refactor(ml): DRY pass — shared sweep helpers + table-driven settings (#161)

Consolidate duplication accrued across the ML tagging + settings backend, behavior-preserving (over-DRY guard applied — the three auto-apply sweep BODIES stay separate; only their shared inner helpers are extracted). - _sigmoid / _conflict_scores / _insert_presentation_review (heads.py): the score→prob transform (6 inlined sites), the presentation conflict signal (2 sites), and the ring-loud PresentationReview insert (2 sites, single- sourced so the mode column can't drift on the shared composite PK). - _applied_or_rejected (training_data.py): the per-tag "applied ∪ rejected" skip-set, byte-identical at 3 sweep sites (heads.py x2, tasks/ml.py ccip). - ccip sweep divergence fixes: import ccip._FIGURE_KINDS + training_data._l2norm instead of local copies that silently drift when the canonical changes. - MLSettings.load / .load_sync classmethods (mirror ImportSettings); route all 8 scalar_one singleton reads through them (the session.get None-path stays). - GET serializers for MLSettings + ImportSettings are now table-driven off the same _EDITABLE tuples PATCH writes, so a new field can't be silently absent from GET (the split that historically dropped fields). - AUTO_APPLY_THRESHOLD_MIN/MAX constant single-sources the [0.5,0.999] operating range across the service clamp + the 5 API validators. - test_ml_dry_helpers.py pins _applied_or_rejected + _sigmoid. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NsmJSQxnNxGgtM5Yz4GAqi
2026-07-13 21:41:24 -04:00
parent d80a5255ed
commit 666b3a2ec8
9 changed files with 187 additions and 172 deletions
@@ -150,9 +150,7 @@ def refresh_character_prototypes(
    """Incrementally refresh the prototype store. `full=True` rebuilds every
    character regardless of the gate/fingerprints (nightly reconcile). Returns
    {skipped, rebuilt, removed}; commits."""
-    settings = session.execute(
-        select(MLSettings).where(MLSettings.id == 1)
-    ).scalar_one()
+    settings = MLSettings.load_sync(session)
    sig = _global_signature(session)
    if not full and settings.ccip_ref_signature == sig:
        return {"skipped": True, "rebuilt": 0, "removed": 0}
@@ -204,9 +202,7 @@ def retract_auto_applied_ccip(session: Session) -> int:
    n_retracted."""
    import numpy as np

-    settings = session.execute(
-        select(MLSettings).where(MLSettings.id == 1)
-    ).scalar_one()
+    settings = MLSettings.load_sync(session)
    if not settings.ccip_auto_apply_enabled:
        return 0
    thr = float(settings.ccip_auto_apply_threshold)
@@ -23,6 +23,7 @@ from datetime import UTC, datetime
 from typing import Any

 from sqlalchemy import delete, exists, func, select
+from sqlalchemy.dialects.postgresql import insert as pg_insert
 from sqlalchemy.ext.asyncio import AsyncSession
 from sqlalchemy.orm import Session

@@ -42,6 +43,7 @@ from ...models import (
 from ...models.tag import CHROME_SYSTEM_TAGS, PROCESS_SYSTEM_TAGS, image_tag
 from .training_data import (
    _AUTO_SOURCES,
+    _applied_or_rejected,
    _auto_apply_point,
    _hygiene_excluded_ids,
    _ids_with_tag,
@@ -61,6 +63,14 @@ MIN_POSITIVES_FLOOR = 8     # hard floor; settings.head_min_positives can raise
 _UNLABELED_POOL = 4000
 _EXAMPLES_MIN = 8           # need at least this many embedded +/- to fit a head

