Skip to main content

GlobalMaxPool

Global Max Pooling

Reduces each channel to a single value by taking the maximum across all spatial positions.

Shape Contract:

  • Input: [batch, channels, height, width] (any spatial size)
  • Output: [batch, channels, 1, 1] single value per channel

Notes:

  • Equivalent to AdaptiveMaxPool(1)
  • Reduces spatial dimensions to 1x1
  • No learnable parameters
  • Preserves strongest activation in each channel
  • Works with any input resolution
  • Less common than GlobalAvgPool but useful for detection tasks

Signature​

neuron GlobalMaxPool()

Ports​

Inputs:

  • default: [batch, channels, *, *]

Outputs:

  • default: [batch, channels, 1, 1]

Implementation​

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