class

LazyInstanceNorm2d

extends_LazyInstanceNormMixin

LazyInstanceNorm2d(eps: float = 1e-05, momentum: float = 0.1, affine: bool = False, track_running_stats: bool = False, device: DeviceLike = None, dtype: DTypeLike = None)

source

InstanceNorm2d with lazy num_features inference.

Identical to InstanceNorm2d except that num_features is inferred from x.shape[1] on the first forward pass.

Parameters

epsfloat= 1e-05

Numerical stability constant. Default: 1e-5.

momentumfloat= 0.1

EMA factor for optional running statistics. Default: 0.1.

affinebool= False

If True, lazily allocates per-channel scale and shift. Default: False.

track_running_statsbool= False

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

deviceDeviceLike= None

Device for lazily allocated tensors. Default: None.

dtypeDTypeLike= None

Data type for lazily allocated tensors. Default: None.

Notes

Input: $(N, C, H, W)$ — same as InstanceNorm2d. C is inferred on the first forward pass.
Output: same shape as the input.

Examples

>>> import lucid
>>> import lucid.nn as nn
>>> norm = nn.LazyInstanceNorm2d()
>>> x = lucid.randn(2, 64, 128, 128)
>>> out = norm(x)   # num_features=64 inferred here
>>> out.shape
(2, 64, 128, 128)
Commonly used in generator networks where the channel dimension is
determined at runtime by a preceding convolution:
>>> norm_track = nn.LazyInstanceNorm2d(track_running_stats=True)
>>> out2 = norm_track(x)
>>> out2.shape
(2, 64, 128, 128)