Merge pull request 'Video tag quality: cadence sampling + min-frame aggregation + ML thread cap (#747)' (#111) from dev into main

alembic/versions/0053_ml_settings_video_tagging.py

@@ -0,0 +1,49 @@
+"""ml_settings: video tagging knobs (cadence sampling + noise floor)
+
+#747. Video tag quality/perf: sample frames at a fixed cadence (interval) so a
+tag's frame-presence reflects real screen time, cap total frames so long videos
+stay bounded, and keep a tag only if it appears in >= min_tag_frames sampled
+frames. Operator-tunable via Settings → ML (replaces the VIDEO_ML_FRAMES env var).
+
+Revision ID: 0053
+Revises: 0052
+Create Date: 2026-06-16
+"""
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "0053"
+down_revision: Union[str, None] = "0052"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.add_column(
+        "ml_settings",
+        sa.Column(
+            "video_frame_interval_seconds", sa.Float(), nullable=False,
+            server_default="4.0",
+        ),
+    )
+    op.add_column(
+        "ml_settings",
+        sa.Column(
+            "video_max_frames", sa.Integer(), nullable=False, server_default="64",
+        ),
+    )
+    op.add_column(
+        "ml_settings",
+        sa.Column(
+            "video_min_tag_frames", sa.Integer(), nullable=False,
+            server_default="3",
+        ),
+    )
+
+
+def downgrade() -> None:
+    op.drop_column("ml_settings", "video_min_tag_frames")
+    op.drop_column("ml_settings", "video_max_frames")
+    op.drop_column("ml_settings", "video_frame_interval_seconds")

backend/app/api/ml_admin.py

@@ -14,6 +14,9 @@ _EDITABLE = (
     "centroid_similarity_threshold",
     "min_reference_images",
     "tagger_store_floor",
+    "video_frame_interval_seconds",
+    "video_max_frames",
+    "video_min_tag_frames",
 )
 
 
@@ -32,6 +35,9 @@ async def get_settings():
             "centroid_similarity_threshold": s.centroid_similarity_threshold,
             "min_reference_images": s.min_reference_images,
             "tagger_store_floor": s.tagger_store_floor,
+            "video_frame_interval_seconds": s.video_frame_interval_seconds,
+            "video_max_frames": s.video_max_frames,
+            "video_min_tag_frames": s.video_min_tag_frames,
             "tagger_model_version": s.tagger_model_version,
             "embedder_model_version": s.embedder_model_version,
         }
@@ -85,6 +91,15 @@ def _validate(p: dict) -> str | None:
                 f"suggestion_threshold_{cat} cannot be below tagger_store_floor "
                 f"({floor}) — predictions below the floor are not stored"
             )
+    # Video tagging (#747).
+    if p["video_frame_interval_seconds"] <= 0:
+        return "video_frame_interval_seconds must be > 0"
+    if p["video_max_frames"] < 1:
+        return "video_max_frames must be >= 1"
+    if p["video_min_tag_frames"] < 1:
+        return "video_min_tag_frames must be >= 1"
+    if p["video_min_tag_frames"] > p["video_max_frames"]:
+        return "video_min_tag_frames cannot exceed video_max_frames"
     return None
 
 

backend/app/models/ml_settings.py

@@ -40,6 +40,21 @@ class MLSettings(Base):
     min_reference_images: Mapped[int] = mapped_column(
         Integer, nullable=False, default=5
     )
+    # Video tagging (#747). Sample one frame every N seconds (fixed CADENCE, not a
+    # fixed count) so a tag's frame-presence reflects real screen time regardless
+    # of video length; cap the total so a long video can't explode into hundreds
+    # of inferences (the cadence stretches past the cap). A tag is kept only if it
+    # appears in >= video_min_tag_frames sampled frames (≈ that many × interval
+    # seconds on screen) — duration-independent noise rejection. Operator-tunable.
+    video_frame_interval_seconds: Mapped[float] = mapped_column(
+        Float, nullable=False, default=4.0
+    )
+    video_max_frames: Mapped[int] = mapped_column(
+        Integer, nullable=False, default=64
+    )
+    video_min_tag_frames: Mapped[int] = mapped_column(
+        Integer, nullable=False, default=3
+    )
     tagger_model_version: Mapped[str] = mapped_column(
         String(128), nullable=False, default="camie-tagger-v2"
     )

backend/app/services/ml/embedder.py

@@ -1,8 +1,8 @@
 """SigLIP SO400M image-embedding wrapper (PyTorch CPU).
 
