Conv3d

3D Convolutional Layer

Applies 3D convolution over volumetric data (video, 3D medical imaging).

Parameters:

• in_channels: Number of channels in input
• out_channels: Number of channels produced by convolution
• kernel_size: Size of the convolving kernel (int or (d, h, w))
• stride: Stride of the convolution (default: 1)
• padding: Zero-padding added to all sides (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, depth, height, width]
• Output: [batch, out_channels, out_depth, out_height, out_width]

Notes:

• Output size per dim: (size + 2padding - dilation(kernel-1) - 1) / stride + 1
• Used for video understanding, 3D medical imaging, point clouds
• Memory intensive due to 5D tensors

Signature

neuron Conv3d(in_channels, out_channels, kernel_size, stride=Int(1), padding=Int(0), dilation=Int(1), groups=Int(1), bias=Bool(true))

Ports

Inputs:

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

Outputs:

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

Implementation

Source { source: "core", path: "convolutions/Conv3d" }