Method → application → architecture organization + hybrid trainer#49
Open
kmheckel wants to merge 2 commits into
Open
Method → application → architecture organization + hybrid trainer#49kmheckel wants to merge 2 commits into
kmheckel wants to merge 2 commits into
Conversation
…d trainer Organizes Spyx around the question a new researcher actually asks — "how do I train this?" — by the order of information each method exploits, then maps that onto applications and architectures. Docs (user-facing): - docs/explanation/training-methods.md — the spine: 0th-order/evolutionary, 1st-order/surrogate, transfer/conversion+QAT, local/bio-inspired (honestly flagged roadmap, #27/#28), and the 0+1 hybrid. Each with when-to-use, trade-off, and Spyx entry point. - docs/explanation/choosing-an-approach.md — method×application and method×architecture decision matrices + a task→app→arch→method flow. - Wired into the mkdocs Explanation nav and linked from the home page. spyx.experimental.zoo — runnable reference recipes keyed by application and tagged by training method × architecture, all on synthetic data: - control-lif-es (evolutionary / LIF-MLP, evosax Open_ES) - classification-rsnn (surrogate / RSNN, axn.arctan) - language-s5 (gradient / S5, spyx.ssm.S5Diag) REGISTRY/list_recipes/get; each recipe exposes build/synthetic_batch/demo. spyx.experimental.hybrid — the 0+1 trainer: a surrogate gradient corrected by an antithetic-NES estimate of the true (hard-spike) loss gradient, projected orthogonal to the surrogate (the complement of Guided-ES). hybrid_gradient, make_hybrid_train_step, es_gradient, hybrid_diagnostics. research/new/ hybrid_evo_surrogate/ runs the three-arm study — honest negative in the smoke regime (surrogate wins; correction needs a small self-normalising lambda). Gate: 202 tests pass (16 new), ruff + ty clean, mkdocs build --strict clean. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 9fcd8582ef
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| :return: list of per-generation mean rollout costs (lower is better). | ||
| """ | ||
| import optax # lazy: optax is a core dep but imported here to keep import light | ||
| from evosax.algorithms import Open_ES |
There was a problem hiding this comment.
This import is only available when the optional [evo] extra is installed, but the new zoo tests call every recipe's demo() unconditionally and the CI test job syncs only --extra quant, so control-lif-es will raise ModuleNotFoundError: evosax in the default test environment. Either include the evo extra in the tested dependency set or skip/gate this recipe when evosax is unavailable.
Useful? React with 👍 / 👎.
Follow-up to the honest negative in the initial hybrid study: the raw-lambda correction can end up much WORSE than plain surrogate descent because the antithetic-NES estimate has norm ||g_orth|| that easily exceeds ||g_s||, so at a fixed lambda the noisy ES term dominates the update. Fix: add normalize=True to hybrid_gradient / hybrid_diagnostics / make_hybrid_train_step. In this mode lambda is a dimensionless FRACTION of the surrogate step — the correction is rescaled by lam_eff = lambda*||g_s||/||g_orth|| so ||applied correction|| = lambda*||g_s|| exactly, immune to the ES variance / sigma scaling. New diagnostics: lam_eff, correction_fraction. New test asserts the correction has norm lambda*||g_s||, stays orthogonal to g_s, and lambda=0 still recovers pure surrogate. Study: run.py is now a four-arm comparison (surrogate / es / hybrid-raw / hybrid-norm); sweep.py sweeps seeds x K x lambda over a harder regime and reports mean delta(true_loss) with seed spread so a "win" must survive noise. Honest verdict (README): self-normalising lambda removes the failure mode and brings hybrid to PARITY with the surrogate (never catastrophically worse) — a real improvement over the raw version — but it does NOT beat the surrogate on these easy synthetic tasks (task ceiling, little surrogate bias left to correct). Parity-plus-safety reported, not a win. Gate: ruff + ty clean; test_hybrid (6) + test_zoo (11) pass. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
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
1 participant
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.
What & why
Spyx had modules but no answer to the first question a new researcher asks: "how do I train this thing?" This PR adds a method-first organization — training methods indexed by the order of information they exploit — and maps it onto applications and architectures, plus two new experimental building blocks it references.
Docs (user-facing)
explanation/training-methods.md— the spine. Five families: 0th-order/evolutionary, 1st-order/surrogate gradient, transfer (conversion + QAT), local/bio-inspired (honestly flagged as roadmap — issues Implement Synthetic Gradients #27/neuron models for e-prop implementation #28), and the 0+1 hybrid. Each with when-to-use, the trade-off, and the concrete Spyx entry point.explanation/choosing-an-approach.md— two decision matrices (method × application, method × architecture) and atask → application → architecture → method"start here" flow.spyx.experimental.zoo— reference recipesRunnable recipes keyed by application, tagged by method × architecture, all on synthetic data (no downloads, sub-second smoke paths):
control-lif-esOpen_ES)classification-rsnnaxn.arctan)language-s5ssm.S5Diag)REGISTRY/list_recipes/get; each recipe exposesbuild/synthetic_batch/demo.spyx.experimental.hybrid— the 0+1 trainerThe novel method: a surrogate gradient corrected by an antithetic-NES estimate of the true (hard-spike) loss gradient, projected orthogonal to the surrogate — the surrogate supplies the cheap bulk descent direction, ES supplies only the correction in the subspace the surrogate is blind to. It is the complement of Guided-ES (which restricts ES search to the gradient's subspace; here we restrict it to the orthogonal complement).
hybrid_gradient,make_hybrid_train_step,es_gradient,hybrid_diagnostics.Honest result (
research/new/hybrid_evo_surrogate/, three-arm study): in the cheap smoke regime the pure surrogate wins —‖g_orth‖ ≈ 41 ≫ ‖g_s‖ ≈ 8.6at λ=0.5, so the high-variance correction dominates. The writeup reports this as a boundary result and proposes a small self-normalising λ (λ·‖g_s‖/‖g_orth‖) with larger K as the fix, not a win it didn't earn.Verification
uv run pytest -m "not network"— 202 passed, 1 skipped (16 new tests across zoo + hybrid)uv run ruff check/uv run ty check— cleanuv run mkdocs build --strict— clean⟨g_orth, g_s⟩ ≈ -3e-7).Everything new lands under
spyx.experimental(unstable-API contract); the stable core is untouched.🤖 Generated with Claude Code