Merge pull request 'Crop-proposer detector config → Settings → Tagging (default-on, lease-announced) (#134)' (#194) from dev into main
Build images / sign-extension (push) Successful in 3s
CI / lint (push) Successful in 3s
Build images / build-agent (push) Successful in 8s
Build images / build-ml (push) Successful in 8s
Build images / build-web (push) Successful in 10s
CI / frontend-build (push) Successful in 21s
CI / backend-lint-and-test (push) Successful in 30s
CI / integration (push) Successful in 3m40s

This commit was merged in pull request #194.
This commit is contained in:
2026-07-05 20:10:02 -04:00
9 changed files with 466 additions and 9 deletions
+50 -9
View File
@@ -265,6 +265,8 @@ class Worker:
self._embedder = None
self._embedder_lock = threading.Lock()
self._proposers = None
self._proposers_sig = None # detector-config signature the current
# proposers were built for (#134)
self._proposers_lock = threading.Lock()
# --- held-lease bookkeeping --------------------------------------------
@@ -796,14 +798,52 @@ class Worker:
self._embedder = CropEmbedder(model_name, self.cfg.embed_dtype)
return self._embedder
def _ensure_proposers(self):
if self._proposers is not None:
return self._proposers
def _effective_cfg(self, det: dict | None):
"""The detector config for a job: the LEASE-ANNOUNCED block (the DB/UI
source of truth, #134) overlaid on the env cfg, or the env cfg unchanged
when the server announced nothing (bootstrap / older backend). An enabled
toggle off maps to empty weights — the 'weights == "" → proposer skipped'
contract in detectors.py. detector_level + ccip_model stay env (separate
settings, not part of the crop-proposer config)."""
if not det:
return self.cfg
def _w(p: str) -> str:
slot = det.get(p) or {}
return (slot.get("weights") or "") if slot.get("enabled", True) else ""
from dataclasses import replace
return replace(
self.cfg,
person_weights=_w("person"),
person_conf=float(det["person"]["conf"]),
anatomy_weights=_w("anatomy"),
anatomy_conf=float(det["anatomy"]["conf"]),
panel_weights=_w("panel"),
panel_conf=float(det["panel"]["conf"]),
max_figures=int(det["max_figures"]),
max_components=int(det["max_components"]),
max_panels=int(det["max_panels"]),
max_regions=int(det["max_regions"]),
dedupe_iou=float(det["dedupe_iou"]),
)
def _proposers_for(self, eff):
"""Proposers built from the effective detector config, rebuilt ONLY when
that config changes — so an operator's UI edit takes effect on the next
lease with no restart, and steady-state reuses the loaded YOLO models."""
sig = (
eff.person_weights, eff.person_conf,
eff.anatomy_weights, eff.anatomy_conf,
eff.panel_weights, eff.panel_conf,
eff.max_figures, eff.max_components, eff.max_panels,
)
with self._proposers_lock:
if self._proposers is None:
if self._proposers is None or self._proposers_sig != sig:
from .detectors import Proposers
self._proposers = Proposers(self.cfg)
return self._proposers
self._proposers = Proposers(eff)
self._proposers_sig = sig
return self._proposers
def _consume(self, job: dict, frames: list, stop_evt: threading.Event) -> bool:
"""Detect + embed the decoded frames and submit the result. Returns True
@@ -856,7 +896,8 @@ class Worker:
else ["figure", "face", "concept", "panel"]
)
embedder = self._ensure_embedder(model_name) if want_siglip else None
proposers = self._ensure_proposers()
eff = self._effective_cfg(job.get("detectors"))
proposers = self._proposers_for(eff)
regions = []
ccip_ev = self.cfg.ccip_model or "ccip-default"
@@ -922,7 +963,7 @@ class Worker:
