dereklstinson / gocunets Goto Github PK

Add support methods to satisfy gonum.graph interfaces.

Either that or my rtx 2080 ti card is junk. It doesn't work with any of my c602 boards. It works with my x79 board (but at a 3x performance loss to the equivalent 1080ti). I might try it on my z97 board. The only solution I can think of is to move some of the auto build stuff to the top.

Layers is like a wrapper of a wrapper. These can move to the nvidia cudnn folder. The nvidia cudnn folder can have all its sub-folders merged. Having all the sub-folders made sense when first created, but to beginners trying to use the API it could hard to find the operations.

The structures within the trainable layers contain have an in-bedded trainer. This needs to go. The trainers need to be a higher level operation.

That also means the trainer interface needs to go. Instead, for all the trainers they should receive the tensors that they use to train. Most trainers would use need the delta weights and weights.

Should this be implemented on GoCuNets? What if custom Modules are built? Maybe a Module will need its own save json. Really it only needs the name of the module, weight values, and alpha,beta values.