From 65fee745859c9cf36eb22daf43b0541a5312e4d3 Mon Sep 17 00:00:00 2001
From: Duncan Tourolle <duncan@tourolle.paris>
Date: Sat, 4 Jul 2026 20:41:54 +0200
Subject: [PATCH] perf(movienet): vectorise eval matching; count frames missing
 from Image.zip
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

movienet_eval: replace the per-element dot() with numpy — actor references are
loaded once as an ndarray and scored with a single matmul, keeping a
whole-library gallery fast.

movienet_prep: count and report frames referenced by annotations but absent
from Image.zip instead of skipping them silently.
---
 scripts/movienet_eval.py | 26 +++++++++++++++-----------
 scripts/movienet_prep.py | 14 +++++++++-----
 2 files changed, 24 insertions(+), 16 deletions(-)

diff --git a/scripts/movienet_eval.py b/scripts/movienet_eval.py
index 4a40b6c..9bfd6f5 100644
--- a/scripts/movienet_eval.py
+++ b/scripts/movienet_eval.py
@@ -19,28 +19,32 @@ import json
 import sys
 from pathlib import Path
 
+import numpy as np
+
 sys.path.insert(0, str(Path(__file__).resolve().parent))
 from sae_embed_loader import load_embedder
 
 
 def load_gallery(path: str) -> dict[str, dict]:
-    """Return {imdb_id: {"name": str, "embeddings": [[float]]}}."""
+    """Return {imdb_id: {"name": str, "refs": np.ndarray[n_refs, dim]}}."""
     with open(path) as f:
         data = json.load(f)
-    return {a["imdb_id"]: {"name": a["name"], "embeddings": a["embeddings"]}
+    return {a["imdb_id"]: {"name": a["name"],
+                           "refs": np.asarray(a["embeddings"], dtype=np.float32)}
             for a in data["actors"]}
 
 
-def dot(a: list[float], b: list[float]) -> float:
-    return sum(x * y for x, y in zip(a, b))
-
-
 def match(embedding: list[float], gallery: dict[str, dict]) -> tuple[str, float, dict[str, float]]:
-    """Return (best_imdb_id, best_similarity, {imdb_id: similarity})."""
-    scores: dict[str, float] = {}
-    for imdb_id, actor in gallery.items():
-        # max similarity across all reference embeddings for this actor
-        scores[imdb_id] = max(dot(embedding, ref) for ref in actor["embeddings"])
+    """Return (best_imdb_id, best_similarity, {imdb_id: similarity}).
+
+    Similarity to an actor is the max dot product over that actor's reference
+    embeddings; vectorised with numpy so a whole-library gallery stays fast.
+    """
+    vec = np.asarray(embedding, dtype=np.float32)
+    scores: dict[str, float] = {
+        imdb_id: float((actor["refs"] @ vec).max())
+        for imdb_id, actor in gallery.items()
+    }
     best_id = max(scores, key=lambda k: scores[k])
     return best_id, scores[best_id], scores
 
diff --git a/scripts/movienet_prep.py b/scripts/movienet_prep.py
index 331cec1..1aff111 100644
--- a/scripts/movienet_prep.py
+++ b/scripts/movienet_prep.py
@@ -133,20 +133,24 @@ def main():
     image_zip = movienet_root / "Image.zip"
     frame_cache: dict[str, np.ndarray] = {}
 
+    n_missing_in_zip = 0
     print(f"[prep] extracting {len(needed_paths)} frames from Image.zip…", file=sys.stderr)
     with zipfile.ZipFile(image_zip) as zf:
         for img_path in needed_paths:
             zip_entry = f"Image/{img_path}"
             try:
                 data = zf.read(zip_entry)
-                arr  = np.frombuffer(data, dtype=np.uint8)
-                img  = cv2.imdecode(arr, cv2.IMREAD_COLOR)
-                if img is not None:
-                    frame_cache[img_path] = img
             except KeyError:
-                pass  # file missing from zip, skip silently
+                n_missing_in_zip += 1  # frame referenced by an annotation but absent from Image.zip
+                continue
+            arr = np.frombuffer(data, dtype=np.uint8)
+            img = cv2.imdecode(arr, cv2.IMREAD_COLOR)
+            if img is not None:
+                frame_cache[img_path] = img
 
     print(f"[prep] frames loaded: {len(frame_cache)}/{len(needed_paths)}", file=sys.stderr)
+    if n_missing_in_zip:
+        print(f"[prep] frames absent from Image.zip: {n_missing_in_zip}", file=sys.stderr)
 
     per_actor_count: dict[str, int] = {}
     gt_entries = []