Fix BatchNorm/LieBN parameter resets and bias being applied when disabled

[lucas-heck]

Problem

SPDBatchNormMean's was still applying an trainable bias when rebias=False.

After parametrization, reset_parameters() was modifying the temporary tensors instead of the underlying parameters, which was preventing biases and scales from resetting correctly in SPDBatchNormMean, SPDBatchNormMeanVar, and SPDBatchNormLie.

Fix

Only create and parametrize the layer’s bias when rebias is enabled.

Reset of parametrized parameters now is done through assignment, which invokes their right_inverse.

Verification

Added tests for disabled bias and LieBN resets after training across all the metrics.

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

❌ Patch coverage is 81.25000% with 6 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
spd_learn/modules/batchnorm.py	64.70%	3 Missing and 3 partials ⚠️

📢 Thoughts on this report? Let us know!

[bruAristimunha]

Thanks @lucas-heck, good catch. With rebias=False, the old code still applied the bias. That bias was created with torch.empty and never reset, so it was uninitialized memory, which makes this a real correctness fix.

Checked before merging:

• Full suite with all four open PRs merged together: green.
• SPDBatchNormMean / SPDBatchNormMeanVar: after perturbing every parameter, reset_parameters() restores the identity bias and unit weight, and the parameter objects stay the same (so optimizer state remains valid). LieBN is covered by your new test.

Minor, non-blocking:

• Old SPDBatchNormMean(rebias=False) checkpoints contain parametrizations.bias.original, so load_state_dict(strict=True) will now report an unexpected key. Those weights were uninitialized anyway, so I'd just mention it in the release notes.
• The "eye → assign if parametrized, else copy_" block now appears three times. A small shared helper could remove the repetition at some point, but it isn't needed for this PR.

I'll merge main into the branch to resolve the CHANGELOG.md conflict, then merge. Thanks!