+# Auto-apply / match confidence operating range. Every graduated auto-apply or
+# CCIP-match threshold the operator can set lives in this band, and the head
+# precision target is clamped to it: below 0.5 "auto-apply" is meaningless, and
+# 1.0 is unachievable so 0.999 is the ceiling. One source shared by the service
+# clamp (_normalize_params) and the API validator (ml_admin._validate).
+AUTO_APPLY_THRESHOLD_MIN = 0.5
+AUTO_APPLY_THRESHOLD_MAX = 0.999
+
 # Only these tag kinds get heads (the surfaced suggestion categories).
 _HEAD_KINDS = (TagKind.general, TagKind.character)
 # tag.kind -> the suggestion category the rail groups under.
@@ -78,6 +88,38 @@ _CATEGORY = {TagKind.general: "general", TagKind.character: "character"}
 _SYSTEM_TAG_SUGGEST_FLOOR = 0.65


+def _sigmoid(z, np):
+    """Logistic sigmoid 1/(1+e^-z): the head score→probability transform. One home
+    for what was inlined at every scoring site (suggest, both sweeps, retract)."""
+    return 1.0 / (1.0 + np.exp(-z))
+
+
+def _conflict_scores(Xn, Wc, bc, np):
+    """The presentation conflict signal (#141): per row, the MAX content-head
+    probability and WHICH head produced it. Shared by the system-tag sweep's guard-2
+    and the soft-wip audit — both ask "does this ALSO look like real content?"."""
+    cprobs = _sigmoid(Xn @ Wc.T + bc, np)
+    return cprobs.max(axis=1), cprobs.argmax(axis=1)
+
+
+def _insert_presentation_review(
+    session, *, image_record_id, tag_id, conflict_tag_id, conflict_score, mode,
+):
+    """Single-source the ring-loud PresentationReview row shape so the two writers
+    (system-tag sweep guard-2 + soft-wip audit) can't drift on columns or `mode` —
+    they share the (image_record_id, tag_id) composite PK, so a divergent `mode`
+    would be a silent first-writer-wins bug."""
+    session.execute(
+        pg_insert(PresentationReview)
+        .values(
+            image_record_id=image_record_id, tag_id=tag_id,
+            conflict_tag_id=conflict_tag_id, conflict_score=conflict_score,
+            mode=mode,
+        )
+        .on_conflict_do_nothing()
+    )
+
+
 class HeadTrainingAlreadyRunning(Exception):
    """Raised by start_head_training_run when a run is already in flight."""

@@ -103,9 +145,7 @@ def start_head_training_run(session: Session, params: dict[str, Any]) -> int:


 def _settings(session: Session) -> MLSettings:
-    return session.execute(
-        select(MLSettings).where(MLSettings.id == 1)
-    ).scalar_one()
+    return MLSettings.load_sync(session)


 def _normalize_params(session: Session, params: dict[str, Any] | None) -> dict[str, Any]:
@@ -124,7 +164,7 @@ def _normalize_params(session: Session, params: dict[str, Any] | None) -> dict[s
    except (TypeError, ValueError):
        cv_folds = DEFAULT_CV_FOLDS
    try:
-        precision_target = min(max(float(params.get("precision_target", s.head_auto_apply_precision)), 0.5), 0.999)
+        precision_target = min(max(float(params.get("precision_target", s.head_auto_apply_precision)), AUTO_APPLY_THRESHOLD_MIN), AUTO_APPLY_THRESHOLD_MAX)
    except (TypeError, ValueError):
        precision_target = s.head_auto_apply_precision
    return {
@@ -536,7 +576,7 @@ async def score_image(
    norms[norms == 0] = 1.0
    Xn = X / norms
    Z = Xn @ heads["W"].T + heads["b"]                  # (B, H)
-    probs_bag = 1.0 / (1.0 + np.exp(-Z))                # (B, H)
+    probs_bag = _sigmoid(Z, np)                         # (B, H)
    probs = probs_bag.max(axis=0)                       # (H,) best over the bag
    # ARGMAX beside the max: WHICH bag row won each head → the region that grounds
    # the tag (bag_meta[win]); None when the whole-image vector won (#1206).
@@ -614,9 +654,7 @@ async def ground_applied_tag(


 async def _settings_async(session: AsyncSession) -> MLSettings:
-    return (
-        await session.execute(select(MLSettings).where(MLSettings.id == 1))
-    ).scalar_one()
+    return await MLSettings.load(session)


