Improving the design
Andy-P opened this issue · comments
Starting a thread to improve the overall design. I think this approach of using graph.jl and solver.jl to create arbitrary deep learning models is really interesting and could evolve into something really useful to the community.
Just putting together a list of possible improvement/ideas in no particular order
- The input to the LSTM and RNN models shouldn't have to be a NNMatrix
- The graph.jl and solver.jl might better as a separate library/dependancy (NNGraph.jl?)
- The Abstract type Model is too generic a name. Change to NNModel?
Will add more as they surface.