-Direct port of ImageRepo's siglip.py. torch/transformers are imported
+torch/transformers are imported lazily inside load() so this module can be
-lazily inside load() so this module can be imported in the web container
+imported in the web container (which never runs inference) without paying the
-(which never runs inference) without paying the torch import cost.
+torch import cost.
 """
 
 import os
@@ -13,6 +13,11 @@ from PIL import Image, ImageFile
 
 ImageFile.LOAD_TRUNCATED_IMAGES = True
 
+# Cap torch's intra-op threads so each ml-worker replica is a bounded core
+# consumer on a shared node (torch otherwise uses all cores). Keep
+# N_replicas × this within the cores allotted to ML to avoid oversubscription.
+_INTRA_OP_THREADS = 4
+
 MODEL_NAME = os.environ.get(
     "SIGLIP_MODEL_NAME", "google/siglip-so400m-patch14-384"
 )
@@ -37,6 +42,9 @@ class Embedder:
         from transformers import AutoModel, SiglipImageProcessor
 
         self._torch = torch
+        # Bound torch's CPU thread pool (see _INTRA_OP_THREADS) so each replica
+        # stays a predictable core consumer on a shared node.
+        torch.set_num_threads(_INTRA_OP_THREADS)
         # FC's embedder only does IMAGE inference — never text. AutoProcessor
         # loads the full processor including SiglipTokenizer, which requires
         # the sentencepiece library at import time even if we never call it.

backend/app/services/ml/tagger.py

@@ -1,8 +1,8 @@
-"""Camie-tagger-v2 ONNX wrapper.
+"""Camie-tagger-v2 ONNX wrapper (CPU).
 
-CPU-only, single-image at a time. Loaded lazily inside the ml-worker
+Single-image at a time. Loaded lazily inside the ml-worker process; NOT
-process; NOT thread-safe — the ml queue worker must run --concurrency=1
+thread-safe — the ml queue worker runs --concurrency=1 per process (scale ML by
-(set by the FC-1 entrypoint).
+running multiple worker replicas, not threads).
 
 v2 layout reference: HuggingFace Camais03/camie-tagger-v2 root has
 camie-tagger-v2.onnx (789 MB) + camie-tagger-v2-metadata.json (7.77 MB)
@@ -19,6 +19,11 @@ from pathlib import Path
 import numpy as np
 from PIL import Image, ImageFile
 
+# Cap inference threads (see Tagger.load) so each ml-worker replica is a bounded
+# core consumer on a shared node — keep N_replicas × this within the cores
+# allotted to ML so replicas don't oversubscribe the box / starve the DB.
+_INTRA_OP_THREADS = 4
+
 # onnxruntime lives in requirements-ml.txt only — it is NOT installed in the
 # lean web image or in CI. Imported lazily inside Tagger.load() so this module
 # imports fine without it (the suggestion service imports SURFACED_CATEGORIES
@@ -117,8 +122,15 @@ class Tagger:
         # without onnxruntime (CI / lean web image).
         import onnxruntime as ort
 
+        # Cap the intra-op thread pool. ONNX Runtime otherwise sizes it to ALL
+        # host cores, so on a shared node each ml-worker replica would grab every
+        # core and oversubscribe (and starve the co-located DB/web). Bounding it
+        # makes each replica a predictable core consumer — run N replicas where
+        # N × _INTRA_OP_THREADS stays within the cores you allot to ML.
+        opts = ort.SessionOptions()
+        opts.intra_op_num_threads = _INTRA_OP_THREADS
         session = ort.InferenceSession(
-            str(model_path), providers=["CPUExecutionProvider"]
+            str(model_path), sess_options=opts, providers=["CPUExecutionProvider"],
         )
         self._input_name = session.get_inputs()[0].name
         # Assign sentinels last so a partial load isn't observable.

backend/app/tasks/ml.py

@@ -49,11 +49,12 @@ def tag_and_embed(self, image_id: int) -> dict:
     """Run Camie + SigLIP on one image; store predictions + embedding;
     then enqueue per-image allowlist application.
 
