Does `--fixed_curvature=True` work? about mvae

[macio232]

Ok, I can see the reason

mvae/mt/mvae/models/train.py

Lines 179 to 184 in 96d851c

	if self.epoch < 10:
	for c in self.model.components:
	if isinstance(c, StereographicallyProjectedSphereComponent) or isinstance(c, SphericalComponent):
	c._pradius.data = torch.ones_like(c._pradius.data) * (11 - self.epoch)
	elif isinstance(c, PoincareComponent) or isinstance(c, HyperbolicComponent):
	c._nradius.data = torch.ones_like(c._nradius.data) * (11 - self.epoch)

And I found a reference for such behavior in the paper

The first 10 epochs for all models are trained with a fixed curvature starting at 0 and increasing in absolute value each epoch.

In my experiments, I have skipped the burn-in phase what resulted in AssestrionError in assert torch.isfinite(z_mean).all() in encode of component.

[oskopek]

Yes, stability is a big reason why we did the burn in. If you want, you can try ignoring assertions by adding the -O (big letter O) flag to python when running training.

Any other stability-related changes might also help, like lowering the learning rate, etc.