SNNPy Alpha 3.0.0

The neural network modelling software currently has the following features.

• Xavier initiation of weights using box mueller transform to generate normally distributed random variables.
• Uniform random initialization of weights.
• Uniform random initialization of weights.
• Simple forward propogation algorithm.
• Simple back propogation algorithm.
• Batch training.
• Batch randomzitaion (ensuring that the training examples arent always in the same order).
• 4 activation functions (linear, sigmoid, ReLU, TanH).
• Possibility to use different activation functions in the input, hidden and output layer.
• 2 cost functions (cross entropy cost, and mean square cost).
• Configurable learning rates.
• Fully configurable feed forward network structures (only requirement being there should be at least 1 input, 1 hidden layer neuron and 1 output neuron).
• Gradient batch averaging.
• L1 and L2 Regularization