MetaTune

MetaTune framework implements simultaneous model selection and hyperparameter tuning of scikit-learn supervised learning algorithms with optuna implemented metaheuristic search algorithms

---

Getting Started

---

MetaTune implements user customizable tuners for all sci-kit-learn supervised learning algorithms, with it one can easily sample the best model out of a search space of all scikit-learn algorithms and their corresponding hyperparameters with relative ease with only a few lines of code.

Importing the MetaTune(...) class is as easy as:

from metatune import MetaTune

tuner = MetaTune(task="classification")

We can further use this object to select the machine learning algorithm and the corresponding parameters that best model our data like so:

import optuna
from sklearn import datasets
from sklearn.preprocessing import MinMaxScaler
from sklearn.model_selection import train_test_split
from sklearn.metrics import f1_score, recall_score, precision_score


# load classification dataset
X, y = datasets.load_iris(return_X_y=True)

# define data scaler
scaler = MinMaxScaler()

# split data to training and evaluation set
X_train, X_eval, y_train, y_eval = train_test_split(X, y, random_state=0, shuffle=True)

#scaler features
scaler.fit(X_train)
X_train = scaler.transform(X_train)
X_eval = scaler.transform(X_eval)


def objective(trial):
    model = tuner.sample_models_with_params(trial)
    model.fit(X_train, y_train)

    pred = model.predict(X_eval)
    f1 = f1_score(y_eval, pred, average="micro")
    recall = recall_score(y_eval, pred, average="micro")
    precision = precision_score(y_eval, pred, average="micro")

    return (f1 + recall + precision) / 3


# create optuna study
study = optuna.create_study(
    study_name="classifier_model",
    direction="maximize",
    pruner=optuna.pruners.MedianPruner(), # pruner is necessary to prune bad parameter combinations of models when sampling
    sampler=optuna.samplers.TPESampler(),
)

study.optimize(objective, n_trials=50)

#output
...
# [I 2023-06-01 07:19:01,678] Trial 27 finished with value: 0.8947368421052632 and parameters: {'model_tuner': 'AdaBoostClassifierTuner', #'AdaBoostClassifierTuner_estimator': None, 'AdaBoostClassifierTuner_n_estimators': 90, 'AdaBoostClassifierTuner_learning_rate': #              0.12391614619885151, 'AdaBoostClassifierTuner_algorithm': 'SAMME.R', 'AdaBoostClassifierTuner_random_state': 7705}. Best is trial 1 with value: # 0.9736842105263158.
...

In this task, we hope to find the best classification algorithm and its hyperparameters that best maximizes the objective, which happens to be the average sum of f1, recall and precision metrics.

After running this, we can retrieve the best optuna trial and build a model out of it for further fine-tuning with more data, like so:

sampled_model = tuner.build_sampled_model(study.best_trial)

NOTE that the models returned are purely scikit-learn models, thus the reason the regular fit(...) and predict(...) methods can be called on them.

NOTE: A thing to note is that for large datasets, model and hyperparameter search ought to be used on a subset of the data rather than the entire dataset. The reason for this being that the aim of hyperparameter tuning is not to fit all of the data, but rather to sample the parameters that best model the data. After tuning, you can instantiate the sampled model with the corresponding sampled parameters and fine-tune the model on the large dataset, this way you avoid the unnecessary computation required to model the entirety of the available dataset.

---

Handling Model Data Incompatibility

---

In some cases, not all scikit-learn models will be compatible with your data. In such cases, you can do one of two things

1. Calling the only_compatible_with_data(...) method:

This method takes in three arguments, X, y and probability_score. X and y correspond to the training data, this attempts a data fit on all models (initialized with their default parameters) in the search space, it exempts models that are not compatible to the dataset, hence reducing the search space.

NOTE:, Some models may not actually be incompatible with your dataset, and a model may be exempted simply because the default parameters being used raises an exception when an attempt to fit the model on the data is made. so it may not be suitable to use this method in some cases, hence the reason to opt for the second option. If you wish to also implement custom tuners that do not exist amongst the implemented library of tuners, then using this method is not advisable.

