Module `Kaun.Pool`

2-D max and average pooling.

2-D pooling.

Pure, parameter-free window reductions over the last two axes — the spatial axes of the NCHW layout [| batch; channels; height; width |] used by Conv. Every function is differentiable through Rune.

Windows are placed wherever they fit entirely inside the input (no padding): pooling height with a window of kh and a stride of sh yields (height - kh) / sh + 1 positions, and likewise for the width. All leading axes are preserved.

val max_pool2d : 
  kernel_size:(int * int) ->
  ?stride:(int * int) ->
  (float, 'b) Nx.t ->
  (float, 'b) Nx.t

max_pool2d ~kernel_size:(kh, kw) x is the maximum of each kh × kw window over x's last two axes.

stride is the (vertical, horizontal) step between windows and defaults to kernel_size (non-overlapping windows).

Raises Invalid_argument if a kernel or stride component is not positive, if x has rank below 2, or if the window does not fit (height < kh or width < kw).

val avg_pool2d : 
  kernel_size:(int * int) ->
  ?stride:(int * int) ->
  (float, 'b) Nx.t ->
  (float, 'b) Nx.t

avg_pool2d ~kernel_size x is like max_pool2d except each window is reduced to its mean. Same defaults and errors as max_pool2d.