When running with --mode single, it loads the dag given by --dag_path (which is a simple json dump) and trains this dag.
In this mode, there is no "autoML" part. The controller is not trained and the same dag is used during the "train_shared" phase and validate phase.
I provided a dag.json file, which is the best dag found in the paper. This mode should allow us to "retrain from scratch" and eventually reach the result of the paper (working on it...).
Major changes are in the Trainer class. I added a single=False optional parameter on the train() method and a dag=None optional parameter on the train_shared() method. This pull does not break backward compatibility.
carpedm20
commented
5 years ago
When running with
--mode single, it loads the dag given by--dag_path(which is a simple json dump) and trains this dag.In this mode, there is no "autoML" part. The controller is not trained and the same dag is used during the "train_shared" phase and validate phase.
I provided a
dag.jsonfile, which is the best dag found in the paper. This mode should allow us to "retrain from scratch" and eventually reach the result of the paper (working on it...).Major changes are in the Trainer class. I added a
single=Falseoptional parameter on thetrain()method and adag=Noneoptional parameter on thetrain_shared()method. This pull does not break backward compatibility.