Function quantize_onnx_model

ppq/api/interface.py:185–276

量化一个 onnx 原生的模型 输入一个 onnx 模型的文件路径 返回一个量化后的 PPQ.IR.BaseGraph quantize onnx model, input onnx model and return quantized ppq IR graph. Args: onnx_import_file (str): 被量化的 onnx 模型文件路径 onnx model location calib_dataloader (DataLoader): 校准数据集 calibration data loader cali

(
    onnx_import_file: str,
    calib_dataloader: DataLoader,
    calib_steps: int,
    input_shape: List[int],
    platform: TargetPlatform,
    input_dtype: torch.dtype = torch.float,
    inputs: List[Any] = None,
    setting: QuantizationSetting = None,
    collate_fn: Callable = None,
    device: str = 'cuda',
    verbose: int = 0,
    do_quantize: bool = True,
)

@ empty_ppq_cache
def quantize_onnx_model(
    onnx_import_file: str,
    calib_dataloader: DataLoader,
    calib_steps: int,
    input_shape: List[int],
    platform: TargetPlatform,
    input_dtype: torch.dtype = torch.float,
    inputs: List[Any] = None,
    setting: QuantizationSetting = None,
    collate_fn: Callable = None,
    device: str = 'cuda',
    verbose: int = 0,
    do_quantize: bool = True,
) -> BaseGraph:
    """量化一个 onnx 原生的模型 输入一个 onnx 模型的文件路径 返回一个量化后的 PPQ.IR.BaseGraph quantize
    onnx model, input onnx model and return quantized ppq IR graph.

    Args:
        onnx_import_file (str): 被量化的 onnx 模型文件路径 onnx model location

        calib_dataloader (DataLoader): 校准数据集 calibration data loader

        calib_steps (int): 校准步数 calibration steps

        collate_fn (Callable): 校准数据的预处理函数 batch collate func for preprocessing

        input_shape (List[int]): 模型输入尺寸，用于执行 jit.trace，对于动态尺寸的模型，输入一个模型可接受的尺寸即可。
            如果模型存在多个输入，则需要使用 inputs 变量进行传参，此项设置为 None
                                a list of ints indicating size of input, for multiple inputs, please use
                                keyword arg inputs for direct parameter passing and this should be set to None

        input_dtype (torch.dtype): 模型输入数据类型，如果模型存在多个输入，则需要使用 inputs 变量进行传参，此项设置为 None
                                the torch datatype of input, for multiple inputs, please use keyword arg inputs
                                for direct parameter passing and this should be set to None

        inputs (List[Any], optional): 对于存在多个输入的模型，在Inputs中直接指定一个输入List，从而完成模型的tracing。
                                for multiple inputs, please give the specified inputs directly in the form of
                                a list of arrays

        setting (OptimSetting): 量化配置信息，用于配置量化的各项参数，设置为 None 时加载默认参数。
                                Quantization setting, default setting will be used when set None

        do_quantize (Bool, optional): 是否执行量化 whether to quantize the model, defaults to True.


        platform (TargetPlatform, optional): 量化的目标平台 target backend platform, defaults to TargetPlatform.DSP_INT8.

        device (str, optional): 量化过程的执行设备 execution device, defaults to 'cuda'.

        verbose (int, optional): 是否打印详细信息 whether to print details, defaults to 0.

    Raises:
        ValueError: 给定平台不可量化 the given platform doesn't support quantization
        KeyError: 给定平台不被支持 the given platform is not supported yet

    Returns:
        BaseGraph: 量化后的IR，包含了后端量化所需的全部信息
                   The quantized IR, containing all information needed for backend execution