CUDA implementation? about btrack HOT 4 OPEN

Additionally, I am looking at the logs and I'm having a bit of a trouble trying to get an understanding for progress based on on the values at each section. The inequality for the optimization problem makes sense, but is there any built in indication of progress?

from btrack.

For anybody who is having slow performance issues, I realized that my issue was due to incorrect array formatting:
I was passing in data in the form (x, y, z) (ex. 512,512,50) rather than (z, x, y) (50,512,512).
This ended up being up to how I went from a list of images as 2D numpy arrays to a 3D numpy array:

• np.dstack(masks) creates an array of the shape (512,512,50)
• np.array(masks) creates an array of the shape (50, 512, 512)

Small mistake, but stumped me for a little bit and created excessively long run times. Still curious about the CUDA implementation though!

from btrack.

Hi @Prehani - sorry it took a while to get back to you, and it seems like you solved the problem! Hope the results are good.

Re: the CUDA implementation, I made a start on this but never got round to finishing it. Parallelizing the bayesian update code is quite tricky.

https://github.com/quantumjot/BayesianTracker/blob/ac397675a425bf4abe6d9c09fd65f80072bf4950/btrack/src/belief.cu#L67-L75

from btrack.

No worries--preliminary testing appears to be very promising! We sent over some UI suggestions to the Napari team for tracks and a few other things, but so far BTrack has been working flawlessly after that small mix up on my end.

Out of curiosity, do you have any recommended segmentation packages for timeseries data? We've had decent luck with Cellpose, but it seems that there could be potential for a temporally aware algorithm using a similar tracking mechanism to generate an updated mask on a frame by frame basis rather than segmenting each image individually.

Re Cuda: totally understandable, I know it can be an absolute nightmare to work with. It seems for now our timeseries data is fine running solely on the CPU end of thing.

Thanks again for all the work you've done on this!

from btrack.