-    Video: sample frames between 10% and 90% of duration (VIDEO_ML_FRAMES,
+    Video (#747): sample frames at a fixed cadence (ml_settings
-    default 6). Max-pool tagger confidences across frames, mean-pool the
+    video_frame_interval_seconds, capped at video_max_frames), keep a tag only if
+    it appears in >= video_min_tag_frames frames and average its confidence over
+    those frames (mean-pool, not max — kills one-frame noise); mean-pool the
     SigLIP embeddings. On no-frames returns status='no_frames' (not an error).
     """
-    import os
     import time
 
     from ..services.ml.embedder import get_embedder
@@ -116,7 +117,9 @@ def tag_and_embed(self, image_id: int) -> dict:
                 phase = "video_sample_frames"
                 t0 = time.monotonic()
                 frames = _sample_video_frames(
-                    src, int(os.environ.get("VIDEO_ML_FRAMES", "6"))
+                    src,
+                    interval=settings.video_frame_interval_seconds,
+                    max_frames=settings.video_max_frames,
                 )
                 log.info(
                     "tag_and_embed sampled %d frame(s) in %.1fs: %s",
@@ -127,13 +130,19 @@ def tag_and_embed(self, image_id: int) -> dict:
                 phase = "video_infer"
                 import numpy as np
 
-                preds = _maxpool_predictions(
+                preds = _aggregate_video_predictions(
                     [tagger.infer(f, store_floor=settings.tagger_store_floor)
-                     for f in frames]
+                     for f in frames],
+                    min_frames=settings.video_min_tag_frames,
                 )
                 embedding = np.mean(
                     [embedder.infer(f) for f in frames], axis=0
                 ).astype("float32")
+                log.info(
+                    "tag_and_embed video aggregated %d tag(s) from %d frame(s) "
+                    "(min_frames=%d): %s",
+                    len(preds), len(frames), settings.video_min_tag_frames, ctx,
+                )
                 for f in frames:
                     f.unlink(missing_ok=True)
             else:
@@ -208,9 +217,17 @@ def tag_and_embed(self, image_id: int) -> dict:
     return {"status": "ok", "image_id": image_id, "tags": len(preds)}
 
 
-def _sample_video_frames(src: Path, n: int) -> list[Path]:
+def _sample_video_frames(
-    """Extract n frames evenly between 10% and 90% of duration via ffmpeg.
+    src: Path, *, interval: float, max_frames: int,
-    Returns temp file paths (caller deletes). Empty list on failure."""
+) -> list[Path]:
+    """Sample frames at a fixed CADENCE — one every `interval` seconds — so a
+    tag's frame-presence reflects real screen time regardless of video length
+    (#747). The count is capped at `max_frames`: a video longer than
+    interval×max_frames stretches the spacing instead of exploding the frame
+    count (keeps cost bounded so a long video can't hog the single ml-worker).
+    Frames are taken across the 5%–95% window (skip intro/outro black/cards) via
+    per-frame fast-seek. Returns temp file paths (caller deletes); [] on failure.
+    """
     import json
     import subprocess
     import tempfile
@@ -229,20 +246,25 @@ def _sample_video_frames(src: Path, n: int) -> list[Path]:
     if duration <= 0:
         return []
 
-    start, end = duration * 0.10, duration * 0.90
+    start, end = duration * 0.05, duration * 0.95
-    step = (end - start) / max(n - 1, 1)
+    span = max(end - start, 0.0)
+    # Cadence count, clamped to [1, max_frames]. int(span/interval)+1 ≈ one frame
+    # per `interval` seconds across the window; the cap stretches spacing on very
+    # long videos.
+    n = max(1, min(int(span / interval) + 1, max(1, max_frames)))
+    step = span / max(n - 1, 1)
     out: list[Path] = []
     tmpdir = Path(tempfile.mkdtemp(prefix="fc_vid_"))
     for i in range(n):
         ts = start + i * step
-        dest = tmpdir / f"frame_{i:02d}.jpg"
+        dest = tmpdir / f"frame_{i:04d}.jpg"
         try:
             subprocess.run(
                 [
                     "ffmpeg", "-ss", f"{ts:.2f}", "-i", str(src),
                     "-frames:v", "1", "-q:v", "3", "-y", str(dest),
                 ],
-                check=True, capture_output=True, timeout=60,
+                check=True, capture_output=True, timeout=30,
             )
             if dest.is_file():
                 out.append(dest)
@@ -251,18 +273,38 @@ def _sample_video_frames(src: Path, n: int) -> list[Path]:
     return out
 
 
-def _maxpool_predictions(per_frame: list[dict]) -> dict:
+def _aggregate_video_predictions(per_frame: list[dict], *, min_frames: int) -> dict:
-    """Aggregate per-frame {name: TagPrediction} dicts by max confidence."""
+    """Aggregate per-frame {name: TagPrediction} into one prediction set (#747).
-    merged: dict[str, dict] = {}
+
+    A tag is kept only if it appears (≥ the tagger store floor, already applied)
+    in at least `min_frames` of the sampled frames — because sampling is at a
+    fixed cadence, that means it was on screen for roughly min_frames×interval
+    seconds, so a single-frame flicker / scene-transition artifact is dropped
+    while a genuine scene-local tag in a long video survives. Confidence is the
+    MEAN over the frames where the tag appears (not max — max re-inflated the
+    one-frame noise this whole change exists to remove).
+
+    `min_frames` is clamped to the number of frames actually sampled so a very
+    short video (1–2 frames) still tags instead of dropping everything.
+    """
+    n = len(per_frame)
+    if n == 0:
+        return {}
+    threshold = max(1, min(min_frames, n))
+    agg: dict[str, dict] = {}
     for frame_preds in per_frame:
         for name, p in frame_preds.items():
-            cur = merged.get(name)
+            cur = agg.get(name)
-            if cur is None or p.confidence > cur["confidence"]:
+            if cur is None:
-                merged[name] = {
+                agg[name] = {"category": p.category, "sum": p.confidence, "count": 1}
-                    "category": p.category,
+            else:
-                    "confidence": p.confidence,
+                cur["sum"] += p.confidence
-                }
+                cur["count"] += 1
-    return merged
+    return {
+        name: {"category": v["category"], "confidence": v["sum"] / v["count"]}
+        for name, v in agg.items()
+        if v["count"] >= threshold
+    }
 
 
 @celery.task(name="backend.app.tasks.ml.backfill", bind=True)

frontend/src/components/settings/MLThresholdSliders.vue

@@ -30,6 +30,39 @@
           </div>
         </v-col>
       </v-row>
+
+      <v-divider class="my-4" />
+
+      <div class="text-subtitle-2 mb-1">Video tagging</div>
+      <div class="text-caption fc-muted mb-3">
+        Videos are tagged by sampling frames at a fixed cadence. A tag is kept
+        only if it shows up in enough frames (≈ that many × the interval in
+        seconds of screen time), which filters one-frame noise without losing
+        tags that only appear in part of a longer video.
+      </div>
+      <v-row>
+        <v-col cols="12" sm="4">
+          <v-text-field
+            v-model.number="local.video_frame_interval_seconds"
+            label="Frame interval (s)" type="number" min="0.5" step="0.5"
+            density="comfortable" hide-details @change="save"
+          />
+        </v-col>
+        <v-col cols="12" sm="4">
+          <v-text-field
+            v-model.number="local.video_max_frames"
+            label="Max frames" type="number" min="1" step="1"
+            density="comfortable" hide-details @change="save"
+          />
+        </v-col>
+        <v-col cols="12" sm="4">
+          <v-text-field
+            v-model.number="local.video_min_tag_frames"
+            label="Min frames per tag" type="number" min="1" step="1"
+            density="comfortable" hide-details @change="save"
+          />
+        </v-col>
+      </v-row>
     </v-card-text>
     <v-card-text v-else><v-skeleton-loader type="paragraph" /></v-card-text>
   </v-card>
@@ -59,9 +92,14 @@ async function save() {
   const floor = local.tagger_store_floor
   local.suggestion_threshold_character = Math.max(local.suggestion_threshold_character, floor)
   local.suggestion_threshold_general = Math.max(local.suggestion_threshold_general, floor)
+  // Mirror the server invariant: a tag can't require more frames than are sampled.
+  local.video_min_tag_frames = Math.min(local.video_min_tag_frames, local.video_max_frames)
   const patch = {}
   for (const f of fields) patch[f.key] = local[f.key]
   patch.tagger_store_floor = local.tagger_store_floor
+  patch.video_frame_interval_seconds = local.video_frame_interval_seconds
+  patch.video_max_frames = local.video_max_frames
+  patch.video_min_tag_frames = local.video_min_tag_frames
   try { await store.patchSettings(patch) }
   catch (e) { toast({ text: e.message, type: 'error' }) }
 }

tests/test_api_ml_admin.py

@@ -56,6 +56,36 @@ async def test_suggestion_threshold_below_store_floor_rejected(client):
     assert "tagger_store_floor" in (await resp.get_json())["error"]
 
 
+@pytest.mark.asyncio
+async def test_video_tagging_settings_default_and_patch(client):
+    """#747: video cadence/noise knobs are exposed + patchable."""
+    body = await (await client.get("/api/ml/settings")).get_json()
+    assert body["video_frame_interval_seconds"] == pytest.approx(4.0)
+    assert body["video_max_frames"] == 64
+    assert body["video_min_tag_frames"] == 3
+
+    resp = await client.patch(
+        "/api/ml/settings",
+        json={"video_frame_interval_seconds": 5, "video_max_frames": 40,
+              "video_min_tag_frames": 4},
+    )
+    assert resp.status_code == 200
+    out = await resp.get_json()
+    assert out["video_frame_interval_seconds"] == pytest.approx(5.0)
+    assert out["video_max_frames"] == 40
+    assert out["video_min_tag_frames"] == 4
+
+
+@pytest.mark.asyncio
+async def test_video_min_tag_frames_above_max_rejected(client):
+    resp = await client.patch(
+        "/api/ml/settings",
+        json={"video_max_frames": 10, "video_min_tag_frames": 20},
+    )
+    assert resp.status_code == 400
+    assert "video_min_tag_frames" in (await resp.get_json())["error"]
+
+
 @pytest.mark.asyncio
 async def test_backfill_and_recompute_trigger(client):
     r1 = await client.post("/api/ml/backfill")

tests/test_tasks_ml.py

@@ -1,6 +1,6 @@
 """tag_and_embed / backfill task tests. Models aren't in CI, so we test
-the pure helpers (_maxpool_predictions, _is_video) as unit tests, and the
+the pure helpers (_aggregate_video_predictions, _is_video) as unit tests, and
-DB-touching backfill query as an integration test with monkeypatched
+the DB-touching backfill query as an integration test with monkeypatched
 inference.
 """
 
@@ -9,7 +9,7 @@ from pathlib import Path
 import pytest
 
 from backend.app.services.ml.tagger import TagPrediction
-from backend.app.tasks.ml import _is_video, _maxpool_predictions
+from backend.app.tasks.ml import _aggregate_video_predictions, _is_video
 
 
 def test_is_video():
@@ -18,15 +18,32 @@ def test_is_video():
     assert _is_video(Path("a.jpg")) is False
 
 
-def test_maxpool_predictions():
+def _pred(name, conf, cat="general"):
-    f1 = {"smile": TagPrediction("smile", "general", 0.6)}
+    return {name: TagPrediction(name, cat, conf)}
-    f2 = {
+
-        "smile": TagPrediction("smile", "general", 0.9),
+
-        "sword": TagPrediction("sword", "general", 0.7),
+def test_aggregate_video_keeps_corroborated_and_means():
-    }
+    # #747: 4 frames; "smile" in 3, "sword" in 1 (noise). min_frames=2.
-    merged = _maxpool_predictions([f1, f2])
+    per_frame = [
-    assert merged["smile"]["confidence"] == 0.9
+        {"smile": TagPrediction("smile", "general", 0.6),
-    assert merged["sword"]["confidence"] == 0.7
+         "sword": TagPrediction("sword", "general", 0.9)},
+        _pred("smile", 0.8),
+        _pred("smile", 0.7),
+        {},
+    ]
+    out = _aggregate_video_predictions(per_frame, min_frames=2)
+    assert "sword" not in out  # one-frame flicker dropped
+    assert abs(out["smile"]["confidence"] - (0.6 + 0.8 + 0.7) / 3) < 1e-9  # mean, not max
+
+
+def test_aggregate_video_clamps_min_frames_to_sample_count():
+    # Short video: 1 frame but min_frames=3 — clamp so it still tags.
+    out = _aggregate_video_predictions([_pred("solo", 0.8)], min_frames=3)
+    assert out["solo"]["confidence"] == 0.8
+
+
+def test_aggregate_video_empty():
+    assert _aggregate_video_predictions([], min_frames=3) == {}
 
 
 @pytest.mark.integration