albumentations.batch.normalization

Model-bound input normalization after accelerator transfer.

Members

classInputNormalizer

InputNormalizerclass

InputNormalizer(
    mode: InputNormalizationMode,
    output_precision: ComputePrecision,
    mean: float | Sequence[float] | Tensor | None,
    standard_deviation: float | Sequence[float] | Tensor | None,
    input_scale: float = 1.0
)

Normalize image or volume batches on their device. Normalization runs outside autograd. Args: mode (InputNormalizationMode): "none" only scales and casts. "fixed" uses supplied statistics. "per_sample" standardizes each channel of each sample independently. "per_sample_nonzero" uses only nonzero positions and keeps zeros as padding. output_precision (ComputePrecision): Output dtype, selected after FP32 normalization. mean (float | Sequence[float] | Tensor | None): Scalar or per-channel means for fixed normalization. standard_deviation (float | Sequence[float] | Tensor | None): Positive scalar or per-channel standard deviations for fixed normalization. input_scale (float): Positive divisor applied before normalization.

Parameters

Name	Type	Default	Description
mode	InputNormalizationMode	-	-
output_precision	ComputePrecision	-	-
mean	One of: float Sequence[float] Tensor None	-	-
standard_deviation	One of: float Sequence[float] Tensor None	-	-
input_scale	float	1.0	-