xnn.gnn.models.gemnet.he_orthogonal_#

xnn.gnn.models.gemnet.he_orthogonal_(weight)[source]#

He-scaled orthogonal initialization with exact standardization (paper section 6).

A random (semi-)orthogonal matrix is standardized per output unit to zero mean and unit variance and scaled to variance 1 / fan_in.

Parameters:

weight (Tensor) – A (fan_out, fan_in) weight matrix, modified in place.

Returns:

weight.

Return type:

Tensor