ConvNeXt Block (1D)

Problem

Forward pass of a simplified 1D ConvNeXt block. For each channel: apply a depthwise convolution with the shared kernel and same padding, normalize the result over the length dimension (LayerNorm), scale by the per-channel gamma, and add the residual: out = x + gamma * LN(DW(x)).

The signature is convnext_block_1d(x, dw_kernel, gamma, eps=1e-5). x has shape (channels, length).

Hint

ConvNeXt mixes a depthwise spatial conv with an inverted-bottleneck channel MLP; this simplified block keeps the depthwise conv, LayerNorm and layer scale.

Test cases

case 1

in: [[[1,2,3]],[1,1,1],[1]]

expected: [[-0.3363019143,3.0690415315,3.2672603829]]

case 2

in: [[[1,2,3],[4,5,6]],[1,1,1],[1,1]]

expected: [[-0.3363019143,3.0690415315,3.2672603829],[2.9309558914,6.3363051357,5.7327389729]]

case 3

in: [[[1,2,3]],[0,1,0],[0.5]]

expected: [[0.387632157,2,3.612367843]]

case 4

in: [[[1,2,3]],[1,1,1],[0]]

expected: [[1,2,3]]

Starter codePython

import math
def convnext_block_1d(x, dw_kernel, gamma, eps=1e-5):
    # Your code here
    pass

4 test cases · run your code against them in the editor