Class QuantizationPolicy

ppq/core/quant.py:213–306

QuantizationPolicy is a core abstraction for PPQ quantization calculation. QuantizationProperty and QuantizationPolicy together build a bitmap to describe quantization policy. A QuantizationPolicy instance contains multiple QuantizationProperty, QuantizationPolicy is used in PPQ

class QuantizationPolicy:
    """QuantizationPolicy is a core abstraction for PPQ quantization
    calculation. QuantizationProperty and QuantizationPolicy together build a
    bitmap to describe quantization policy.

    A QuantizationPolicy instance contains multiple QuantizationProperty,
        QuantizationPolicy is used in PPQ (alone with other configuration) to describe how a tensor is quantized.

    During simulating, executor will quantize tensor corresponding to its QuantizationPolicy.
        (QuantizationPolicy is included by TensorQuantizationConfig)

    There are 8 different quantization property(s) supported by PPQ now.

        PER_TENSOR: Also known as per-layer quantization, which mean all parameters of this layer share the same scale and offset.
            (For Convulution layer and Gemm layer which has bias, bias layer will be negative quantized, they do not have a valid scale)

        PER_CHANNEL: parameters are quantized alone channel axis, each channel has a stand-alone scale and offset.

        LINEAR: Linear quantization, follow formula: quant(x) = clip(round(x / scale))

        EXPONENTIAL: Exponential quantization, not yet used.

        SYMMETRICAL: Symmetrical quantization, offset is deactivated in this mode.

        ASYMMETRICAL: Asymmetrical quantization, offset is activated in this mode.

        POWER_OF_2: Power-of-2 quantization, scale must be pow(2, k) in this mode.

        DYNAMIC: Dynamic Activation Quantization, scale is computed on the fly.

    ATTENTION: Not all combinations of all 8 QuantizationProperty are valid, see QuantizationPolicy.__check_valid
    ATTENTION: QuantizationPolicy is read-only, user can only assign its value when created, the only interface of
        QuantizationPolicy is function QuantizationPolicy.has_property.
    """
    def __init__(self, policy: int) -> None:
        if not QuantizationPolicy.__check_valid(policy):
            raise ValueError(
                'invalid quantization pattern, valid partterns are listed in '
                'ppq.core.OperationQuantizationPolicy.__check_valid'
            )
        self._policy = policy

    def has_property(self, property: QuantizationProperty) -> bool:
        return (self._policy & property.value) != 0

    def __eq__(self, o: object) -> bool:
        if not isinstance(o, QuantizationPolicy):
            raise TypeError('Can only compare QuantizationPolicy object '
                            'with another QuantizationPolicy object.')
        return self._policy == o._policy

    @ classmethod
    def __check_valid(cls, policy):
        return policy in {
            # Standard Int Quantization
            QuantizationProperty.ASYMMETRICAL | QuantizationProperty.LINEAR | QuantizationProperty.PER_CHANNEL,
            QuantizationProperty.ASYMMETRICAL | QuantizationProperty.LINEAR | QuantizationProperty.PER_TENSOR,
            QuantizationProperty.SYMMETRICAL | QuantizationProperty.LINEAR | QuantizationProperty.PER_CHANNEL,