Initialisers

Weight initialisation is crucial for training neural networks effectively. Proper initialisation helps prevent vanishing or exploding gradients and can significantly impact convergence speed. The athena library provides various initialisation strategies suited for different activation functions and network architectures.

Available Initialisers

• Glorot Uniform Initialiser
• Glorot Normal Initialiser
• He Uniform Initialiser
• He Normal Initialiser
• LeCun Uniform Initialiser
• LeCun Normal Initialiser
• Zeros Initialiser
• Ones Initialiser
• Identity Initialiser
• Gaussian Initialiser

Creating custom initialisers

The athena library is designed with extensibility in mind, allowing users to create custom initialisers by extending the base_initialiser_type.

See the tutorial: Creating Custom Initialisers