TransposedConv

Transposed Convolution (Deconvolution)

Upsampling convolution that increases spatial dimensions. Also known as deconvolution or fractionally-strided convolution.

Parameters:

• in_channels: Number of input channels
• out_channels: Number of output channels
• kernel_size: Size of the convolving kernel
• stride: Stride of the convolution (default: 1)
• padding: Zero-padding added to both sides (default: 0)
• output_padding: Additional size added to output (default: 0)
• dilation: Spacing between kernel elements (default: 1)
• groups: Number of blocked connections (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:

• Output size: (input - 1) * stride - 2*padding + kernel_size + output_padding
• Used in decoder networks, GANs, segmentation
• Learnable upsampling (vs interpolation)
• Can cause checkerboard artifacts if not used carefully

Signature

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

Ports

Inputs:

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

Outputs:

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

Implementation

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