ToPoincare parameters riemannian = False

[Yaolu1998]

Hyp-OW/models/hypow_deformable_detr.py

Line 222 in ea93b74

self.tpc=ToPoincare(c=self.args.hyperbolic_c,ball_dim=hidden_dim,riemannian=False,clip_r=self.args.clip_r)

Good morning!

I am wondering why when projecting euclidean embeddings to hyperbolic space, the riemannian parameter is set to false in ToPoincare. If the loss is computed in hyperbolic space, then the scaling should be considered. When riemanian is set to false, the distance will not be updated correctly during back propogation because the it is not scaled by the Riemannian metric of the Poincare ball. Instead it is being optimized as if the geometry is still euclidean. Is there a specific reason to set it to false?

Thanks in advance!

[tldoan]

Hi,

Thanks for pointing this out, it is indeed a typo, I will fix it.

Best,
Thang,