He Normal Initialiser¶
he_normal_init_type
he_normal_init_type()
Draws samples from a normal distribution designed for layers with ReLU activation.
\[W \sim \mathcal{N}\left(0, \sqrt{\frac{2}{n_{in}}}\right)\]
where \(n_{in}\) is the number of input units.
This initialisation is specifically designed for networks using ReLU activations and helps prevent vanishing/exploding gradients.
Shape:¶
Initialises weights based on the shape provided during layer setup.