quantization

AirFixedPointScaleclassFixed-point form of a real-valued rescale factor.
AirQuantizationDetailsclassBase class for quantization details.
AirBlockwiseQuantizationclassBlockwise quantization details.
AirQuantizationParametersclassParameters for quantization of a tensor.

AirFixedPointScale()

Fixed-point form of a real-valued rescale factor.

The pair reproduces a real multiplier as x * real ~ (x * multiplier) >> (31 - shift), where multiplier is a Q31 value (see quantize_multiplier). Following the CMSIS-NN convention, a positive shift is applied as a left shift before the Q31 multiply and a negative one as a right shift after it.

multiplier: int = Field(..., description='The fixed point multiplier')

Q31 mantissa of the rescale factor.

shift: int = Field(..., description='The fixed point shift')

Power-of-two exponent, in [-31, 30].

from_real_multiplier(real_multiplier: float)

classmethod

Create a fixed point scale from a real multiplier.

AirQuantizationDetails()

Base class for quantization details.

AirBlockwiseQuantization()

Blockwise quantization details.

scales: int = Field(0, description='Number of scales')

zero_points: int = Field(0, description='Number of zero points')

block_size: int = Field(0, description='Size of the blocks')

AirQuantizationParameters()

Parameters for quantization of a tensor.

min: list[float] | None = Field(None, description='Minimum values for quantization')

max: list[float] | None = Field(None, description='Maximum values for quantization')

scale: list[float] | None = Field(None, description='Scale factors for quantization')

zero_point: list[int] | None = Field(None, description='Zero points for quantization')

details_type: int = Field(0, description='Type of quantization details')

details: AirQuantizationDetails | None = Field(None, description='Quantization details')

quantization_dimension: int = Field(0, description='Quantization dimension')

scales: list[float]

Per-tensor or per-channel scales; raises if the tensor is not quantized.

zero_points: list[int]

Per-tensor or per-channel zero points; raises if the tensor is not quantized.