actually solving the linked bug, I amended this PR to include only the Qhull update and nothing else (so, no mutex patch).