utils

quantize_multiplierfunctionConvert a floating point multiplier into a fixed-point multiplier and shift.
compute_fused_activation_rangefunctionCompute the activation min and max values based on the fused activation function.
compute_padding_with_offsetfunctionComputes a single padding value (top/left) for one dimension with offset, as LiteRT does.
compute_paddingfunctionComputes a single padding value (top/left) for one dimension.
reduce_multiplier_q31_to_q15functionReduce a multiplier value for int16 quantization.

quantize_multiplier(real_multiplier: float) -> tuple[int, int]

Convert a floating point multiplier into a fixed-point multiplier and shift.

Provisional: importable, but outside the compatibility promise; a minor release can change or remove it.

Given a real multiplier, this function computes a pair (quantized_multiplier, shift) such that the fixed-point multiplication approximates the real multiplier:

real_multiplier ~ quantized_multiplier / (2^(31 - shift))

This is typically used in quantized inference to convert floating-point scaling factors to fixed-point representation.

The CMSIS-NN arm_nn_divide_by_power_of_two function only supports exponent values in [0, 31], so the resulting shift is clamped to [-31, 30]. When the real multiplier is so small that the shift would fall below -31 the contribution is negligible and the multiplier is flushed to zero.

Parameters

Parameters of quantize_multiplier
Name	Type	Default	Description
`real_multiplier`	`float`	Required	The floating-point multiplier.

Returns

Returns of quantize_multiplier
Value	Type	Description
`tuple`	`tuple[int, int]`	(quantized_multiplier, shift)

compute_fused_activation_range(
    activation_function: AirActivationType,
    output_zero_point: int = 0,
    output_scale: float = 1.0,
    qmin: int = -128,
    qmax: int = 127,
) -> tuple[int, int]

Compute the activation min and max values based on the fused activation function.

Provisional: importable, but outside the compatibility promise; a minor release can change or remove it.

Parameters

Parameters of compute_fused_activation_range
Name	Type	Default	Description
`activation_function`	`litert.ActivationFunctionType`	Required	The activation function type.
`output_zero_point`	`int`	0	The zero point for the output tensor. Defaults to 0.
`output_scale`	`float`	1.0	The scale for the output tensor. Defaults to 1.0.
`qmin`	`int`	-128	Minimum quantized value. Defaults to -128.
`qmax`	`int`	127	Maximum quantized value. Defaults to 127.

Returns

Returns of compute_fused_activation_range
Type	Description
`tuple[int, int]`	tuple[int, int]: The activation min and max values.

compute_padding_with_offset(
    in_size: int,
    filter_size: int,
    stride: int,
    dilation_rate: int = 1,
    padding_type: int = AirPaddingType.SAME,
    operator: str = 'A window operator',
) -> tuple[int, int]

Computes a single padding value (top/left) for one dimension with offset, as LiteRT does.

Provisional: importable, but outside the compatibility promise; a minor release can change or remove it.

This is LiteRT’s ComputePaddingWithOffset over :func:compute_out_size. VALID padding is therefore 0 except when the filter is wider than in_size plus a stride, as a transposed convolution’s output can be.

Parameters

Parameters of compute_padding_with_offset
Name	Type	Default	Description
`in_size`	`int`	Required	Input dimension (height or width).
`filter_size`	`int`	Required	Kernel dimension (height or width).
`stride`	`int`	Required	Stride for that dimension.
`dilation_rate`	`int`	1	Dilation rate for that dimension (default is 1).
`padding_type`	`int`	AirPaddingType.SAME	0 for SAME padding, 1 for VALID padding (default is SAME).
`operator`	`str`	'A window operator'	The operator, as an error names it.

Returns

Returns of compute_padding_with_offset
Type	Description
`tuple[int, int]`	tuple[int, int]: The top or left padding, and the extra bottom or right padding (0 or 1).

Raises

Errors raised by compute_padding_with_offset
Type	Description
`UnsupportedModelError`	If the stride or dilation is below 1, or the padding type is neither SAME nor VALID.

compute_padding(
    in_size: int,
    filter_size: int,
    stride: int,
    dilation_rate: int = 1,
    padding_type: int = AirPaddingType.SAME,
    operator: str = 'A window operator',
) -> int

Computes a single padding value (top/left) for one dimension.

Provisional: importable, but outside the compatibility promise; a minor release can change or remove it.

Parameters

Parameters of compute_padding
Name	Type	Default	Description
`in_size`	`int`	Required	Input dimension (height or width).
`filter_size`	`int`	Required	Kernel dimension (height or width).
`stride`	`int`	Required	Stride for that dimension.
`dilation_rate`	`int`	1	Dilation rate for that dimension (default is 1).
`padding_type`	`int`	AirPaddingType.SAME	0 for SAME padding, 1 for VALID padding (default is SAME).
`operator`	`str`	'A window operator'	The operator, as an error names it.

Returns

Returns of compute_padding
Value	Type	Description
`int`	`int`	The computed top or left padding value.

Raises

Errors raised by compute_padding
Type	Description
`UnsupportedModelError`	If the stride or dilation is below 1, or the padding type is neither SAME nor VALID.

reduce_multiplier_q31_to_q15(multiplier: int) -> int

Reduce a multiplier value for int16 quantization.

Provisional: importable, but outside the compatibility promise; a minor release can change or remove it.

Certain CMSIS functions for int16 quantized models require the multiplier to be reduced from Q31 format to Q15 format. This function performs that reduction.

Parameters

Parameters of reduce_multiplier_q31_to_q15
Name	Type	Default	Description
`multiplier`	`int`	Required	The original multiplier value.

Returns

Returns of reduce_multiplier_q31_to_q15
Value	Type	Description
`int`	`int`	The reduced multiplier value.