We're wasting time training from bad initialisations (probably). Why not start with a high scoring initialisation? How this is specified in the YAML is described here.
We should have an example YAML file that can do this with our current best and hopefully we can then tune from there (increasing learning rate and decaying again?) and get a better score?
Giving a saved model in the yaml file to start training from now works. The example of how to do this is (for the time being) in alexnet_based_load_pkl.yaml.
We're wasting time training from bad initialisations (probably). Why not start with a high scoring initialisation? How this is specified in the YAML is described here.
We should have an example YAML file that can do this with our current best and hopefully we can then tune from there (increasing learning rate and decaying again?) and get a better score?