instance-norm.md

instance_norm

Normalizes each (sample, channel) slice independently across its spatial dims — mean and variance are computed per-sample rather than across the batch.

ATen source: aten.instance_norm (preserved as composite by get_decomp_table())

Inputs

Name	Shape	Description
input	(N, C, *spatial)	1, 2, or 3 spatial dims (e.g., (N, C, H, W) for 2D)
gamma	(C,)	Per-channel scale applied after normalization
beta	(C,)	Per-channel shift applied after the scale

Attributes

Name	Type	Description
eps	float	Numerical-stability epsilon
version	int	Composite op version

Output

Name	Shape	Description
output	(N, C, *spatial)	Same shape as input

Data types

fp16, fp32.

PyTorch example

import torch

N, C, H, W = 2, 6, 10, 10
input = torch.randn(N, C, H, W)
gamma = torch.randn(C)
beta = torch.randn(C)

output = torch.ops.aten.instance_norm.default(
    input,
    gamma,
    beta,
    None,
    None,  # running_mean / running_var unused in inference
    True,  # use_input_stats
    0.1,  # momentum (ignored in inference)
    1e-5,  # eps
    True,  # cudnn_enabled (ignored)
)

Reference

torch.nn.InstanceNorm2d