Skip to content

Conversation

@bonevbs
Copy link
Collaborator

@bonevbs bonevbs commented Jan 9, 2026

Draft commit to implement unified handling of perturbation modes - either through noisy S2 signals or through scalar noise vectors. Requires a rework of the preprocessor.

@bonevbs bonevbs requested a review from azrael417 January 9, 2026 14:25
Copy link
Collaborator

@azrael417 azrael417 left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

some minor remarks


self.noise_base_seed = 333 + comm.get_rank("data")

self.input_noise = GaussianVectorNoise(
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I am not happy with the name. Technically that is scalar noise with respect to coords. Since we noise the channels, shall we call it feature noise instead?

# For concatenation, we must expand explicitly if dimensions match but size doesn't
if (n.dim() == 5) and (n.shape[-2] == 1) and (n.shape[-1] == 1):
if (x.shape[-2] != 1) or (x.shape[-1] != 1):
n = n.expand(-1, -1, -1, x.shape[-2], x.shape[-1])
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why not expand the noise tensor emitted by the layer? That could also be a base noise S2 but diagonal in H and W>

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.

2 participants