SeparableConv

Separable Convolution (Depthwise Separable)

Factorizes standard convolution into depthwise and pointwise operations. Depthwise: per-channel spatial convolution. Pointwise: 1x1 channel mixing.

Parameters:

• in_channels: Number of input channels
• out_channels: Number of output channels
• kernel_size: Size of the depthwise 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, in_channels, height, width]
• Output: [batch, out_channels, out_height, out_width]

Notes:

• Reduces computation: from kernel^2 * in * out to kernel^2 * in + in * out
• Core building block of MobileNet, Xception, EfficientNet
• Combines DepthwiseConv + 1x1 Conv (pointwise)
• Similar accuracy to standard conv with far fewer parameters

Signature

neuron SeparableConv(in_channels, out_channels, kernel_size, stride=Int(1), padding=Int(0), dilation=Int(1), bias=Bool(true))

Ports

Inputs:

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

Outputs:

• default: [batch, out_channels, *, *]

Implementation

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