# we never SigLIP the same region twice — saves GPU and a
# slot against max_regions. High-IoU + kind-aware, so
# intentional nested crops (figure ⊃ head) survive.
pending = dedupe_crops(pending, self.cfg.dedupe_iou)
pending = dedupe_crops(pending, eff.dedupe_iou)
vecs = embedder.embed_batch([c for c, _ in pending])
for (_c, tmpl), vec in zip(pending, vecs, strict=True):
tmpl["siglip_embedding"] = vec
@@ -932,7 +973,7 @@ class Worker:
# mp4, 64 sampled frames × ~32 regions) would otherwise burn
# ~38s of GPU across every frame before the submit is even
# truncated. Bounds the WORK, not just the POST body.
if len(regions) >= self.cfg.max_regions:
if len(regions) >= eff.max_regions:
break
# A Stop mid-frame-loop leaves partial regions — don't submit those;
@@ -0,0 +1,83 @@
"""ml_settings crop-proposer / detector config (#134)
Move the WHERE-to-crop detector config (per-proposer enable + weights + conf,
plus caps + dedupe IoU) into the DB so it's UI-tunable and announced to the GPU
agent in the lease (like the embedder model) — no restart, agent env is now
bootstrap-only. All server_defaults are the working values so existing rows +
fresh installs crop out-of-the-box with all three proposers ON.
Revision ID: 0078
Revises: 0077
Create Date: 2026-07-05
"""
from typing import Sequence, Union
import sqlalchemy as sa
from alembic import op
revision: str = "0078"
down_revision: Union[str, None] = "0077"
branch_labels: Union[str, Sequence[str], None] = None
depends_on: Union[str, Sequence[str], None] = None
_ANATOMY_DEFAULT = (
"https://github.com/aperveyev/booru_yolo/raw/main/models/yolov11m_aa22.pt"
)
_PANEL_DEFAULT = "mosesb/best-comic-panel-detection::best.pt"
def upgrade() -> None:
op.add_column("ml_settings", sa.Column(
"detector_person_enabled", sa.Boolean(), nullable=False,
server_default=sa.true()))
op.add_column("ml_settings", sa.Column(
"detector_person_weights", sa.String(512), nullable=False,
server_default="yolo11n.pt"))
op.add_column("ml_settings", sa.Column(
"detector_person_conf", sa.Float(), nullable=False,
server_default=sa.text("0.35")))
op.add_column("ml_settings", sa.Column(
"detector_anatomy_enabled", sa.Boolean(), nullable=False,
server_default=sa.true()))
op.add_column("ml_settings", sa.Column(
"detector_anatomy_weights", sa.String(512), nullable=False,
server_default=_ANATOMY_DEFAULT))
op.add_column("ml_settings", sa.Column(
"detector_anatomy_conf", sa.Float(), nullable=False,
server_default=sa.text("0.30")))
op.add_column("ml_settings", sa.Column(
"detector_panel_enabled", sa.Boolean(), nullable=False,
server_default=sa.true()))
op.add_column("ml_settings", sa.Column(
"detector_panel_weights", sa.String(512), nullable=False,
server_default=_PANEL_DEFAULT))
op.add_column("ml_settings", sa.Column(
"detector_panel_conf", sa.Float(), nullable=False,
server_default=sa.text("0.30")))
op.add_column("ml_settings", sa.Column(
"detector_max_figures", sa.Integer(), nullable=False,
server_default=sa.text("8")))
op.add_column("ml_settings", sa.Column(
"detector_max_components", sa.Integer(), nullable=False,
server_default=sa.text("8")))
op.add_column("ml_settings", sa.Column(
"detector_max_panels", sa.Integer(), nullable=False,
server_default=sa.text("8")))
op.add_column("ml_settings", sa.Column(
"detector_max_regions", sa.Integer(), nullable=False,
server_default=sa.text("128")))
op.add_column("ml_settings", sa.Column(
"detector_dedupe_iou", sa.Float(), nullable=False,
server_default=sa.text("0.85")))
def downgrade() -> None:
for col in (
"detector_person_enabled", "detector_person_weights", "detector_person_conf",
"detector_anatomy_enabled", "detector_anatomy_weights", "detector_anatomy_conf",
"detector_panel_enabled", "detector_panel_weights", "detector_panel_conf",
"detector_max_figures", "detector_max_components", "detector_max_panels",
"detector_max_regions", "detector_dedupe_iou",
):
op.drop_column("ml_settings", col)
+28
View File
@@ -269,6 +269,33 @@ async def lease():
).scalars()
} if ids else {}
await session.commit()
