Rework the agent skill into a general GP diagnosis skill by bwengals · Pull Request #62 · pymc-devs/ptgp

bwengals · 2026-10-07T02:32:46Z

Closes #34.

Renames the agent skill ptgp-vfe to ptgp and makes it general to all GP methods. Each pitfall page now describes a problem and its fix, the Tier A-E workflow (reference/workflow.md) is removed, and pages are tagged with applies_to (all, VFE, SVGP, VFE, or Unapproximated, VFE).

The content is also updated to match main (signatures, diagnostic fields, pg.fit), and two mechanisms are corrected:

Shrinking sigma increases the VFE trace penalty, so sigma collapse is not a VFE loophole. It's the usual GP interpolation failure, which VFE only allows when Z is dense.
Eta collapse pairs with sigma inflation, not sigma collapse.

The loop selected M points but only conditioned on the first M-1, so d_final, trace_curve[-1], and the repr's variance explained described one point fewer than was returned, and an early stop on threshold returned one unconditioned extra point. It now conditions on every selected point: trace_curve has shape (M + 1,), and early stop returns the smallest prefix below threshold. d_final was also returned in the internally shuffled row order; it is now aligned to the rows of X.

Rename docs/agents/ptgp-vfe to docs/agents/ptgp. Pages now describe each problem and its fix instead of a tiered escalation workflow: workflow.md is removed and tier references are replaced by the concrete action. Pages carry an honest applies_to (all, VFE/SVGP, or VFE) and VFE-only remarks are labelled. Also bring the content up to date with main: function names instead of line numbers, current signatures (check_init, Points(Z, Z_init), compile_kwargs, options vs scipy_options), the new VFEDiagnostics fields and excess_fit_per_n formula, KernelHealthDiagnostics, pg.fit / pg.predict, floatX, and the greedy_variance_init trace_curve / d_final semantics. Rewrite the sigma collapse mechanism (shrinking sigma grows the trace penalty; collapse needs Q close to Kff), fix the eta_collapse pairing and prior advice, and fix the check_inducing.py crash on KernelHealthDiagnostics.

bwengals · 2026-10-07T02:35:49Z

Merge after #61, and it'll need a some changes.

jessegrabowski

retarget to main

jessegrabowski · 2026-10-16T02:40:20Z

+  *interrupted* phase, so `compile_predict` wires up to the
+  partially-trained state.
+- **`compile_training_step(objective_fn, gp, X_var, y_var, model=None,
+  optimizer_fn=None, extra_vars=None, extra_init=None, frozen_vars=None,


stale on main after #65: optimizer_fn= and **optimizer_kwargs are gone, it's optimizer= (a pytensor_ml.optim transform, default adam(1e-2)). point to the ptgp-pytensor-ml skill

jessegrabowski · 2026-10-16T02:40:20Z

+
+## Optim — `ptgp/optim/training.py`
+
+- **`compile_scipy_objective(objective_fn, gp, X_var, y_var, model=None,


nit: the parameter is gp_model, not gp. gp= as a keyword raises TypeError

jessegrabowski · 2026-10-16T02:40:20Z

+  was built as `Points(Z_var, Z_init=Z0)`, nothing otherwise). Pass empty
+  tuples to exclude all extras. `init` defaults to `"prior_median"`
+  (improper priors fall back per RV; see the function docstring).
+- **`compile_scipy_diagnostics(diagnostic_fn, gp, X_var, y_var,


nit: gp_model, same as line 46

jessegrabowski · 2026-10-16T02:40:20Z

+  `result.status == 99`; `unpack`/`sp`/`se` correspond to the
+  *interrupted* phase, so `compile_predict` wires up to the
+  partially-trained state.
+- **`compile_training_step(objective_fn, gp, X_var, y_var, model=None,


nit: gp_model, same as line 46

jessegrabowski · 2026-10-16T02:40:20Z

+  param_groups=None, include_prior=True, compile_kwargs=None,
+  **optimizer_kwargs) -> (train_step, sp, se)`**: Adam / SGD step for
+  minibatch training. Mostly SVGP; listed for completeness.
+- **`compile_predict(gp, X_new_var, model, shared_params,


