pyhgf.model.conv.avg_pool_adapter

pyhgf.model.conv.avg_pool_adapter(pool_size=(2, 2), stride=None)

Build a frozen average-pool: block-mean forward, uniform-split backward.

Forward, each non-overlapping (or strided) window is averaged via jax.lax.reduce_window. The map is linear, so the backward is its transpose in closed form, as in linear_adapter(): the incoming gradient at one pooled position is split evenly back across every position that fed it. That backward is continuous everywhere, unlike max_pool_adapter()’s, whose routing jumps between taps as the winner changes.

Parameters:

• pool_size (tuple[int, int])

• stride (tuple[int, int] | None)

Return type:

EquinoxAdapter