torch.js has not been released yet.

torch.nn.LPPool1d

class LPPool1d extends Module

new LPPool1d(norm_type: number, kernel_size: number, options?: LPPool1dOptions)

readonlynorm_type(number)
readonlykernel_size(number)
readonlystride(number)
readonlyceil_mode(boolean)

1D Lp-norm pooling: reduces sequence length using Lp-norm instead of max or mean.

Applies pooling using Lp-norm (power average): out = (sum(|x|^p))^(1/p). Generalized pooling operation where different norm_type values produce different behaviors:

norm_type=1 (L1): Sum of absolute values (Manhattan distance)
norm_type=2 (L2): Euclidean norm (square root of sum of squares)
Between max (sparse) and mean (dense) pooling

Useful for robustness: less extreme than max pooling but less smooth than average. Essential for:

Robust feature extraction (between max and mean)
Learning robust representations to outliers
Domain-specific downsampling requiring norm-based pooling

L1 vs L2: L1 = sum(|x|), L2 = sqrt(sum(x²))
Robustness: Between max (sensitive to outliers) and mean (loses peaks)
ceil_mode: Affects output size calculation at boundaries

Examples

// L2 (Euclidean) norm pooling
const pool = new torch.nn.LPPool1d(2, 3);  // L2-norm, kernel=3
const x = torch.randn([32, 64, 100]);
const y = pool.forward(x);  // [32, 64, ~34] - L2-norm pooling

// L1 (Manhattan) norm pooling
const pool = new torch.nn.LPPool1d(1, 2);  // L1-norm, kernel=2
const x = torch.randn([16, 128, 50]);
const y = pool.forward(x);  // [16, 128, 25] - sum of absolute values

See Also

PyTorch torch.nn.LPPool1d

LogSoftmaxOptions

LPPool1dOptions