feat: implement resume functionality in batch feature extraction to skip processed samples

app/training/extract_features_batch.py

@@ -232,11 +232,28 @@ def extract_batch(
     failed = 0
     t_start = _time.time()
 
-    tasks = [(s["file_path"], int(s["label_int"])) for s in samples]
-
-    with open(output_path, "w", newline="", encoding="utf-8") as f:
+    done_paths: set[str] = set()
+    resume = output_path.exists() and output_path.stat().st_size > 0
+    if resume:
+        with open(output_path, "r", encoding="utf-8") as f_prev:
+            reader = csv.DictReader(f_prev)
+            for r in reader:
+                done_paths.add(r["file_path"])
+        print(f"  Resuming: {len(done_paths)} samples already processed, skipping", flush=True)
+
+    tasks = [
+        (s["file_path"], int(s["label_int"]))
+        for s in samples
+        if s["file_path"] not in done_paths
+    ]
+    total_remaining = len(tasks)
+    print(f"  Remaining: {total_remaining} samples to process", flush=True)
+
+    file_mode = "a" if resume else "w"
+    with open(output_path, file_mode, newline="", encoding="utf-8") as f:
         writer = csv.DictWriter(f, fieldnames=out_columns)
-        writer.writeheader()
+        if not resume:
+            writer.writeheader()
         f.flush()
 
         with mp.Pool(processes=n_workers) as pool:
@@ -253,9 +270,9 @@ def extract_batch(
                     f.flush()
                     elapsed = _time.time() - t_start
                     rate = i / elapsed if elapsed > 0 else 0
-                    eta = (len(samples) - i) / rate if rate > 0 else 0
+                    eta = (total_remaining - i) / rate if rate > 0 else 0
                     print(
-                        f"  [{i}/{len(samples)}] "
+                        f"  [{i}/{total_remaining}] "
                         f"ok={success} fail={failed} "
                         f"rate={rate:.1f}/s eta={eta / 60:.1f}m",
                         flush=True,