Closed hughperkins closed 9 years ago
Started work on this, in 'momentum' branch. https://github.com/hughperkins/DeepCL/tree/momentum
So, I'm thinking about how this would work / be used.
Maybe something like, on the commandline one has an option like trainer=sgd
, which at the moment would have the only one possible option, ie sgd
, but in the future could change.
Then, depending on which trainer is chosen, additional options could be available, eg momentum
, learningrate
etc, and these would be handed to the sgd somehow.
I seem to have created a layer-specific sgd class, that has to be attached to each conv and fc layer, twice actually, once for biases, and once for non-biases, but this can be done simply by the commandline parser, ie:
Done :-) In master branch.