Skip to main content

DepthwiseConv

Depthwise Convolution

Applies separate convolution to each input channel independently. Each channel is convolved with its own set of filters.

Parameters:

  • channels: Number of input/output channels (same for depthwise)
  • kernel_size: Size of the convolving kernel
  • stride: Stride of the convolution (default: 1)
  • padding: Zero-padding added to both sides (default: 0)
  • dilation: Spacing between kernel elements (default: 1)
  • bias: If true, adds learnable bias (default: true)

Shape Contract:

  • Input: [batch, channels, height, width]
  • Output: [batch, channels, out_height, out_width]

Notes:

  • Equivalent to Conv2d with groups=in_channels
  • Much fewer parameters than standard conv: kernel_size^2 * channels
  • Used in MobileNet, EfficientNet for efficient computation
  • Often followed by pointwise (1x1) convolution

Signature​

neuron DepthwiseConv(channels, kernel_size, stride=Int(1), padding=Int(0), dilation=Int(1), bias=Bool(true))

Ports​

Inputs:

  • default: [batch, channels, height, width]

Outputs:

  • default: [batch, channels, *, *]

Implementation​

Source { source: "core", path: "convolutions/DepthwiseConv" }