TODO.md

🏁 TODOs

IMPORTANT

• New costs functions
• Batch in train algorithm
• Add optimizers
• Add Dropout layer

• FIX INPUTS/OUTPUTS DIMENSIONS
• Create custom Cost
• Create custom Activation functions
• Fix problems with activation register
• Add early stopping
• Check if clone is necessary
• Change train API to use config struct instead of arguments (TrainConfig)
• Flatten layer
• Separates from/to MSGPACK logic from Layer trait
• Try to use only one global register isntead of three
• Separates builder and gloabl register logic
• Create register! macro to register all layers, activations and costs
• Create derive macro for CostFunction and ActivationFunction
• Create derive macro for MSGPackFormatting
• Generalize CostFunction and ActivationFunction traits (NNUtils)
• Check docs
• Check README

Layers implementations

• Add BatchNorm layer
• Add Conv layer
• Add Pooling layer
• Add Deconv layer
• Add Embedding layer
• Add Recurrent layer

• Improve backpropagation (resilient propagation)
• Multithreading
• Try to use GPU (WGPU, torch, etc)

NOT IMPORTANT

• See erased-serde
• Refactoring train algorithm
• Fix Adam optimizer
• Web for docs and examples
• Allow users to set type of the numbers used in the neural network (f32 or f32)