# Crop-proposer config, announced FROM THE SETTING like embed_model_name
# (#134): the agent builds its detectors from this, rebuilding live when
# it changes — so tuning is a DB/UI edit, never an agent restart. Same
# block for every job in the batch (it's global), built once. An enabled
# toggle off is carried through so the agent skips that proposer.
detectors = {
"person": {
"enabled": ml.detector_person_enabled,
"weights": ml.detector_person_weights,
"conf": ml.detector_person_conf,
},
"anatomy": {
"enabled": ml.detector_anatomy_enabled,
"weights": ml.detector_anatomy_weights,
"conf": ml.detector_anatomy_conf,
},
"panel": {
"enabled": ml.detector_panel_enabled,
"weights": ml.detector_panel_weights,
"conf": ml.detector_panel_conf,
},
"max_figures": ml.detector_max_figures,
"max_components": ml.detector_max_components,
"max_panels": ml.detector_max_panels,
"max_regions": ml.detector_max_regions,
"dedupe_iou": ml.detector_dedupe_iou,
}
out = []
for j in jobs:
img = imgs.get(j.image_record_id)
@@ -290,6 +317,7 @@ async def lease():
# re-embed, never an agent change.
"embed_model_name": ml.embedder_model_name,
"embed_version": ml.embedder_model_version,
"detectors": detectors,
})
return jsonify({"jobs": out})
+35
View File
@@ -8,6 +8,26 @@ from ..models import MLSettings
ml_admin_bp = Blueprint("ml_admin", __name__, url_prefix="/api/ml")
# Crop-proposer / detector config (#134). Announced to the GPU agent in the lease
# → tunable here with no restart. weights = ultralytics name | URL | hf_repo::file
# (empty, or enabled off, skips that proposer).
_DETECTOR_FIELDS = (
"detector_person_enabled",
"detector_person_weights",
"detector_person_conf",
"detector_anatomy_enabled",
"detector_anatomy_weights",
"detector_anatomy_conf",
"detector_panel_enabled",
"detector_panel_weights",
"detector_panel_conf",
"detector_max_figures",
"detector_max_components",
"detector_max_panels",
"detector_max_regions",
"detector_dedupe_iou",
)
_EDITABLE = (
"cpu_embed_enabled",
"video_frame_interval_seconds",
@@ -21,6 +41,7 @@ _EDITABLE = (
"ccip_auto_apply_threshold",
"embedder_model_name",
"embedder_model_version",
*_DETECTOR_FIELDS,
)
@@ -76,6 +97,7 @@ async def get_settings():
"ccip_auto_apply_enabled": s.ccip_auto_apply_enabled,
"ccip_auto_apply_threshold": s.ccip_auto_apply_threshold,
"embedder_model_name": s.embedder_model_name,
**{f: getattr(s, f) for f in _DETECTOR_FIELDS},
}
)
@@ -133,6 +155,19 @@ def _validate(p: dict) -> str | None:
for key in ("embedder_model_name", "embedder_model_version"):
if not str(p[key]).strip():
return f"{key} must not be empty"
# Crop proposers (#134). Weights may be empty (that proposer is just off);
# confidences are probabilities; caps are positive counts; IoU is [0,1].
for key in ("detector_person_conf", "detector_anatomy_conf", "detector_panel_conf"):
if not (0.0 <= float(p[key]) <= 1.0):
return f"{key} must be between 0 and 1"
for key in (
"detector_max_figures", "detector_max_components",
"detector_max_panels", "detector_max_regions",
):
if int(p[key]) < 1:
return f"{key} must be >= 1"
if not (0.0 <= float(p["detector_dedupe_iou"]) <= 1.0):
return "detector_dedupe_iou must be between 0 and 1"
return None
+64
View File
@@ -91,6 +91,70 @@ class MLSettings(Base):
embedder_model_name: Mapped[str] = mapped_column(
String(128), nullable=False, default="google/siglip2-so400m-patch16-512"
)
