Hyperas indention Error: Can't define inner functions

[Naghipourfar]

Dear @maxpumperla,

Hyperas can't actually support inner functions in python and it's frustrating because when you want to write code for a single model in order to hyperopt for multiple datasets. It is impossible. I will be very grateful if this feature can add to hyperas or if it already exists, you help me how to use it.

Thanks, Mohsen Please see the following code:

def outer_data(data_name):
    def inner_data():
        train_data = pd.read_csv(f"./data/{data_name}/train_{data_name}.csv").values
        valid_data = pd.read_csv(f"./data/{data_name}/valid_{data_name}.csv").values
        return train_data, valid_data
    return inner_data

def model(train_data, valid_data):
    # some keras model 

best_run, best_model = optim.minimize(model=model,
                                      data=outer_data("mnist"),
                                      algo=tpe.suggest,
                                      max_evals=50,
                                      trials=Trials())

[maxpumperla]

@Naghipourfar I'm not entirely sure what you're trying to do here, but why can't you flatten this?

def data():
    data_name = # some data name
    train_data = pd.read_csv(f"./data/{data_name}/train_{data_name}.csv").values
    valid_data = pd.read_csv(f"./data/{data_name}/valid_{data_name}.csv").values
    return train_data, valid_data

def model(train_data, valid_data):
    # some keras model 

best_run, best_model = optim.minimize(model=model,
                                      data=data,
                                      algo=tpe.suggest,
                                      max_evals=50,
                                      trials=Trials())

I don't quite see why it makes much of a difference between changing your data_name in optim.minimize or in the data function. You have to modify it in one place regardless.

[maxpumperla]

p.s. I do understand why you want this behavior, but hyperas' design has some limitations. If it's just about functionality, use my snippet above.

[Naghipourfar]

@maxpumperla actually, I had no choice so I wrote something like your snippet. Thanks for your help.