phase 10

detect/checkpoint/frames.py

@@ -0,0 +1,80 @@
+"""Frame image storage — save/load to S3/MinIO as JPEGs."""
+
+from __future__ import annotations
+
+import logging
+import os
+import tempfile
+
+import numpy as np
+from PIL import Image
+
+from detect.models import Frame
+
+logger = logging.getLogger(__name__)
+
+BUCKET = os.environ.get("S3_BUCKET_OUT", "mpr-media-out")
+CHECKPOINT_PREFIX = "checkpoints"
+
+
+def save_frames(job_id: str, frames: list[Frame]) -> dict[int, str]:
+    """
+    Save frame images to S3 as JPEGs.
+
+    Returns manifest: {sequence: s3_key}
+    """
+    from core.storage.s3 import upload_file
+
+    manifest = {}
+
+    for frame in frames:
+        key = f"{CHECKPOINT_PREFIX}/{job_id}/frames/{frame.sequence}.jpg"
+
+        with tempfile.NamedTemporaryFile(suffix=".jpg", delete=False) as tmp:
+            img = Image.fromarray(frame.image)
+            img.save(tmp, format="JPEG", quality=85)
+            tmp_path = tmp.name
+
+        try:
+            upload_file(tmp_path, BUCKET, key)
+        finally:
+            os.unlink(tmp_path)
+
+        manifest[frame.sequence] = key
+
+    logger.info("Saved %d frames to s3://%s/%s/%s/frames/",
+                len(frames), BUCKET, CHECKPOINT_PREFIX, job_id)
+    return manifest
+
+
+def load_frames(manifest: dict[int, str], frame_metadata: list[dict]) -> list[Frame]:
+    """
+    Load frame images from S3 and reconstitute Frame objects.
+
+    frame_metadata: list of dicts with sequence, chunk_id, timestamp, perceptual_hash.
+    """
+    from core.storage.s3 import download_to_temp
+
+    meta_map = {m["sequence"]: m for m in frame_metadata}
+    frames = []
+
+    for seq, key in manifest.items():
+        tmp_path = download_to_temp(BUCKET, key)
+        try:
+            img = Image.open(tmp_path).convert("RGB")
+            image_array = np.array(img)
+        finally:
+            os.unlink(tmp_path)
+
+        meta = meta_map.get(seq, {})
+        frame = Frame(
+            sequence=seq,
+            chunk_id=meta.get("chunk_id", 0),
+            timestamp=meta.get("timestamp", 0.0),
+            image=image_array,
+            perceptual_hash=meta.get("perceptual_hash", ""),
+        )
+        frames.append(frame)
+
+    frames.sort(key=lambda f: f.sequence)
+    return frames