nit: gp_model, same as line 46

fixed, and compile_predict now takes X_train / y_train through **data, after #66

jessegrabowski · 2026-10-16T02:40:20Z

+Each page under `pitfalls/` describes one problem: how to detect it, what
+causes it, and how to fix it. Find the symptom below and open its page.
+`applies_to` in each page's frontmatter says which models it concerns:
+`all`, the inducing-point methods (`VFE`, `SVGP`), or `VFE` only. The


nit: excess_fit_per_n_negative uses [Unapproximated, VFE], which isn't one of these three values

added Unapproximated, VFE to the list in SKILL.md, since SVGP histories have no excess_fit_per_n

jessegrabowski · 2026-10-16T02:40:20Z

 2. **Loosen scipy tolerances**:

   ```python
-   scipy_options={"ftol": 1e-7, "gtol": 1e-5}
+   options={"ftol": 1e-7, "gtol": 1e-5}  # pg.fit / tracked_minimize kwargs


nit: gtol=1e-5 is already the L-BFGS-B default, so this only loosens ftol

jessegrabowski · 2026-10-16T02:40:20Z

+- **Per-step improvement**: divide
+  `(history[-1].elbo - history[-50].elbo)` (or the objective field your
+  history records) by the step count. If it's
+  > 1e-3 in absolute terms, real progress; bump maxiter. If it's


not a blocker: elbo is a sum over N, so fixed 1e-3 / 1e-6 per-step cutoffs depend on dataset size. divide by N

jessegrabowski · 2026-10-16T02:40:20Z

-Claude-specific convenience that symlinks it into a Claude Code skill
-directory (`~/.claude/skills/` or a project-local `.claude/skills/`);
-other AI tools can read `docs/agents/` directly.
+The repo ships backend-agnostic agent-skill docs: `docs/agents/ptgp/`


after the rebase docs/agents/ also has ptgp-pytensor-ml (#65). list it here

jessegrabowski · 2026-10-16T02:40:20Z

@@ -86,7 +86,7 @@ PTGP tries to distill some of the approaches of existing GP libraries and make t
 PTGP is set up to work nicely with AI coding assistants:

 - **[`AGENTS.md`](AGENTS.md)** — project-level instructions for AI coding assistants (architecture, conventions, where things live, how to run tests). Follows the [AGENTS.md](https://agents.md/) cross-tool convention used by Codex, Cursor, Aider, and others.
- **[`docs/agents/`](docs/agents/)** — backend-agnostic agent-skill docs covering folk wisdom and training-debug recipes. Currently includes [`ptgp-vfe`](docs/agents/ptgp-vfe/) (VFE diagnostic skill: pitfalls, escalation workflow, interpretation of `VFEDiagnostics` and `GreedyVarianceDiagnostics`).
+- **[`docs/agents/`](docs/agents/)** — backend-agnostic agent-skill docs covering folk wisdom and training-debug recipes. Currently includes [`ptgp`](docs/agents/ptgp/) (GP fit diagnosis: one page per problem with its cause and fix, plus interpretation of the diagnostic fields).


list ptgp-pytensor-ml too, same as AGENTS.md:190

fixed, and updated the summary bullet near the top too

bwengals added 2 commits October 2, 2026 13:13

This was referenced Oct 10, 2026

Add the Variational Nearest Neighbor GP (VNNGP) #69

Open

Add VGP point diagnostics and ptgp-vgp skill #58

Open

jessegrabowski reviewed Oct 16, 2026

View reviewed changes

bwengals added 2 commits November 5, 2026 13:36

Merge remote-tracking branch 'origin/main' into docs/agent-skill-refresh

39dfa79

Address review: match the skill to main's API and fix detection rules

337fab2

bwengals changed the base branch from fix/greedy-trace-curve to main November 5, 2026 21:58

bwengals requested a review from jessegrabowski November 5, 2026 23:17

jessegrabowski approved these changes Nov 6, 2026

View reviewed changes


		## Optim — `ptgp/optim/training.py`

		- **`compile_scipy_objective(objective_fn, gp, X_var, y_var, model=None,

Conversation

bwengals commented Oct 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

bwengals commented Oct 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

jessegrabowski left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

This branch has not been deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

bwengals commented Oct 7, 2026 •

edited

Loading

bwengals commented Oct 7, 2026 •

edited

Loading