Exports a checkpoint to a TensorRT engine at a given precision, skipping export if a valid engine already exists.
def export_engine(model_path, model_name, precision):
engine_path = OUTPUT_DIR / f'{model_name}_{precision}.engine'
if engine_path.exists() and engine_path.stat().st_size > 1e6:
return engine_path
model = YOLO(str(model_path))
model.export(
format='engine',
imgsz=IMGSZ,
batch=BATCH,
half=(precision == 'fp16'),
int8=False,
workspace=4.0,
device=0
)
candidates = [
model_path.with_suffix('.engine'),
model_path.parent / f'{model_path.stem}.engine',
]
for cand in candidates:
if cand.exists():
cand.rename(engine_path)
return engine_path
raise FileNotFoundError(f'Engine not found for {model_name}')
TensorRT engines are hardware-specific — compile on the exact GPU you plan to deploy to, not on a development machine.