Skip to content

Conversation

rockerBOO
Copy link
Contributor

@rockerBOO rockerBOO commented Jun 19, 2025

https://arxiv.org/abs/2506.14280
https://github.com/team-approx-bayes/ivon
https://arxiv.org/abs/2402.17641

I'm just trying it out so I dunno if it's good or not.

Requires ivon-opt but only works with my branch because of device, mixed precision, and gradient accumulation issues.

pip install git+https://github.com/rockerBOO/ivon@gradient-accumulation
unet_lr = 1e-4 # Same LR you'd use for Adam
optimizer_type = "ivon.IVON"
optimizer_args = [
    "ess=400" # ess should be the number of items in your dataset. Maybe just the number of items per epoch?
]

enable_pruning=true # Optional, but might improve performance by pruning the high variance values

Default parameters:

        ess: float,
        hess_init: float = 1.0,
        beta1: float = 0.9,
        beta2: float = 0.99999,
        weight_decay: float = 1e-4,
        mc_samples: int = 1,
        hess_approx: str = 'price',
        clip_radius: float = float("inf"),
        sync: bool = False,
        debias: bool = True,
        rescale_lr: bool = True

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant