feat(trip-cluster): cluster orchestration (seed derivation + coverage) + CLI

2026-06-27 17:28:27 +02:00
parent a9b443cd53
commit 5256df4117
3 changed files with 131 additions and 1 deletions
@@ -0,0 +1,57 @@
+from collections import defaultdict
+
+from photoflow.core.models import Cluster
+from photoflow.immich import pipeline
+from app.clustering import cluster_assets, _epoch
+from app.coverage import coverage_members
+
+# A tag seeds only if its photos dominate the time window they span: among all
+# assets whose taken_at falls in [min, max] of the tag's photos, at least this
+# fraction must carry the tag. A name/event tag sprinkled among many untagged
+# photos in the same window fails this gate.
+SEED_DOMINANCE_MIN = 0.5
+
+
+def _seed_tags(assets, tags_by_asset, max_span_days: int,
+               dominance_min: float = SEED_DOMINANCE_MIN) -> set:
+    epochs = {a.immich_id: _epoch(a.taken_at) for a in assets}
+    tagged = defaultdict(list)  # tag -> epochs of its assets
+    for a in assets:
+        for t in tags_by_asset.get(a.immich_id, []):
+            if not pipeline.is_pipeline_tag(t):
+                tagged[t].append(epochs[a.immich_id])
+
+    seeds = set()
+    for tag, ts in tagged.items():
+        if len(ts) < 2:
+            continue
+        lo, hi = min(ts), max(ts)
+        # Loose safety cap only — reject absurdly long spans, but keep genuine
+        # long trips (this library has one > 60 days), so the default is high.
+        if hi - lo > max_span_days * 86400:
+            continue
+        # Dominance / temporal contiguity gate.
+        in_window = sum(1 for e in epochs.values() if lo <= e <= hi)
+        if in_window and len(ts) / in_window >= dominance_min:
+            seeds.add(tag)
+    return seeds
+
+
+def run_cluster(store, *, gap_factor: float = 6.0, seed_max_span_days: int = 400) -> dict:
+    assets = store.all_assets(include_processed=False)
+    tags_by_asset = {a.immich_id: store.asset_tags(a.immich_id) for a in assets}
+    seed_tags = _seed_tags(assets, tags_by_asset, seed_max_span_days)
+    candidates = cluster_assets(assets, tags_by_asset, seed_tags, gap_factor=gap_factor)
+    assets_by_id = {a.immich_id: a for a in assets}
+
+    store.clear_clusters()
+    n = 0
+    for cand in candidates:
+        members = coverage_members(cand, assets_by_id)
+        store.insert_cluster(Cluster(
+            start_at=cand.start_at, end_at=cand.end_at,
+            count=sum(1 for m in members if m.included),
+            suggested_name=cand.suggested_name, confidence=cand.confidence,
+            kind_guess=cand.kind_guess, status="pending"), members)
+        n += 1
+    return {"clusters": n}