# -- Crop proposers / detectors (#1202, #134) --------------------------
# WHERE-to-crop YOLO detectors feeding the crop→SigLIP bag + CCIP. Config
# lives HERE (DB) and is announced to the GPU agent in the lease — same as
# the embedder model — so it is UI-tunable with NO restart, and the agent's
# env is bootstrap-only. Each weights spec is an ultralytics builtin name,
# an http(s) URL, or "hf_repo::file" (agent's _resolve). enabled off (or an
# empty weights) skips that proposer. All ON by default (operator 2026-07-05)
# so a fresh install crops out-of-the-box.
# person: general COCO figure detector for Western/realistic art the anime
# person-detector misses → NMS-merged with imgutils → CCIP + concept.
detector_person_enabled: Mapped[bool] = mapped_column(
Boolean, nullable=False, default=True
)
detector_person_weights: Mapped[str] = mapped_column(
String(512), nullable=False, default="yolo11n.pt"
)
detector_person_conf: Mapped[float] = mapped_column(
Float, nullable=False, default=0.35
)
# anatomy: booru_yolo anime/furry/NSFW torso components → concept crops.
# Default = yolov11m_aa22 (26 classes, best mAP50-95 0.96), committed in the
# upstream repo so the URL resolves. License UNSTATED — fine for a private
# homelab (operator accepted #1202).
detector_anatomy_enabled: Mapped[bool] = mapped_column(
Boolean, nullable=False, default=True
)
detector_anatomy_weights: Mapped[str] = mapped_column(
String(512), nullable=False,
default=(
"https://github.com/aperveyev/booru_yolo/raw/main/models/"
"yolov11m_aa22.pt"
),
)
detector_anatomy_conf: Mapped[float] = mapped_column(
Float, nullable=False, default=0.30
)
# panel: comic page → panel regions → concept crops (Apache-2.0, YOLOv12x).
detector_panel_enabled: Mapped[bool] = mapped_column(
Boolean, nullable=False, default=True
)
detector_panel_weights: Mapped[str] = mapped_column(
String(512), nullable=False,
default="mosesb/best-comic-panel-detection::best.pt",
)
detector_panel_conf: Mapped[float] = mapped_column(
Float, nullable=False, default=0.30
)
# Per-frame caps bound the crop→embed explosion; max_regions is the hard
# per-job backstop; dedupe_iou drops near-duplicate crops before the embed.
detector_max_figures: Mapped[int] = mapped_column(
Integer, nullable=False, default=8
)
detector_max_components: Mapped[int] = mapped_column(
Integer, nullable=False, default=8
)
detector_max_panels: Mapped[int] = mapped_column(
Integer, nullable=False, default=8
)
detector_max_regions: Mapped[int] = mapped_column(
Integer, nullable=False, default=128
)
detector_dedupe_iou: Mapped[float] = mapped_column(
Float, nullable=False, default=0.85
)
updated_at: Mapped[datetime] = mapped_column(
DateTime(timezone=True), nullable=False, server_default=func.now()
)
@@ -0,0 +1,151 @@
<template>
<MaintenanceTile
icon="mdi-crop"
title="Crop proposers (where to crop)"
blurb="Detectors that decide WHERE to crop each image before embedding — general figures, anime/furry anatomy, and comic panels — so a tag can ground to a region, not just the whole image."
:open="false"
>
<p class="fc-muted text-body-2 mb-4">
Each proposer finds regions that the crop SigLIP bag scores alongside the
whole image; the best-scoring region is what grounds a tag. Weights accept
an <strong>ultralytics name</strong> (<code>yolo11n.pt</code>), a
<strong>URL</strong>, or <strong>hf_repo::file</strong>. Edits reach the GPU
agent on its next lease no restart. After turning a proposer on, re-process
the library so existing images gain its regions.
