Efficiency problem due to the for loops about s2cnn

[tscohen]

Yes, running these can take a long time, and it is likely that they can be made faster. However, they should only have to run once at the beginning of your script, because the results are cached using the @lru_cache(maxsize=32) statement. It should also be possible to implement caching that works across multiple runs.

That said, I don't know if the current code (including the stuff you run each epoch) would be fast enough for your use case. 256x256 is pretty big, even for a 2D CNN. A spherical CNN with SO(3) feature maps is more like a 3D CNN, and there a feature map of 256x256x256 is currently considered very large. Moreover, the SO(3) convolution code is less optimized than CUDA's 3D-Conv code. It's certainly not impossible to scale this up, but will take some engineering effort.

[mariogeiger]

Now in af8768c these functions are cached on disk