ml.models.norms

Defines general-purpose helper functions for initializing norm layers.

from ml.models.norms import get_norm_linear, get_norm_1d, get_norm_2d, get_norm_3d, cast_norm_type

linear = nn.Sequential(nn.Linear(32, 32), get_norm_linear("layer", dim=32))
conv_1d = nn.Sequential(nn.Conv1d(32, 32, 3), get_norm_1d("layer", dim=32, groups=4))
conv_2d = nn.Sequential(nn.Conv2d(32, 32, 3), get_norm_2d("layer", dim=32, groups=4))
conv_3d = nn.Sequential(nn.Conv3d(32, 32, 3), get_norm_3d("layer", dim=32, groups=4))

# This lets you parametrize the norm type as a string.
linear = nn.Sequential(nn.Linear(32, 32), get_norm_linear(cast_norm_type(my_norm), dim=32))

Choices for the norm type are:

  • "no_norm": No normalization

  • "batch" or "batch_affine": Batch normalization

  • "instance" or "instance_affine": Instance normalization

  • "group" or "group_affine": Group normalization

  • "layer" or "layer_affine": Layer normalization

Note that instance norm and group norm are not available for linear layers.

ml.models.norms.cast_norm_type(s: str) → Literal['no_norm', 'batch', 'batch_affine', 'instance', 'instance_affine', 'group', 'group_affine', 'layer', 'layer_affine'][source]
ml.models.norms.cast_parametrize_norm_type(s: str) → Literal['no_norm', 'weight', 'spectral'][source]
class ml.models.norms.LastBatchNorm(channels: int, momentum: float = 0.99, affine: bool = True, eps: float = 0.0001, device: device | None = None, dtype: dtype | None = None)[source]

Bases: Module

Applies batch norm along final dimension without transposing the tensor.

The normalization is pretty simple, it basically just tracks the running mean and variance for each channel, then normalizes each channel to have a unit normal distribution.

Input:

x: Tensor with shape (…, N)

Output:

The tensor, normalized by the running mean and variance

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

mean: Tensor
var: Tensor
forward(x: Tensor) → Tensor[source]

Defines 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 ml.models.norms.ConvLayerNorm(channels: int, *, dims: int | None = None, eps: float = 1e-05, elementwise_affine: bool = True, device: device | None = None, dtype: dtype | None = None)[source]

Bases: Module

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

reset_parameters() → None[source]
forward(inputs: Tensor) → Tensor[source]

Defines 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.

ml.models.norms.get_norm_1d(norm: Literal['no_norm', 'batch', 'batch_affine', 'instance', 'instance_affine', 'group', 'group_affine', 'layer', 'layer_affine'], *, dim: int | None = None, groups: int | None = None, device: device | None = None, dtype: dtype | None = None) → Module[source]

Returns a normalization layer for tensors with shape (B, C, T).

Parameters:

  • norm – The norm type to use

  • dim – The number of dimensions in the input tensor

  • groups – The number of groups to use for group normalization

  • device – The device to use for the layer

  • dtype – The dtype to use for the layer

Returns:

A normalization layer

Raises:

NotImplementedError – If norm is not a valid 1D norm type

ml.models.norms.get_norm_2d(norm: Literal['no_norm', 'batch', 'batch_affine', 'instance', 'instance_affine', 'group', 'group_affine', 'layer', 'layer_affine'], *, dim: int | None = None, groups: int | None = None, device: device | None = None, dtype: dtype | None = None) → Module[source]

Returns a normalization layer for tensors with shape (B, C, H, W).

Parameters:

  • norm – The norm type to use

  • dim – The number of dimensions in the input tensor

  • groups – The number of groups to use for group normalization

  • device – The device to use for the layer

  • dtype – The dtype to use for the layer

Returns:

A normalization layer

Raises:

NotImplementedError – If norm is not a valid 2D norm type

ml.models.norms.get_norm_3d(norm: Literal['no_norm', 'batch', 'batch_affine', 'instance', 'instance_affine', 'group', 'group_affine', 'layer', 'layer_affine'], *, dim: int | None = None, groups: int | None = None, device: device | None = None, dtype: dtype | None = None) → Module[source]

Returns a normalization layer for tensors with shape (B, C, D, H, W).

Parameters:

  • norm – The norm type to use

  • dim – The number of dimensions in the input tensor

  • groups – The number of groups to use for group normalization

  • device – The device to use for the layer

  • dtype – The dtype to use for the layer

Returns:

A normalization layer

Raises:

NotImplementedError – If norm is not a valid 3D norm type

ml.models.norms.get_norm_linear(norm: Literal['no_norm', 'batch', 'batch_affine', 'instance', 'instance_affine', 'group', 'group_affine', 'layer', 'layer_affine'], *, dim: int | None = None, device: device | None = None, dtype: dtype | None = None) → Module[source]

Returns a normalization layer for tensors with shape (B, …, C).

Parameters:

  • norm – The norm type to use

  • dim – The number of dimensions in the input tensor

  • device – The device to use for the layer

  • dtype – The dtype to use for the layer

Returns:

A normalization layer

Raises:

NotImplementedError – If norm is not a valid linear norm type

ml.models.norms.get_parametrization_norm(module: T_module, norm: Literal['no_norm', 'weight', 'spectral'], *, name: str = 'weight', n_power_iterations: int = 1, eps: float = 1e-12, weight_dim: int = 0, spectral_dim: int | None = None) → T_module[source]

Returns a parametrized version of the module.

Parameters:

  • module – The module to parametrize

  • norm – The parametrization norm type to use

  • name – The name of the parameter to use for the parametrization; this should reference the name on the module (for instance, weight for a nn.Linear module)

  • n_power_iterations – The number of power iterations to use for spectral normalization

  • eps – The epsilon value to use for spectral normalization

  • weight_dim – The dimension of the weight parameter to normalize when using weight normalization

  • spectral_dim – The dimension of the weight parameter to normalize when using spectral normalization

Returns:

The parametrized module