cuai / differentiable-frechet-mean Goto Github PK

I did the same setup as in the example but replace Poincare with Lorentz. This leads to a dimension mismatch and the forward fails.

line 21, in _ldot
    m = u * v
RuntimeError: The size of tensor a (3) must match the size of tensor b (4) at non-singleton dimension 0

Hi, thanks for your great work that provides a solution to calculate the Frechet mean.

The code you released only supports to calculate one 'mean' for a set of points on some manifold. Do you plan to realize a version for batch compution.

By the way, it seems to be time-consuming while the hyper-parameter "max_iter" is set to 1000 defaultly. Is there any suggestion or experiment evaluation of trade-off between the time complexity and accuracy?

Best regards.

Hi, thank you for sharing this work. I found that using Lorentz model returns nan most of the time. I am not sure if this is because the Lornetz model is unbounded.

Thanks for this beautiful work.
Sorry I'm new to torch. I'm not sure if this is how you modified HGCN to incorporate Frechet mean, note line 72(unchanged) and 78(changed):

But this comes with errors. Could you tell me how exactly you did?
Thanks!

Applying frechet mean on the Lorentz model with K not equal to -1 cannot get the correct answer.

I found the bug in frechetmean/forward/hyperboloid_forward.py. Where the original code calculate variable mu with
mu = u / torch.sqrt(K * Lorentz._ldot(u, u, keepdim=True))

Changing it to
mu = - K * u / torch.sqrt(K * Lorentz._ldot(u, u, keepdim=True))
solves the problem.