Repository navigation
Forward mean-field learning rule options by name - #816
Open
AHMETHAKANBEZIR1 wants to merge 1 commit into
Open
AHMETHAKANBEZIR1 wants to merge 1 commit into
AHMETHAKANBEZIR1 wants to merge 1 commit into
Conversation
Co-authored-by: OpenAI Codex <codex@openai.com>
This was referenced Oct 23, 2026
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
MeanFieldConnection passes its weight_decay positionally into AbstractConnection's reduction slot. The learning rule therefore gets zero weight decay, so a requested decay silently does nothing; providing reduction explicitly also raises a multiple-values TypeError in the constructor.
Forward nu and weight_decay by name so the inherited reduction option can bind correctly. For example, a shared weight0.5 with weight_decay0.1 remains0.5 after three Network.run steps on master, whereas the common learning-rule contract gives0.5*0.9**3=0.3645.
Validation against exact unmodified master78ec02b, Windows / Python3.12 / Torch2.14.0+cpu:
Eight regression/control cases: six fail and two zero-decay controls pass on master; all eight pass with the fix. Public Network.run checks repeated decay for scalar and per-target weights, including zero and full decay, while preserving the Parameter. Explicit sum/mean callbacks are accepted and reach the learning rule's reduction.
Complete local suite:263passed,1skipped. Package import path verified for this checkout in the Poetry environment.
Both configured all-file pre-commit hooks, normal commit hooks and staged diff check pass; changelog updated.
This is independent of #813's normalization and #815's CSRM state handling; no matching open fix or claimed issue was found. No CUDA, speed or task-accuracy benchmark is claimed. AI disclosure: the account owner requested OSS work; OpenAI Codex performed the source diagnosis, implementation, tests and description. The account owner has not independently reviewed the code. This is a constructed training reproducer, not a claimed production incident.
Local compatibility check: the updated heads of #813, #815 and #816 were applied together in a private checkout based on master78ec02b. The complete combined suite passed (297 passed, 1 skipped; Torch2.14 CPU; import path verified). Only the changelog additions needed a conflict resolution, which retained all entries; the implementation changes applied cleanly. This is local integration evidence, not an upstream merge or CUDA result.