braindecode.functional.masked_group_norm#

braindecode.functional.masked_group_norm(x, num_groups, mask=None, weight=None, bias=None, eps=1e-05)[source]#

Group normalization with statistics over the entries a mask flags.

For a group norm whose statistics span several channels of a sample, so that padded channels do not shift those of the real ones.

Parameters:

x (Tensor) – Tensor of shape (batch, channels, ...), as for torch.nn.functional.group_norm().
num_groups (int) – Number of groups the channels are split into.
mask (Optional[Tensor]) – Boolean mask over the batch and the leading axes after the channels (e.g. (batch, rows) for x of shape (batch, channels, rows, width)), True for the entries that enter the statistics; it is broadcast over the remaining axes. None is torch.nn.functional.group_norm().
weight (Optional[Tensor]) – Per-channel affine parameters, of shape (channels,).
bias (Optional[Tensor]) – Per-channel affine parameters, of shape (channels,).
eps (float) – Added to the variance for numerical stability.

Returns:

x normalized, same shape. Masked entries are zero, including after the affine transform, and receive zero input gradients. Their values (even NaN) affect neither the statistics nor the affine gradients.

Return type:

Tensor