I appreciate the NEURAL and such features in Zorro. My request is to expand the functionality as follows:
- Provide a default Levenburg Marquadt neural network file with deep learning topology (e.g 2 Hidden Layers) and parameters *preconfigured * as a generic file in the Zorro folder. ( Basically a preconfigured Neural Network file or dll like some apps use today. )
- Allow the User to easily load this configured neural network file in the script, then specify inputs as usual, and train it, with Zorro saving the determined weights in a copy of the Net (the trained Net under some provided filena,e) and then easily use it in the future by specifying the trained Net name in a script. So long as it is in the root folder, Zorro will load it ( the trained weights are already in the file.)
- This will give users an easier way to work with a defined Topology without having to deal with the nitty gritty code configuration. Maybe even the ability to train and save multiple networks under different filenames for later retrieval.
- This would eliminate the painpoint of having to install any packages or frameworks, etc like I have to do today.
- In time, maybe even expand the list of available pre-configured Topologies in the root folder (TDNN, etc) so that all we would be doing in our script essentially is selecting which NN type we want and letting the training determine the weights.. all within the existing framework for selecting DTREE or NEURAL today.
The net benefit is ease of utilizing advanced and known Topologies as well as training, saving, and retrieving your Net. For instance, if I ran some solid training sessions and saved my work, I could retain the best 10 nets as 10 files (delete the others) in my root folder so that I can load up anyone I want in the script just by typing the corresponding file name.
Last edited by OptimusPrime; 08/09/19 00:44.