Make COCO export streaming to avoid OOM on large configs

tools/export_coco.py

@@ -4,7 +4,7 @@ import orjson
 from datasets import load_dataset
 
 
-def export_split(ds, out_dir: Path, split_name: str):
+def export_streaming(split_iter, out_dir: Path, split_name: str):
     images_dir = out_dir / split_name / "images"
     ann_dir = out_dir / split_name / "annotations"
     images_dir.mkdir(parents=True, exist_ok=True)
@@ -14,14 +14,14 @@ def export_split(ds, out_dir: Path, split_name: str):
     annotations = []
     categories = None
     ann_id = 1
+    n_images = 0
 
-    for ex in ds:
+    for ex in split_iter:
         image_id = int(ex["image_id"])
         filename = ex["filename"]
         width = int(ex["width"])
         height = int(ex["height"])
 
-        # write image bytes
         (images_dir / filename).write_bytes(ex["image_bytes"])
 
         images.append({
@@ -41,6 +41,10 @@ def export_split(ds, out_dir: Path, split_name: str):
             ann_id += 1
             annotations.append(a)
 
+        n_images += 1
+        if n_images % 500 == 0:
+            print(f"[{split_name}] exported {n_images} images...")
+
     coco = {
         "images": images,
         "annotations": annotations,
@@ -60,18 +64,18 @@ def main():
     out_dir = Path(args.out) / args.config
     out_dir.mkdir(parents=True, exist_ok=True)
 
-    # streaming=False ensures export writes locally
-    train = load_dataset(args.repo, args.config, split="train", streaming=False)
-    idt   = load_dataset(args.repo, args.config, split="id_test", streaming=False)
-    ood   = load_dataset(args.repo, args.config, split="ood_test", streaming=False)
+    # Streaming avoids Arrow memory-mapping OOM on large configs
+    train = load_dataset(args.repo, args.config, split="train", streaming=True)
+    idt   = load_dataset(args.repo, args.config, split="id_test", streaming=True)
+    ood   = load_dataset(args.repo, args.config, split="ood_test", streaming=True)
 
-    n_i, n_a = export_split(train, out_dir, "train")
+    n_i, n_a = export_streaming(train, out_dir, "train")
     print(f"[train] images={n_i} annotations={n_a}")
 
-    n_i, n_a = export_split(idt, out_dir, "id_test")
+    n_i, n_a = export_streaming(idt, out_dir, "id_test")
     print(f"[id_test] images={n_i} annotations={n_a}")
 
-    n_i, n_a = export_split(ood, out_dir, "ood_test")
+    n_i, n_a = export_streaming(ood, out_dir, "ood_test")
     print(f"[ood_test] images={n_i} annotations={n_a}")
 
     print(f"COCO export written to: {out_dir}")