Function main

deploy/ios_demo/scripts/quantize_onnx_model.py:213–381

()

def main() -> None:
    p = argparse.ArgumentParser(description="Quantize PaddleOCR ONNX model.")
    p.add_argument(
        "--input-model-dir",
        required=True,
        type=Path,
        help="Input model directory",
    )
    p.add_argument(
        "--output-model-dir",
        required=True,
        type=Path,
        help="Output model directory",
    )
    p.add_argument(
        "--mode",
        required=True,
        choices=("dynamic", "static"),
        help="dynamic: weight-only (quantize_dynamic). static: QDQ (quantize_static, needs --calib-data-dir).",
    )
    p.add_argument(
        "--calib-data-dir",
        type=Path,
        default=None,
        help="Directory of float32 .npy calibration samples (static mode only; one tensor per file).",
    )
    p.add_argument(
        "--per-channel",
        action=argparse.BooleanOptionalAction,
        default=True,
        help="Per-channel weight quantization (default: true).",
    )
    p.add_argument(
        "--calibration-method",
        default="MinMax",
        help="ORT CalibrationMethod name (e.g. MinMax, Entropy, Percentile).",
    )
    p.add_argument(
        "--no-verify",
        action="store_true",
        help="Skip ONNX checker validation of the output model after quantization.",
    )
    p.add_argument(
        "--ort-preprocess",
        action=argparse.BooleanOptionalAction,
        default=True,
        help=(
            "Before quantize_static / quantize_dynamic, run ORT quant_pre_process (shape infer + "
            "optional graph optimization) and attach onnx.quant metadata."
        ),
    )
    p.add_argument(
        "--onnx-opset-convert",
        action=argparse.BooleanOptionalAction,
        default=True,
        help=(
            "After ORT writes the output, run onnx.version_converter when the graph's declared "
            "default-domain opset is below --onnx-target-opset. Use --no-onnx-opset-convert to keep "