Method register_quantize_delegate

ppq/executor/torch.py:296–323

Since PPQ 0.6.2, Interface TorchQuantizeDelegate is introduced to customize quantization logic: To be specific, you are suppose to inherit this class, and define your own computation logic within function __call__. Pass your Delegate to TorchExecutor by TorchExec

(
        self, config: TensorQuantizationConfig,
        delegator: TorchQuantizeDelegator)

        self.deploy()

    def register_quantize_delegate(
        self, config: TensorQuantizationConfig,
        delegator: TorchQuantizeDelegator):
        """Since PPQ 0.6.2, Interface TorchQuantizeDelegate is introduced to
        customize quantization logic: To be specific, you are suppose to
        inherit this class, and define your own computation logic within
        function __call__.

            Pass your Delegate to TorchExecutor by TorchExecutor.register_quantize_delegate(c, d)
                Where c is the target quantization config, d is your delegator class.
                Once you invoke this function, PPQ execution system will hand the quantization
                computation of config c over to your delegate. PPQ execution system will no
                longer quantize variable related with config c anymore.

        Notice that a delegate replaces quantization computation only, it still under the control of PPQ quantization
        System, so to say if your config has an invalid state like DEQUANTIZED, PPQ execution system will never been
        required to quantize related tensor and so your delegate class will take no effects on config c.

        Remove delegate function by TorchExecutor.remove_quantize_delegate(c)
        """
        if not isinstance(delegator, TorchQuantizeDelegator):
            raise TypeError(
                f'You can only register a TorchQuantizeDelegate as quantization delegator function,'
                f' however a/an {type(delegator)} was given')
        if not isinstance(config, TensorQuantizationConfig):
            raise TypeError(
                f'Except a TensorQuantizationConfig instance, however {type(config)} was passed.')
        self._delegates[config] = delegator

    def remove_quantize_delegate(
        self, config: TensorQuantizationConfig):