</p>
<div v-for="p in proposers" :key="p.key" class="fc-proposer">
<div class="d-flex align-center mb-1" style="gap: 10px;">
<v-icon size="18" :color="p.on ? 'accent' : undefined">{{ p.icon }}</v-icon>
<span class="fc-section-h">{{ p.label }}</span>
<v-switch
v-model="p.on" :loading="busy" hide-details density="compact"
color="success" class="ml-auto"
@update:model-value="v => saveToggle(p, v)"
/>
</div>
<p class="fc-muted text-body-2 mb-2">{{ p.help }}</p>
<div class="d-flex flex-wrap mb-4" style="gap: 12px;">
<v-text-field
v-model="p.weights" label="Weights" density="compact" hide-details
style="min-width: 300px; flex: 1;" :disabled="busy || !p.on"
placeholder="name | URL | hf_repo::file"
@change="save({ [`detector_${p.key}_weights`]: p.weights })"
/>
<v-text-field
v-model.number="p.conf" label="Confidence" type="number"
min="0" max="1" step="0.05" density="compact" hide-details
style="max-width: 140px;" :disabled="busy || !p.on"
@change="save({ [`detector_${p.key}_conf`]: Number(p.conf) })"
/>
</div>
</div>
<div class="fc-caps">
<div class="fc-section-h mb-1">Caps</div>
<p class="fc-muted text-body-2 mb-3">
Bound how many crops each image can produce — protects GPU time and
storage. Dedupe IoU drops near-duplicate crops before embedding.
</p>
<div class="d-flex flex-wrap" style="gap: 12px;">
<v-text-field
v-for="c in caps" :key="c.key"
v-model.number="c.val" :label="c.label" type="number"
:min="c.min" :max="c.max" :step="c.step || 1" density="compact"
hide-details style="max-width: 165px;" :disabled="busy"
@change="save({ [c.key]: Number(c.val) })"
/>
</div>
</div>
</MaintenanceTile>
</template>
<script setup>
import { toast } from '../../utils/toast.js'
import { onMounted, ref } from 'vue'
import MaintenanceTile from '../common/MaintenanceTile.vue'
import { useMLStore } from '../../stores/ml.js'
const mlSettings = useMLStore()
const busy = ref(false)
const proposers = ref([])
const caps = ref([])
// Static copy per proposer; the mutable state (on/weights/conf) is seeded from
// settings on mount so the card reflects the live DB config (rule 26 — it works
// untouched).
const DEFS = [
{
key: 'person', label: 'Figures (person)', icon: 'mdi-account-outline',
help: 'General figure detector for the Western/realistic art the anime ' +
'detector misses → character identity + a whole-figure crop.',
},
{
key: 'anatomy', label: 'Anatomy (booru)', icon: 'mdi-dog-side',
help: 'Anime / furry / NSFW torso components (heads, busts, hips, …) → ' +
'localized concept crops.',
},
{
key: 'panel', label: 'Comic panels', icon: 'mdi-view-grid-outline',
help: 'Splits a comic page into its panels → one concept crop each.',
},
]
const CAP_DEFS = [
{ key: 'detector_max_figures', label: 'Max figures', min: 1 },
{ key: 'detector_max_components', label: 'Max components', min: 1 },
{ key: 'detector_max_panels', label: 'Max panels', min: 1 },
{ key: 'detector_max_regions', label: 'Max regions / job', min: 1 },
{ key: 'detector_dedupe_iou', label: 'Dedupe IoU', min: 0, max: 1, step: 0.05 },
]
onMounted(async () => {
try { await mlSettings.loadSettings() } catch { /* non-fatal — defaults show */ }
const s = mlSettings.settings || {}
proposers.value = DEFS.map(d => ({
...d,
on: !!s[`detector_${d.key}_enabled`],
weights: s[`detector_${d.key}_weights`] ?? '',
conf: s[`detector_${d.key}_conf`] ?? 0.3,
}))
caps.value = CAP_DEFS.map(c => ({ ...c, val: s[c.key] ?? 0 }))
})
async function save(patch, revert) {
busy.value = true
try {
await mlSettings.patchSettings(patch)
toast({ text: 'Saved', type: 'success' })
} catch (e) {
if (revert) revert()
toast({ text: `Could not save: ${e.message}`, type: 'error' })
} finally {
busy.value = false
}
}
function saveToggle (p, v) {
// Revert the switch on failure so it never lies about the persisted state.
