Orthogonal¶
-
class
Orthogonal(scale=1.414, rand_type='uniform')¶ Initialize weight as orthogonal matrix.
This initializer implements Exact solutions to the nonlinear dynamics of learning in deep linear neural networks, available at https://arxiv.org/abs/1312.6120.
Parameters: - scale : float optional
Scaling factor of weight.
- rand_type: string optional
Use “uniform” or “normal” random number to initialize weight.