Pooling modules

Pooling modules wrap local relation pooling and batch-wise global pooling. Local pooling returns sparse tensors; global pooling returns dense batch rows. TrilinearUpsample3d provides parameter-free normalized interpolation on generated or explicit target support.

Related pages

• Functional pooling API: Pooling operations

• Pooling backend routes: Pooling routes

• Relation model: Coordinates and relations

• Batch metadata: Sparse tensor model

Module summary

Module	Reduction	Output type
Pool3d	Configurable sum, max, or avg.	SparseTensor
SumPool3d / MaxPool3d / AvgPool3d	Fixed local reduction mode.	SparseTensor
PoolTranspose3d	Contributor average on generated or explicit target support.	SparseTensor
TrilinearUpsample3d	Normalized separable linear interpolation.	SparseTensor
GlobalSumPool / GlobalAvgPool / GlobalMaxPool	Batch-wise dense reduction.	MLX array

class mlx_lattice.nn.pool.AvgPool3d(*, kernel_size=2, stride=2, padding=0, dilation=1)

Bases: Pool3d

Local sparse average-pooling module.

Parameters:

• kernel_size (int | Sequence[int])

• stride (int | Sequence[int])

• padding (int | Sequence[int])

• dilation (int | Sequence[int])

class mlx_lattice.nn.pool.GlobalAvgPool

Bases: Module

Batch-wise global average-pooling module returning dense (B, C) rows.

class mlx_lattice.nn.pool.GlobalMaxPool

Bases: Module

Batch-wise global max-pooling module returning dense (B, C) rows.

class mlx_lattice.nn.pool.GlobalSumPool

Bases: Module

Batch-wise global sum-pooling module returning dense (B, C) rows.

class mlx_lattice.nn.pool.MaxPool3d(*, kernel_size=2, stride=2, padding=0, dilation=1)

Bases: Pool3d

Local sparse max-pooling module.

Parameters:

• kernel_size (int | Sequence[int])

• stride (int | Sequence[int])

• padding (int | Sequence[int])

• dilation (int | Sequence[int])

class mlx_lattice.nn.pool.Pool3d(*, mode='sum', kernel_size=2, stride=2, padding=0, dilation=1)

Bases: Module

Configurable local sparse 3D pooling module.

mode selects sum, max, or avg reduction over a sparse kernel relation. The module returns a sparse tensor with output stride multiplied by the pooling stride.

Parameters:

• mode (PoolMode)

• kernel_size (int | Sequence[int])

• stride (int | Sequence[int])

• padding (int | Sequence[int])

• dilation (int | Sequence[int])

class mlx_lattice.nn.pool.PoolTranspose3d(*, kernel_size=2, stride=2, padding=0, dilation=1)

Bases: Module

Average-pooling transpose onto generated or explicit target support.

Parameters:

• kernel_size (int | Sequence[int])

• stride (int | Sequence[int])

• padding (int | Sequence[int])

• dilation (int | Sequence[int])

class mlx_lattice.nn.pool.SumPool3d(*, kernel_size=2, stride=2, padding=0, dilation=1)

Bases: Pool3d

Local sparse sum-pooling module.

Parameters:

• kernel_size (int | Sequence[int])

• stride (int | Sequence[int])

• padding (int | Sequence[int])

• dilation (int | Sequence[int])

class mlx_lattice.nn.pool.TrilinearUpsample3d(*, stride=2)

Bases: Module

Normalized trilinear upsampling on generated or target support.

Parameters:

stride (int | Sequence[int])