ray-project / tune-sklearn

A drop-in replacement for Scikit-Learn’s GridSearchCV / RandomizedSearchCV -- but with cutting edge hyperparameter tuning techniques.
https://docs.ray.io/en/master/tune/api_docs/sklearn.html
Apache License 2.0
467 stars 51 forks source link

"training_iteration" from TuneSearchCV is always 1, and accuracy does not improve over time #263

Open gbrookshire opened 1 year ago

gbrookshire commented 1 year ago

I'm using tune_sklearn.TuneSearchCV to tune a fairly large number of hyperparameters using the "hyperopt" optimization method. When I look at the results in the cv_results_ attribute, the model performance does not improve over the course of hyperparameter tuning. I also see that "training_iteration" is always 1.

Does this mean that TuneSearchCV is not taking prior search trials into account? Is it essentially behaving like a random search, and ignoring the Bayesian optimization?

Here's an illustration of the kind of code I'm using:

estimator = << an sklearn pipeline >>
search_spaces = {
        "mdl__alpha": ray.tune.loguniform(1e-7, 1e2)  # Lots of hyperparameters like this
}
search = TuneSearchCV(
            estimator,
            search_spaces,
            search_optimization="hyperopt",
            n_trials=64,
            n_jobs=7,
            scoring="balanced_accuracy"
            refit=True)

search.fit(X, y)  # X and y are the data and labels

cv_results = pd.DataFrame(search.cv_results_)  # "search" is a fitted instance of TuneSearchCV
cv_results['training_iteration']

## RESULT
0     1
1     1
2     1
3     1
4     1
     ..
59    1
60    1
61    1
62    1
63    1
Name: training_iteration, Length: 64, dtype: int64

cv_results['mean_test_balanced_accuracy']

## RESULT
0     0.773810
1     0.761905
2     0.654762
3     0.714286
4     0.790476
        ...   
59    0.747619
60    0.733333
61    0.771429
62    0.680952
63    0.745238
Name: mean_test_balanced_accuracy, Length: 64, dtype: float64