class

InstanceNorm2d

extends_InstanceNormBase

InstanceNorm2d(num_features: int, eps: float = 1e-05, momentum: float = 0.1, affine: bool = False, track_running_stats: bool = False, device: DeviceLike = None, dtype: DTypeLike = None)

source

Instance normalization for 4-D input (N, C, H, W).

Normalises each (n, c) slice over its spatial height and width independently:

y_{n,c} = \frac{x_{n,c} - \mu_{n,c}} {\sqrt{\sigma_{n,c}^2 + \varepsilon}} \cdot \gamma_c + \beta_c

where $\mu_{n,c}$ and $\sigma_{n,c}^2$ are computed over the $H \times W$ spatial positions of sample $n$ , channel $c$ .

Unlike BatchNorm2d, the statistics do not mix across the batch, making the output of each image entirely independent of the other images in the batch. This property makes InstanceNorm2d well-suited to image style transfer, generative models, and any scenario where per-image normalization is desirable.

Parameters

num_featuresint

Number of feature channels

C

epsfloat= 1e-05

Numerical stability constant. Default: 1e-5.

momentumfloat= 0.1

EMA factor for optional running stats. Default: 0.1.

affinebool= False

If True, learns per-channel

\gamma

and

\beta

. Default: False.

track_running_statsbool= False

If True, maintains running mean/variance buffers. Default: False.

deviceDeviceLike= None

Device for parameters/buffers. Default: None.

dtypeDTypeLike= None

Data type for parameters/buffers. Default: None.

Attributes

weightParameter or None

Per-channel scale

\gamma

, shape (num_features,). None when affine=False.

biasParameter or None

Per-channel shift

\beta

, shape (num_features,). None when affine=False.

running_meanTensor or None

Running mean per channel, shape (num_features,). None when track_running_stats=False.

running_varTensor or None

Running variance per channel, shape (num_features,). None when track_running_stats=False.

Notes

Input: $(N, C, H, W)$
Output: $(N, C, H, W)$ — same shape.

Instance Norm was introduced for style transfer, where each image should be normalized independently so that stylistic statistics from other images do not bleed through.
At inference time, weight and bias can be replaced by style-conditioned tensors to implement Adaptive Instance Normalization (AdaIN).

Examples

>>> import lucid
>>> import lucid.nn as nn
>>> norm = nn.InstanceNorm2d(64)
>>> x = lucid.randn(4, 64, 128, 128)
>>> out = norm(x)
>>> out.shape
(4, 64, 128, 128)
With affine parameters for conditional style transfer:
>>> norm_affine = nn.InstanceNorm2d(64, affine=True)
>>> out2 = norm_affine(x)
>>> out2.shape
(4, 64, 128, 128)