save({ [`detector_${p.key}_enabled`]: !!v }, () => { p.on = !v })
}
</script>
<style scoped>
.fc-muted { color: rgb(var(--v-theme-on-surface-variant)); }
.fc-section-h {
font-size: 13px; font-weight: 700; letter-spacing: 0.03em;
text-transform: uppercase; color: rgb(var(--v-theme-on-surface));
}
.fc-proposer {
border-top: 1px solid rgb(var(--v-theme-surface-light)); padding-top: 14px;
}
.fc-proposer:first-of-type { border-top: none; padding-top: 0; }
.fc-caps {
border-top: 1px solid rgb(var(--v-theme-surface-light)); padding-top: 16px;
}
code {
font-family: 'JetBrains Mono', monospace; font-size: 0.85em;
background: rgb(var(--v-theme-surface-light)); padding: 1px 5px;
border-radius: 4px;
}
</style>
@@ -36,6 +36,7 @@
</p>
<div class="fc-tile-stack">
<MLThresholdSliders />
<CropProposersCard />
<HeadsCard />
<AliasTable />
</div>
@@ -75,6 +76,7 @@ import MissingFileRepairCard from './MissingFileRepairCard.vue'
import GpuTriageCard from './GpuTriageCard.vue'
import DbMaintenanceCard from './DbMaintenanceCard.vue'
import MLThresholdSliders from './MLThresholdSliders.vue'
import CropProposersCard from './CropProposersCard.vue'
import HeadsCard from './HeadsCard.vue'
import GpuAgentCard from './GpuAgentCard.vue'
import AliasTable from './AliasTable.vue'
+24
View File
@@ -105,6 +105,30 @@ async def test_lease_announces_embed_model_then_submit_embedding(client, db):
assert img.siglip_embedding is not None and len(list(img.siglip_embedding)) == 1152
@pytest.mark.asyncio
async def test_lease_announces_detector_config(client, db):
# #134: the lease carries the crop-proposer config from MLSettings, so the
# agent builds its detectors from the DB/UI (no restart). Defaults are all-on
# with the pinned working weights.
img = await _img(db, "c" * 64)
await GpuJobService(db).enqueue(img.id, "embed")
await db.commit()
token = (await (await client.post("/api/gpu/token/rotate")).get_json())["token"]
hdr = {"Authorization": f"Bearer {token}"}
leased = await client.post(
"/api/gpu/jobs/lease", json={"agent_id": "a1", "batch_size": 5}, headers=hdr,
)
det = (await leased.get_json())["jobs"][0]["detectors"]
assert det["person"]["enabled"]
assert det["anatomy"]["enabled"]
assert det["panel"]["enabled"]
assert "yolov11m_aa22" in det["anatomy"]["weights"] # booru_yolo default
assert det["panel"]["weights"].endswith("::best.pt") # mosesb default
assert det["max_regions"] == 128
assert det["dedupe_iou"] == 0.85
@pytest.mark.asyncio
async def test_submit_with_stale_lease_is_409(client, db):
img = await _img(db, "b" * 64)
+29
View File
@@ -52,6 +52,35 @@ async def test_embedder_model_default_settable_and_empty_rejected(client):
assert bad.status_code == 400
@pytest.mark.asyncio
async def test_detector_settings_defaults_patch_and_validation(client):
# #134: crop-proposer config is exposed + editable here (announced to the
# agent via the lease). Defaults are all-on with the pinned weights.
body = await (await client.get("/api/ml/settings")).get_json()
assert body["detector_person_enabled"] is True
assert body["detector_anatomy_enabled"] is True
assert "yolov11m_aa22" in body["detector_anatomy_weights"]
assert body["detector_panel_weights"].endswith("::best.pt")
assert body["detector_max_regions"] == 128
ok = await client.patch("/api/ml/settings", json={
"detector_anatomy_enabled": False,
"detector_person_conf": 0.5,
"detector_max_panels": 4,
})
assert ok.status_code == 200
out = await ok.get_json()
assert out["detector_anatomy_enabled"] is False
assert out["detector_person_conf"] == 0.5
assert out["detector_max_panels"] == 4
# Out-of-range confidence + non-positive cap are rejected.
assert (await client.patch(
"/api/ml/settings", json={"detector_panel_conf": 1.5})).status_code == 400
assert (await client.patch(
"/api/ml/settings", json={"detector_max_regions": 0})).status_code == 400
@pytest.mark.asyncio
async def test_embedder_models_list(client):
# #1203: the dropdown reads the supported-model list from the server.