Files
FabledCurator/tests/test_training_hygiene.py
T
bvandeusen e6f128c894
CI / lint (push) Successful in 3s
CI / frontend-build (push) Successful in 20s
CI / backend-lint-and-test (push) Successful in 34s
CI / integration (push) Failing after 4m28s
feat(ml): training hygiene — system-tagged images are absent from other concepts training
Step 2 of milestone #128. _hygiene_excluded_ids (training_data.py) is the
one shared predicate: images carrying any system tag are dropped from
every OTHER concepts head training — not positives (a rough wip tagged
as a character drags the head toward generic-sketch) and not rejection
or sampled negatives (a wip OF character X is not evidence against X).
A system tags own head trains on them unfiltered; that is what makes
auto-flagging banners work. Selection is split out of train_head as the
sklearn-free head_training_ids so CI (no sklearn) can pin the behavior.

CCIP: reference prototypes skip hygiene-tagged images — a faceless wip
figure region must never become an identity reference — and the ref
cache signature now counts hygiene applications, since tagging an image
wip changes the reference set without touching character/region counts.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CDgx8bQS5YrGRK76v8HUnM
2026-07-02 23:19:41 -04:00

126 lines
4.4 KiB
Python

"""Training hygiene (#128): system-tagged images are ABSENT from other
concepts' training data and from CCIP reference prototypes.
sklearn only exists in the ml image, so these pin head_training_ids (the
sklearn-free selection split out of train_head) rather than a full fit —
the exclusion lives entirely in that selection.
"""
import pytest
from sqlalchemy import insert, select
from backend.app.models import ImageRecord, ImageRegion, Tag, TagKind
from backend.app.models.tag import image_tag
from backend.app.services.ml import ccip
from backend.app.services.ml.heads import head_training_ids
from backend.app.services.ml.training_data import _hygiene_excluded_ids
from backend.app.services.tag_service import TagService
pytestmark = pytest.mark.integration
_CFG = {"min_positives": 2, "neg_ratio": 1}
async def _system_wip(db) -> Tag:
return (await db.execute(
select(Tag).where(Tag.is_system.is_(True), Tag.name == "wip")
)).scalar_one()
async def _img(db, sha, *, embedded=True):
rec = ImageRecord(
path=f"/images/{sha}.jpg", sha256=sha, size_bytes=1, mime="image/jpeg",
width=1, height=1, origin="imported_filesystem",
integrity_status="unknown",
siglip_embedding=([0.1] * 1152 if embedded else None),
)
db.add(rec)
await db.flush()
return rec
async def _apply(db, image_id, tag_id):
await db.execute(insert(image_tag).values(
image_record_id=image_id, tag_id=tag_id, source="manual",
))
@pytest.mark.asyncio
async def test_hygiene_excluded_ids(db):
wip = await _system_wip(db)
flagged = await _img(db, "a" * 64)
await _img(db, "b" * 64)
await _apply(db, flagged.id, wip.id)
excluded = await db.run_sync(lambda s: _hygiene_excluded_ids(s))
assert excluded == {flagged.id}
@pytest.mark.asyncio
async def test_head_selection_drops_hygiene_from_both_sides(db):
"""A wip-tagged image of concept X is neither a positive NOR a sampled
negative for X — it is absent entirely."""
wip = await _system_wip(db)
concept = await TagService(db).find_or_create("concept_x", TagKind.general)
clean_a = await _img(db, "a" * 64)
clean_b = await _img(db, "b" * 64)
flagged_pos = await _img(db, "c" * 64) # X + wip: dropped positive
negative_pool = await _img(db, "d" * 64) # untagged: legit negative
flagged_pool = await _img(db, "e" * 64) # wip only: must not be sampled
for img in (clean_a, clean_b, flagged_pos):
await _apply(db, img.id, concept.id)
await _apply(db, flagged_pos.id, wip.id)
await _apply(db, flagged_pool.id, wip.id)
ids = await db.run_sync(lambda s: head_training_ids(s, concept.id, _CFG))
assert ids is not None
pos_ids, neg_ids = ids
assert set(pos_ids) == {clean_a.id, clean_b.id}
assert negative_pool.id in neg_ids
assert flagged_pos.id not in neg_ids
assert flagged_pool.id not in neg_ids
@pytest.mark.asyncio
async def test_system_tags_own_head_keeps_hygiene_positives(db):
"""The wip head itself trains ON wip-tagged images — that's what makes
auto-flagging work."""
wip = await _system_wip(db)
one = await _img(db, "a" * 64)
two = await _img(db, "b" * 64)
await _img(db, "c" * 64) # negative pool
await _apply(db, one.id, wip.id)
await _apply(db, two.id, wip.id)
ids = await db.run_sync(lambda s: head_training_ids(s, wip.id, _CFG))
assert ids is not None
pos_ids, _ = ids
assert set(pos_ids) == {one.id, two.id}
@pytest.mark.asyncio
async def test_ccip_references_skip_hygiene_images(db):
"""A wip's figure region must never become an identity prototype, even on
a single-character image."""
ccip._REF_CACHE.update(sig=None, refs=None)
wip = await _system_wip(db)
char = await TagService(db).find_or_create("Char A", TagKind.character)
clean = await _img(db, "a" * 64)
flagged = await _img(db, "b" * 64)
for img, slot in ((clean, 0), (flagged, 1)):
vec = [0.0] * 768
vec[slot] = 1.0
db.add(ImageRegion(
image_record_id=img.id, kind="figure",
rx=0.0, ry=0.0, rw=1.0, rh=1.0,
ccip_embedding=vec, embedding_version="ccip-test",
))
await _apply(db, img.id, char.id)
await _apply(db, flagged.id, wip.id)
await db.commit()
refs = await ccip.character_references(db)
vectors = refs.get(char.id, [])
assert len(vectors) == 1
assert float(vectors[0][0]) == pytest.approx(1.0)
assert float(vectors[0][1]) == pytest.approx(0.0)