foscat.SphereDownGeo#

Classes#

SphereDownGeo

Geometric HEALPix downsampling operator (NESTED indexing).

Module Contents#

class foscat.SphereDownGeo.SphereDownGeo(nside_in: int, mode: str = 'smooth', radius_deg: float | None = None, sigma_deg: float | None = None, weight_norm: str = 'l1', cell_ids_out: numpy.ndarray | list[int] | None = None, in_cell_ids: numpy.ndarray | list[int] | torch.Tensor | None = None, use_csr=True, device=None, dtype: torch.dtype = torch.float32)[source]#

Bases: torch.nn.Module

Geometric HEALPix downsampling operator (NESTED indexing).

This module reduces resolution by a factor 2:

nside_out = nside_in // 2

Input conventions#

  • If in_cell_ids is None:

    x is expected to be full-sphere: [B, C, N_in] output is [B, C, K_out] with K_out = len(cell_ids_out) (or N_out if None).

  • If in_cell_ids is provided (fine pixels at nside_in, NESTED):
    x can be either:

    • compact: [B, C, K_in] where K_in = len(in_cell_ids), aligned with in_cell_ids order

    • full-sphere: [B, C, N_in] (also supported)

    output is [B, C, K_out] where cell_ids_out is derived as unique(in_cell_ids // 4), unless you explicitly pass cell_ids_out (then it will be intersected with the derived set).

Modes#

  • mode=”smooth”: linear downsampling y = M @ x (M sparse)

  • mode=”maxpool”: non-linear max over available children (fast)

device = None#
dtype = Ellipsis#
nside_in#
nside_out#
N_in#
N_out#
mode = ''#
weight_norm = ''#
in_cell_ids = None#
has_in_subset#
K_out#
K_in#
forward(x: torch.Tensor)[source]#
Parameters:

x (torch.Tensor) –

If has_in_subset:

  • [B,C,K_in] (compact, aligned with in_cell_ids) OR [B,C,N_in] (full sphere)

Else:

  • [B,C,N_in] (full sphere)

Returns:

  • y (torch.Tensor) – [B,C,K_out]

  • cell_ids_out (torch.Tensor) – [K_out] coarse pixel ids (nside_out), aligned with y last dimension.