Hyper-parameters

The hyper-parameter optimization is performed using Optuna. It can be done in two steps as follows:

1. Declare the hyper-parameter set

Declare the set of parameter to be optimized inside the model through the definition of a static function return a dictionnary of values. The following Listing provide an example of such function suggesting parameters for the window size, the normlization type, as well as the :

@staticmethod
def suggest_hparams(self, trial):

    norm_ = trial.suggest_categorical('normalize', ['z-norm', 'lognorm'])
    window_length = trial.suggest_int('in_size', low=50, high=1800)
    window_length += 1 if window_length % 2 == 0 else 0
    return {
        'input_norm': norm_,
        'output_norm':norm_,
        'in_size': window_length,
        }

2. Specify the use of Optuna

 experiment = {
   'power': {'mains': ['active'],'appliance': ['active']},
   'sample_rate': 6,
   'appliances': [
         'fridge',
         'washing machine',
             'dish washer',
       ],
   'artificial_aggregate': False,
   'DROP_ALL_NANS': True,
   'methods': {
           'WAVENILM': NILMExperiment({
                "model_name": 'WAVENILM',
                'context_size': 481,
                'input_norm':'z-norm',
                'target_norm':'z-norm',
                'feature_type':'mains',
                'max_nb_epochs':max_nb_epochs,
                'use_optuna':True,
                }),
   },

   'train': {
     'datasets': {
      data: {
         'path': data_path,
         'buildings': {
               1: {
                     'start_time': '2015-01-04',
                     'end_time': '2015-03-30'
           }
         }
      }
     }
   },
     'test': {
     'datasets': {

       data: {
         'path': data_path,
         'buildings': {
               1: {
                     'start_time': '2015-04-16',
                     'end_time': '2015-05-15'
                 }
         }
       }
     },
         'metrics':['mae','nde','f1score', 'rmse']
     }

 }

Note

The use of optuna will generate several models and they will all have saved as checkpoints in the corresponding folder splitted into different subfolders labelled according to the trial ID.