Matteo Visconti di Oleggio Castello

oh cool, thanks for letting us now! I didn't know that pytorch supports the M1. Since I have an M1 based machine, I will try to implement this in himalaya...

Unfortunately, it looks like the MPS support in pytorch is far from being completed. Many linear algebra operators are not implemented yet (https://github.com/pytorch/pytorch/issues/77764). For example, using `torch.linalg.eigh` returns the following...

Tom can correct me if I'm wrong, but I think that using that flag will defeat the purpose of using GPU support, and it will be equivalent to running himalaya...

I will experiment a bit more and see what speedup we get even if we use `PYTORCH_ENABLE_MPS_FALLBACK=1` :)

Well, I'm happy I was wrong (by a lot). I ran the voxelwise tutorial that fits the banded ridge model (it's actually one banded ridge model + one ridge model)....

Another test: we don't get a noticeable speedup when running a simple ridge model. ## Backend `torch`, `svd` diag ``` python 05_plot_motion_energy_model.py 236.70s user 50.63s system 177% cpu 2:41.66 total...

See also this issue https://github.com/gallantlab/pycortex/issues/488 The surfaces shared with the NSD dataset may not conform to the standard freesurfer format, and so they require fixes prior to visualizing them in...

Hi @mingxue1204, have you manually flattened the surfaces and imported them with `cortex.freesurfer.import_subj`? `cortex.quickshow` requires flattened surfaces. If you don't have flattened surfaces, you can use the webgl viewer with...

Oh, thanks for sharing that. We need to fix our code to adapt to Nibabel's new version. In the meanwhile, you could downgrade nibabel to < 4.0 (`python -m pip...

Hi, what type of data are you trying to visualize? is it volumetric or surface-based? it looks like you're using the "fsaverage" subject, but we use that to plot surface-based...