Method __init__

ppq/quantization/algorithm/training.py:319–364

(
        self, config: TensorQuantizationConfig, var: Variable, 
        is_parameter_trainable: bool = True, 
        is_scale_trainable:     bool = True, 
        is_offset_trainable:    bool = True
    )

class LSQDelegator(TorchQuantizeDelegator):
    def __init__(
        self, config: TensorQuantizationConfig, var: Variable, 
        is_parameter_trainable: bool = True, 
        is_scale_trainable:     bool = True, 
        is_offset_trainable:    bool = True
    ) -> None:
        self.config        = config
        self.is_parameter  = var.is_parameter
        self.var           = var
        self.policy        = config.policy
        self.passive       = config.state == QuantizationStates.PASSIVE

        self.param_backup = None
        if self.is_parameter and is_parameter_trainable:
            self.param_backup = self.var.value.clone()

        # There is 4 checks for scale training:
        #   1. scale is valid
        #   2. state is active
        #   3. do not have POWER_OF_2 policy but Must have Linear policy
        #   4. is_scale_trainable = True
        self.scale_backup       = None
        self.is_scale_trainable = False
        if is_scale_trainable:
            policy_check = not config.policy.has_property(QuantizationProperty.POWER_OF_2)
            linear_check = config.policy.has_property(QuantizationProperty.LINEAR)
            state_check  = ((config.state == QuantizationStates.ACTIVATED) and (config.dominated_by == config))
            value_check  = isinstance(config.scale, torch.Tensor)
            if policy_check and state_check and value_check and linear_check:
                self.is_scale_trainable = True
                self.scale_backup = self.config.scale.detach().clone()

        # There is 4 checks for offset training:
        #   1. offset is valid
        #   2. state is active
        #   3. do not have SYMMETRICAL policy
        #   4. is_scale_trainable = True
        self.offset_backup       = None
        self.is_offset_trainable = False
        if is_offset_trainable:
            policy_check = not config.policy.has_property(QuantizationProperty.SYMMETRICAL)
            state_check  = ((config.state == QuantizationStates.ACTIVATED) and (config.dominated_by == config))
            value_check  = isinstance(config.offset, torch.Tensor)
            if policy_check and state_check and value_check:
                self.is_offset_trainable = True
                self.offset_backup = self.config.offset.detach().clone()

    def trainable_tensors(self) -> List[torch.Tensor]:
        params = []