Skip to main content

AvgPool

2D Average Pooling

Downsamples by computing the average value in each pooling window. Provides smoother downsampling compared to max pooling.

Parameters:

  • kernel_size: Size of the pooling window
  • stride: Stride of the pooling window (default: 1)
  • padding: Zero-padding added to input (default: 0)

Shape Contract:

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

Notes:

  • Output size: floor((input + 2*padding - kernel_size) / stride) + 1
  • Smoother than MaxPool, preserves more spatial information
  • Used in ResNet final pooling and some attention mechanisms
  • No learnable parameters
  • Preserves channel count

Signature​

neuron AvgPool(kernel_size, stride=Int(1), padding=Int(0))

Ports​

Inputs:

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

Outputs:

  • default: [batch, channels, *, *]

Implementation​

Source { source: "core", path: "pooling/AvgPool" }