Skip to main content

AdaptiveMaxPool

Adaptive Max Pooling

Pools input to a fixed output size using max operation. Automatically calculates kernel size and stride to achieve target output.

Parameters:

  • output_size: Target spatial size (height and width will both be this value)

Shape Contract:

  • Input: [batch, channels, height, width] (any spatial size)
  • Output: [batch, channels, output_size, output_size]

Notes:

  • Accepts any input spatial size, outputs fixed size
  • output_size=1 is common for global max pooling
  • Used before fully-connected layers to handle variable input sizes
  • No learnable parameters
  • Preserves strongest activations (max values)

Signature​

neuron AdaptiveMaxPool(output_size)

Ports​

Inputs:

  • default: [batch, channels, *, *]

Outputs:

  • default: [batch, channels, output_size, output_size]

Implementation​

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