RevIN

class sdm.nn.RevIN(affine_channels: int | None = None, eps: float = 1e-05, device: device | str | None = None, dtype: dtype | None = None)

Bases: Module

Reversible instance normalization from the “Reversible Instance Normalization for Accurate Time-Series Forecasting against Distribution Shift” paper.

Parameters:

• affine_channels (int | None) – The number of channels for learned affine parameters. If None, no affine transformation is applied.

• eps (float) – The value added to the supplied scale for numerical stability.

• device (device | str | None) – The device.

• dtype (dtype | None) – The dtype.

compute_stats(x: Tensor, mask: Tensor | None = None) → tuple[Tensor, Tensor]

Compute mean and standard deviation over the final dimension.

NaNs and masked values are excluded.

Parameters:

• x (Tensor) – Tensor with shape [..., C, R].

• mask (Tensor | None) – Boolean observed-value mask with shape [..., C, R].

Returns:

Mean and standard deviation with shape [..., C, 1].

Return type:

tuple[Tensor, Tensor]

forward(x: Tensor, mean: Tensor, scale: Tensor) → Tensor

The forward pass.

Parameters:

• x (Tensor) – Tensor with shape [..., C, R].

• mean (Tensor) – The mean with shape [..., C, 1].

• scale (Tensor) – The scale with shape [..., C, 1].

Return type:

Tensor

inverse(x: Tensor, mean: Tensor, scale: Tensor) → Tensor

Invert the forward pass.

Parameters:

• x (Tensor) – Normalized values or forecasts. With affine parameters, shape

• x – Tensor with shape [..., C, R].

• mean (Tensor) – The mean with shape [..., C, 1].

• scale (Tensor) – The scale with shape [..., C, 1].

Return type:

Tensor