GPU Support on Apple M3 Pro Chip

[mshallow]

I'm not sure if your codebase supports silicon GPUs at this time, but I am currently trying to un the segment_motifs step on your example data, and pytorch does not seem capable of using the GPU on my computer. According to pytorch's documentation, it should be capable of using the GPU for training and evaluation, but when I run the segment_motifs, at step of fitting the HMM, the code only uses my computer's CPU.
I tried the suggestion from the windows GPU issue, and changing from python 3.13 to 3.13 didn't make any difference.

Computer specs:
MacBook Pro with Apple M3 Pro Chip, MacOS Sequoia 15.5

Environment:
Conda environment with python=3.12.11, installed lisbet from the developer instructions on your installation guide.

I've verified that pytorch was correctly installed using the tests on their pages.

[gchindemi]

Hello @mshallow! GPU support is available for Apple silicon, but segment_motifs relies on multiprocessing for parallelism because it’s based on the CPU-only library hmmlearn.

Currently, LISBET supports GPU acceleration only during training and “direct” inference (i.e., annotate_behavior and compute_embeddings).

We plan to investigate alternatives to hmmlearn if it becomes a bottleneck for most users. Any feedback is welcome.

Please note that LISBET supports training HMM models on a reduced dataset using —fit_frac. In combination with the default multiprocessing, this usually strikes a good balance between the quality of the results and the time taken.

[mshallow]

Hi @gchindemi thanks for getting back to me! That's good to know that it shouldn't be trying to use the GPU, I would need to check, but I think that this step is where most of us in our lab trying this out ave run into issues. I'll try it out on a couple of different machines as well as looking into how long it takes on an HPC since we also have someone trying it out there.

[gchindemi]

Sounds good! Please let me know if there is anything I could help with.