Method create_default_quant_config

ppq/quantization/quantizer/base.py:107–204

为你的算子创建一个默认量化信息 对于一个 Onnx 算子而言，它总是会有几个输入和输出 Variable 你需要为每一个相连的 Variable 初始化量化信息 TensorQuantConfig 这个函数就是用来帮你初始化这些信息的。 一个麻烦的问题是： 对于很多 onnx 算子而言，他们的部分输入都是不需要量化的: 如 Clip 算子的三个输入 value, min, max, 大

(
        op: Operation, 
        num_of_bits: int = 8,
        quant_min: Union[int, float] = -127, 
        quant_max: Union[int, float] = 128, 
        observer_algorithm: str = 'percentile', 
        policy: QuantizationPolicy = 
            QuantizationPolicy(
                QuantizationProperty.PER_TENSOR + 
                QuantizationProperty.LINEAR + 
                QuantizationProperty.SYMMETRICAL), 
        rounding: RoundingPolicy = RoundingPolicy.ROUND_HALF_EVEN, 
        exponent_bits: int = 0,
    )

    @ staticmethod
    def create_default_quant_config(
        op: Operation, 
        num_of_bits: int = 8,
        quant_min: Union[int, float] = -127, 
        quant_max: Union[int, float] = 128, 
        observer_algorithm: str = 'percentile', 
        policy: QuantizationPolicy = 
            QuantizationPolicy(
                QuantizationProperty.PER_TENSOR + 
                QuantizationProperty.LINEAR + 
                QuantizationProperty.SYMMETRICAL), 
        rounding: RoundingPolicy = RoundingPolicy.ROUND_HALF_EVEN, 
        exponent_bits: int = 0,
    ) -> OperationQuantizationConfig:
        """
        为你的算子创建一个默认量化信息
        
        对于一个 Onnx 算子而言，它总是会有几个输入和输出 Variable
        你需要为每一个相连的 Variable 初始化量化信息 TensorQuantConfig
        
        这个函数就是用来帮你初始化这些信息的。
        
        一个麻烦的问题是：

            对于很多 onnx 算子而言，他们的部分输入都是不需要量化的:
            
            如 Clip 算子的三个输入 value, min, max, 大部分框架不要求量化 min, max
            如 Reshape 算子的两个输入 value, shape, 其中 shape 不能够被量化

            PPQ 的算子接线器中记录了这些信息

            算子接线器中记录了所有标准 onnx 的默认量化策略
            该函数将使用预定义的算子量化策略初始化量化信息
        
        你可以在 Quantizer 中对默认量化策略进行进一步修改

        Create a default quantization configuration for given op.
        For each onnx op, there will be some input and output variables.
        
        You are required to create tensor quantization config for every
        input and output variables.
        
        This function is designed for creating a default quantization config for you.

            The created OQC(Op Quantization Config) is based on OpSocket.

            In fact, there are some rules or templates when creating the OQC:
            For Clip Op which has 3 input variable, namely value, min and max
                most framework does not require a quantization config for min and max.
            For Reshape Op which has 2 input variable, namely value and shape
                the input shape can never be quantized.

        Those rules are pre-defined within OpSocket, thus ppq will create
        OQC based on underlying OpSocket of your Op.
        
        After the default OQC got created, you can overwrite its state in quantizer.
        """
        assert isinstance(op, Operation), (