 # --- Earned auto-apply (sync, ml worker) ---------------------------------
@@ -687,7 +725,6 @@ def auto_apply_sweep(
    embeddings in chunks; commits per chunk on a real run. Returns
    {n_applied, concepts:[{tag_id,name,applied,scanned,threshold}]}."""
    import numpy as np
-    from sqlalchemy.dialects.postgresql import insert as pg_insert

    settings = _settings(session)
    rows = _auto_apply_heads(
@@ -704,18 +741,7 @@ def auto_apply_sweep(
    names = [r.name for r in rows]

    # Skip images that already carry, or have rejected, each tag.
-    skip = {tid: set() for tid in tag_ids}
-    for tid in tag_ids:
-        for (iid,) in session.execute(
-            select(image_tag.c.image_record_id).where(image_tag.c.tag_id == tid)
-        ):
-            skip[tid].add(iid)
-        for (iid,) in session.execute(
-            select(TagSuggestionRejection.image_record_id).where(
-                TagSuggestionRejection.tag_id == tid
-            )
-        ):
-            skip[tid].add(iid)
+    skip = _applied_or_rejected(session, tag_ids)

    applied = [0] * len(rows)
    scanned = 0
@@ -729,7 +755,7 @@ def auto_apply_sweep(
        if not cids:
            continue
        Xn = _l2norm(np.vstack([emb[i] for i in cids]).astype(np.float32), np)
-        probs = 1.0 / (1.0 + np.exp(-(Xn @ W.T + b)))   # (N, H)
+        probs = _sigmoid(Xn @ W.T + b, np)              # (N, H)
        scanned += len(cids)
        for h in range(len(rows)):
            tid = tag_ids[h]
@@ -840,7 +866,6 @@ def system_tag_auto_apply_sweep(
    enabled flag is set. numpy-only (no sklearn). Returns {n_applied, n_flagged,
    concepts}."""
    import numpy as np
-    from sqlalchemy.dialects.postgresql import insert as pg_insert

    cfg = _SWEEP_MODES[mode]
    settings = _settings(session)
@@ -869,18 +894,7 @@ def system_tag_auto_apply_sweep(
    valued = _valued_image_ids(session)

    # Skip images that already carry, or have rejected, each presentation tag.
-    skip = {tid: set() for tid in pres_tag_ids}
-    for tid in pres_tag_ids:
-        for (iid,) in session.execute(
-            select(image_tag.c.image_record_id).where(image_tag.c.tag_id == tid)
-        ):
-            skip[tid].add(iid)
-        for (iid,) in session.execute(
-            select(TagSuggestionRejection.image_record_id).where(
-                TagSuggestionRejection.tag_id == tid
-            )
-        ):
-            skip[tid].add(iid)
+    skip = _applied_or_rejected(session, pres_tag_ids)

    applied = [0] * len(pres)
    n_flagged = 0
@@ -895,11 +909,9 @@ def system_tag_auto_apply_sweep(
        if not cids:
            continue
        Xn = _l2norm(np.vstack([emb[i] for i in cids]).astype(np.float32), np)
-        probs = 1.0 / (1.0 + np.exp(-(Xn @ Wp.T + bp)))       # (N, P)
+        probs = _sigmoid(Xn @ Wp.T + bp, np)                  # (N, P)
        if Wc is not None:
-            cprobs = 1.0 / (1.0 + np.exp(-(Xn @ Wc.T + bc)))  # (N, C)
-            max_c = cprobs.max(axis=1)
-            arg_c = cprobs.argmax(axis=1)
+            max_c, arg_c = _conflict_scores(Xn, Wc, bc, np)   # (N,), (N,)
        scanned += len(cids)
        for p in range(len(pres)):
            tid = pres_tag_ids[p]
@@ -924,15 +936,12 @@ def system_tag_auto_apply_sweep(
                if Wc is not None and float(max_c[idx]) >= conflict_thr:
                    n_flagged += 1
                    if not dry_run:
-                        session.execute(
-                            pg_insert(PresentationReview)
-                            .values(
-                                image_record_id=iid, tag_id=tid,
-                                conflict_tag_id=conf_tag_ids[int(arg_c[idx])],
-                                conflict_score=float(max_c[idx]),
-                                mode=mode,
-                            )
-                            .on_conflict_do_nothing()
+                        _insert_presentation_review(
+                            session,
+                            image_record_id=iid, tag_id=tid,
+                            conflict_tag_id=conf_tag_ids[int(arg_c[idx])],
+                            conflict_score=float(max_c[idx]),
+                            mode=mode,
                        )
        if not dry_run:
            session.commit()
@@ -956,7 +965,6 @@ def soft_wip_conflict_audit(session: Session, dry_run: bool = False) -> dict:
    NOT remove the tag; the operator decides. No-op when there are no content heads.
    numpy-only. Returns {n_scanned, n_flagged}."""
    import numpy as np
-    from sqlalchemy.dialects.postgresql import insert as pg_insert

