fovi.arch.norm

class fovi.arch.norm.KNNChannelNorm(num_coords, num_channels, device='cuda')[source]

Bases: LayerNorm

Simple wrapper to perform normalization over the channel dimension.

Parameters:

  • num_coords (int) – Number of coordinates (unused but required for interface).

  • num_channels (int) – Number of channels to normalize over.

  • device (str, optional) – Device to place the layer on. Defaults to ‘cuda’.

__init__(num_coords, num_channels, device='cuda')[source]

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(x)[source]

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class fovi.arch.norm.KNNBatchNorm(num_coords, num_channels, device='cuda')[source]

Bases: BatchNorm1d

Simple wrapper to perform normalization over the batch dimension.

Parameters:

  • num_coords (int) – Number of coordinates (unused but required for interface).

  • num_channels (int) – Number of channels to normalize over.

  • device (str, optional) – Device to place the layer on. Defaults to ‘cuda’.

__init__(num_coords, num_channels, device='cuda')[source]

Initialize internal Module state, shared by both nn.Module and ScriptModule.

class fovi.arch.norm.KNNCoordNorm(num_coords, num_channels, device='cuda')[source]

Bases: LayerNorm

Simple wrapper to perform normalization over the coordinate dimension.

Parameters:

  • num_coords (int) – Number of coordinates to normalize over.

  • num_channels (int) – Number of channels (unused but required for interface).

  • device (str, optional) – Device to place the layer on. Defaults to ‘cuda’.

__init__(num_coords, num_channels, device='cuda')[source]

Initialize internal Module state, shared by both nn.Module and ScriptModule.

class fovi.arch.norm.KNNLayerNorm(num_coords, num_channels, device='cuda')[source]

Bases: LayerNorm

Simple wrapper to perform normalization over both coordinate and channel dimensions.

Parameters:

  • num_coords (int) – Number of coordinates to normalize over.

  • num_channels (int) – Number of channels to normalize over.

  • device (str, optional) – Device to place the layer on. Defaults to ‘cuda’.

__init__(num_coords, num_channels, device='cuda')[source]

Initialize internal Module state, shared by both nn.Module and ScriptModule.

class fovi.arch.norm.KNNGroupNorm(num_coords, num_channels, base_groups=32, min_channels_per_group=4, device='cuda')[source]

Bases: GroupNorm

Simple wrapper to perform group normalization.

Parameters:

  • num_coords (int) – Number of coordinates (unused but required for interface).

  • num_channels (int) – Number of channels to normalize over.

  • base_groups (int, optional) – Base number of groups. Defaults to 32.

  • min_channels_per_group (int, optional) – Minimum channels per group. Defaults to 4.

  • device (str, optional) – Device to place the layer on. Defaults to ‘cuda’.

__init__(num_coords, num_channels, base_groups=32, min_channels_per_group=4, device='cuda')[source]

Initialize internal Module state, shared by both nn.Module and ScriptModule.

fovi.arch.norm.find_valid_num_groups(num_channels, base_groups, min_channels_per_group=4)[source]

Find a valid number of groups for group normalization.

Parameters:

  • num_channels (int) – Total number of channels.

  • base_groups (int) – Desired base number of groups.

  • min_channels_per_group (int, optional) – Minimum channels per group. Defaults to 4.

Returns:

Valid number of groups that divides num_channels evenly.

Return type:

int

fovi.arch.norm.get_norm(norm_type, num_coords, num_channels, base_groups=32, min_channels_per_group=4, device='cuda')[source]

Get a normalization layer for a KNN layer with shape (batch, num_channels, num_coords).

Parameters:

  • norm_type (str) – Type of normalization (‘layer_chan’, ‘channel’, ‘layer_space’, ‘coord’, ‘layer_all’, ‘layer’, ‘group’, ‘batch’, or ‘none’).

  • num_coords (int) – Number of coordinates.

  • num_channels (int) – Number of channels.

  • base_groups (int, optional) – Base number of groups for group norm. Defaults to 32.

  • min_channels_per_group (int, optional) – Minimum channels per group. Defaults to 4.

  • device (str, optional) – Device to place the layer on. Defaults to ‘cuda’.

Returns:

Normalization layer or None if norm_type is ‘none’.

Return type:

nn.Module or None

Raises:

NotImplementedError – If norm_type is not implemented.