torch.js has not been released yet.

torch.nn.InstanceNorm2d

class InstanceNorm2d extends _InstanceNorm

Instance Normalization for 2D inputs (images): normalizes each image independently.

Per-sample normalization across spatial dimensions. Essential for style transfer (CycleGAN, MUNIT) and generative models. Each image's channels normalized independently, not across batch.

When to use InstanceNorm2d:

Style transfer and image-to-image translation networks
GANs and generative models sensitive to instance statistics
Domain adaptation with unpaired images
Single-image inference (batch size 1)
When batch statistics are unreliable (small batches)

\begin{aligned} \mu_{n,c} = \frac{1}{H \cdot W} \sum_{h,w} x[n,c,h,w], \quad \sigma_{n,c}^2 = \frac{1}{H \cdot W} \sum_{h,w} (x[n,c,h,w] - \mu_{n,c})^2 \\ x_{\text{norm}} = \frac{x - \mu_{n,c}}{\sqrt{\sigma_{n,c}^2 + \epsilon}}, \quad y = \gamma x_{\text{norm}} + \beta \end{aligned}

2D spatial normalization: Normalizes across height and width per image per channel
Batch-independent: No batch statistics used
Perfect for single-image: Works with batch size 1 unlike BatchNorm
Standard in style transfer: Used in CycleGAN, MUNIT, Pix2Pix

Loses batch info: Ignores batch statistics completely
May hurt supervised learning: BatchNorm typically better for classification

Examples

// Style transfer network (CycleGAN-style)
const norm = new torch.nn.InstanceNorm2d(256);  // 256 channels
const x = torch.randn([16, 256, 64, 64]);  // [batch, channels, height, width]
const y = norm.forward(x);  // Each image normalized independently

// Image-to-image translation independent of batch size
const norm = new torch.nn.InstanceNorm2d(64);
const x_batch_1 = torch.randn([1, 64, 256, 256]);  // Single image
const y1 = norm.forward(x_batch_1);
const x_batch_32 = torch.randn([32, 64, 256, 256]);  // Batch of 32
const y32 = norm.forward(x_batch_32);  // Same normalization behavior

See Also

PyTorch torch.nn.InstanceNorm2d