    from ..wip_title import WIP_TITLE_SOFT_SOURCE, resolve_wip_tag_id

@@ -993,22 +1001,17 @@ def soft_wip_conflict_audit(session: Session, dry_run: bool = False) -> dict:
            continue
        scanned += len(cids)
        Xn = _l2norm(np.vstack([emb[i] for i in cids]).astype(np.float32), np)
-        cprobs = 1.0 / (1.0 + np.exp(-(Xn @ Wc.T + bc)))
-        max_c = cprobs.max(axis=1)
-        arg_c = cprobs.argmax(axis=1)
+        max_c, arg_c = _conflict_scores(Xn, Wc, bc, np)
        for k in range(len(cids)):
            if float(max_c[k]) >= conflict_thr:
                n_flagged += 1
                if not dry_run:
-                    session.execute(
-                        pg_insert(PresentationReview)
-                        .values(
-                            image_record_id=cids[k], tag_id=wip_id,
-                            conflict_tag_id=conf_tag_ids[int(arg_c[k])],
-                            conflict_score=float(max_c[k]),
-                            mode="process",
-                        )
-                        .on_conflict_do_nothing()
+                    _insert_presentation_review(
+                        session,
+                        image_record_id=cids[k], tag_id=wip_id,
+                        conflict_tag_id=conf_tag_ids[int(arg_c[k])],
+                        conflict_score=float(max_c[k]),
+                        mode="process",
                    )
        if not dry_run:
            session.commit()
@@ -1062,7 +1065,7 @@ def retract_auto_applied_heads(session: Session) -> int:
            continue
        Xn = _l2norm(np.vstack([emb[i] for i in cids]).astype(np.float32), np)
        w = np.asarray(weights, dtype=np.float32)
-        probs = 1.0 / (1.0 + np.exp(-(Xn @ w + float(bias))))
+        probs = _sigmoid(Xn @ w + float(bias), np)
        below = [cids[k] for k in np.where(probs < float(thr))[0]]
        for iid in below:
            session.execute(
@@ -94,6 +94,24 @@ def _rejected_ids(session: Session, tag_id: int) -> list[int]:
    ]


+def _applied_or_rejected(session: Session, tag_ids) -> dict[int, set[int]]:
+    """Per-tag skip set for the auto-apply sweeps: every image that ALREADY carries
+    the tag (ANY source — not just training positives) OR has rejected it. A sweep
+    never re-applies to these. Shared by auto_apply_sweep + system_tag_auto_apply_sweep
+    (heads.py) and scheduled_ccip_auto_apply (tasks/ml.py). Callers mutate the returned
+    sets in-place to also dedupe within a single run."""
+    skip: dict[int, set[int]] = {}
+    for tid in tag_ids:
+        ids = {
+            r[0] for r in session.execute(
+                select(image_tag.c.image_record_id).where(image_tag.c.tag_id == tid)
+            ).all()
+        }
+        ids.update(_rejected_ids(session, tid))
+        skip[tid] = ids
+    return skip
+
+
 def _sample_unlabeled(session: Session, exclude: set[int], limit: int) -> list[int]:
    """Random image ids (with an embedding) NOT carrying the tag. Concepts are
    sparse, so an untagged image is almost always a true negative."""