Skip to content

Fix dropout inference and terminal/singleton PPO updates - #3

Open
jadenfix wants to merge 3 commits into
codex/readiness-20261012-pythonfrom
codex/audio-qa-20261013-inference-ppo
Open

jadenfix wants to merge 3 commits into
codex/readiness-20261012-pythonfrom
codex/audio-qa-20261013-inference-ppo

Conversation

@jadenfix

Copy link
Copy Markdown
Owner

Follow-up stacked on #2 at b0515f9; keep its implementation owner and merge order.

Greedy action scoring previously left Dropout active even with training=False, so identical inputs could select different actions. Temporarily use eval mode and restore the caller's mode. PPO singleton batches previously normalized with unbiased std and made parameters NaN; use population std. Mask advantage recursion at done boundaries: two terminal rewards [1,100] now produce critic targets [1,100], rather than [100,100]. Keep zero-variance explained variance finite and preserve device placement for bootstrap/diagnostics.

Validation: full offline Python suite 30 passed. All three new real-network regressions fail against the #2 baseline. Existing DQN optimizer/checkpoint workflow remains green. git diff --check passes.

Scope: mathematical and action-scoring repairs. Environment rewards/observations remain simulated; live speech quality is not validated. The separate SER training path is still blocked by missing packaged configuration and model initialization/configuration defects. PPO replay/on-policy design and bootstrap handling across arbitrary shuffled transitions are not established by these terminal-boundary tests. Launcher #1 remains independent.

@vercel

vercel Bot commented Oct 13, 2026 •

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated
human-voice-ai Error Error Oct 13, 2026 1:09am UTC

@jadenfix

Copy link
Copy Markdown
Owner Author

Implementation owner handoff: SER completion is now reviewable in #4 at 68dc02b813626a29c162c5f8726059f1a5fb3bed. It branches from your exact #3 head and preserves all dropout/PPO/terminal fixes, plus the conflict-free #1 launcher patch. Four new real computation tests bring the integrated suite to 34 passing tests. Actual synthetic audio features -> three SER Adam updates -> strict complete-config checkpoint restore -> identical probabilities -> greedy policy action also passes from an installed wheel outside the checkout. No semantic emotion accuracy/live speech claim. Owner branches remain unchanged; merge coordinator will preserve support changes when consolidating.

This branch had an error being deployed

1 failed deployment
Preview — 60fb4983 Deployed Oct 13, 2026 by vercel[bot]
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