Inconsistency in ARFClassifier's drift_detector and warning_detector Attributes When Saving and Loading

[kdMoura]

Versions

river version: 0.19.0
Python version: 3.9.18
Operating system: Ubuntu 18.04.6 LTS, Ubuntu 22.04.2 LTS

Describe the bug

I have noticed an inconsistency when using the ARFClassifier from the river library in Python. When I create an ARFClassifier object and set its drift_detector and warning_detector attributes to None, I expect them to be None when I load the object from a saved file using pickle. However, loaded_model.drift_detector and loaded_model.warning_detector are not None as I expected. Is there something I'm missing, or is this an issue?

Steps/code to reproduce

from river import forest
import pickle

model = forest.ARFClassifier(n_models=15, seed=33, leaf_prediction="mc", drift_detector=None, warning_detector=None)
# Save the model
with open('model.pkl', 'wb') as f:
    pickle.dump(model, f)

# Load the model
with open('model.pkl', 'rb') as model_file:
    loaded_model = pickle.load(model_file)

# I expect loaded_model.drift_detector and loaded_model.warning_detector to be None, but they're not.
print(loaded_model.drift_detector)
#ADWIN
print(loaded_model.warning_detector)
#ADWIN

[smastelini]

Hi @kdMoura, thanks for reporting that.

Does the problem also happen in the development version of River? We recently made some changes in ARF.

Update: as per @MaxHalford comment.

[MaxHalford]

I don't think this has anything to do with pickle. The attributes are populated when the class is instantiated:

from river import forest

model = forest.ARFClassifier(drift_detector=None, warning_detector=None)
model.drift_detector

ADWIN (
  delta=0.001
  clock=32
  max_buckets=5
  min_window_length=5
  grace_period=10
)

However, isn't this expected behavior? Using None means that default values are used. See the docs:

[smastelini]

Actually, this led me to think. We need to revisit the way we set the detectors by default to ensure users can switch off the drift detection part.

I can take care of this.

[smastelini]

Hey @kdMoura, please check out #1350 :)