Upload scripts/train_job.py with huggingface_hub

scripts/train_job.py

@@ -0,0 +1,64 @@
+import os, sys, json, glob
+
+EPOCHS = int(os.getenv('EPOCHS', '150'))
+BATCH = int(os.getenv('BATCH', '64'))
+IMGSZ = int(os.getenv('IMGSZ', '320'))
+PUSH = os.getenv('PUSH', '0') == '1'
+DATA = os.getenv('DATA_DIR', '/data/ds')
+
+import torch
+print('torch', torch.__version__, 'cuda', torch.cuda.is_available(), flush=True)
+assert torch.cuda.is_available(), 'no CUDA device'
+
+n_tr = len(glob.glob(DATA + '/images/train/*'))
+n_vl = len(glob.glob(DATA + '/images/val/*'))
+print('train', n_tr, 'val', n_vl, flush=True)
+assert n_tr > 1000 and n_vl > 50, 'dataset incomplete'
+
+try:
+    import trackio
+    trackio.init(project='microduck-detector')
+    print('trackio ok', flush=True)
+except Exception as e:
+    trackio = None
+    print('trackio failed (ignored):', e, flush=True)
+
+from ultralytics import YOLO
+model = YOLO('yolo11n.pt')
+n_params = sum(p.numel() for p in model.model.parameters())
+print('param count:', n_params, flush=True)
+assert n_params < 5_000_000, 'model exceeds 5M params'
+
+results = model.train(
+    data=DATA + '/data.yaml',
+    imgsz=IMGSZ, epochs=EPOCHS, batch=BATCH, device=0,
+    project='/data/runs', name='microduck', exist_ok=True,
+    workers=4, patience=50,
+)
+print('TRAIN_DONE', flush=True)
+
+best = '/data/runs/microduck/weights/best.pt'
+assert os.path.exists(best), 'best.pt missing'
+
+metrics = model.val(data=DATA + '/data.yaml', imgsz=IMGSZ, device=0)
+print('VAL_MAP50', metrics.box.map50, 'MAP', metrics.box.map, flush=True)
+
+if trackio is not None:
+    try:
+        trackio.log({'val/mAP50': float(metrics.box.map50), 'val/mAP': float(metrics.box.map), 'params': sum(p.numel() for p in model.model.parameters())})
+        trackio.finish()
+    except Exception as e:
+        print('trackio log failed (ignored):', e, flush=True)
+
+if PUSH:
+    from huggingface_hub import HfApi
+    api = HfApi()
+    api.create_repo('pngwn/microduck-detector', repo_type='model', exist_ok=True)
+    api.upload_file(path_or_fileobj=best, path_in_repo='microduck_yolo11n.pt', repo_id='pngwn/microduck-detector', repo_type='model')
+    print('PUSHED model', flush=True)
+
+# qualitative predictions on real val photos (also useful after smoke)
+real_imgs = sorted(glob.glob(DATA + '/images/val/real_*.jpg'))
+print('real val images:', len(real_imgs), flush=True)
+model.predict(real_imgs, imgsz=IMGSZ, conf=0.35, save=True, project='/data/pred', name='real', exist_ok=True)
+print('PRED_DONE', flush=True)