The probability_score (default=False) when set to True remove models that do not implement the predict_proba(...) method, this can be handy if you wish to maximise an objective that relies on the predicted probability scores for each class rather than predicted labels. This argument is only effective if task = "classification".

NOTE: Some models may not have the predict_proba(...) readily available until certain conditions are met in the parameter combination used to initialise the model object. For example, the SVC class objects will only have this method if the probability parameter is set to True.

2. raising an optuna.exceptions.TrialPruned(e)

Calling the .fit(...) predict(...) or predict_proba(...) method on your dataset may throw an exception regardless of whether the combination of sampled parameters is correct, for example, calling the .fit(...) method of a naive bayes classifier model on dataset with non-positive features will raise an exception, or in some cases, the sampled nu parameter for the NuSVC or NuSVR model may not be compatible with your dataset and may raise an exception. In cases like these the optimization process will be terminated due to these unforseen errors. To handle them, you will need to catch the raised exception and reraise it as an optuna.exceptions.TrialPruned exception instead. You can do this like so:

...
try:
    model.fit(X_train, y_train)

except Exception as e:
   raise optuna.exceptions.TrialPruned(e)

Depending on the error, you might also want to put the pred = model.predict(X_eval) line in the try block for this same reason, this is because some models may fit on your data but due to data incompatibility may throw an exception during prediction. An example of this can be observed when trying to fit a categorical Naive Bayes model on data with continous numerical feature column vectors instead of categorical feature column vectors.

This way, instead of terminating the metaheuristic search program, optuna simply skips to the next trial and gives you a log relating to why the trial was pruned. Similarly, if you happen to want to optimize an objective the relies on predicted probability scores for each class, rather than the class label, you can call the predict_proba(...) method on X_train in the try block, just after called the fit(...) method.

NOTE: You must first ensure that your dataset is properly formatted and correct before utilising the try block this way.

---

Advanced Configurations

In the MetaTuner(..) class, there are four more arguments alongside the task argument used for advanced configuration when initialising a MetaTuner(..) class, namely: custom_tuners, custom_only, excluded and single_tuner.

1. custom_tuners Argument: Optional[Iterable[BaseTuner]] (Default=None)

The custom_tuners argument is used to specify custom tuners or overwrite the default parameter space of existing tuners in the sample space. If you wish to extend or shrink the sample space of a given parameter in tuner, you can do so as shown:

from metatune import MetaTune
from metatune.tune_classifier import NuSVCTuner, SGDClassifierTuner

nu_svc_tuner = NuSVCTuner(
    nu_space={"low":0.1, "high":0.4, "step":None, "log":False}
)

sgd_classifier_tuner = SGDClassifierTuner(
    loss_space=('hinge', 'log_loss', 'modified_huber', 'squared_hinge')
)

tuner = MetaTune(task="classification", custom_tuners=[nu_svc_tuner, sgd_classifier_tuner])

You can view all the tuners in the search space with the search_space attribute of the MataTune instance like so:

tuner.search_space

2. custom_only Argument: bool (Default=False)

If set to True, this argument makes it such that only the list of tuners specified in the custom_tuners argument make up the search space

3. excluded: Optional[Iterable[Union[str, Callable]]] (default=None)

This argument is used to specify a list of tuners to exempt from the search space. It takes as value, a list of strings or Callable types corresponding to class names or class of tuners being excluded, it can be used as shown:

from metatune.tune_classifier import NuSVCTuner

tuner = MetaTune(task="classification", excluded=[NuSVCTuner, "SGDClassifierTuner"])

4. single_tuner: Optional[BaseTuner] (default=None)

This is used when you wish to perform parameter searching for a single specific model class with its corresponding tuner. If specified, only the tuner passed as argument will exist in the search space. You can use this to specify a custom tuner or an already implemented tuner with or without overwriting the parameter search spaces as shown:

from metatune.tune_classifier import GradientBoostingClassifierTuner

tuner = MetaTune(task="classification", single_tuner=GradientBoostingClassifierTuner())

---

Creating a Custom Tuner

You may want to perform hyperparameter sampling for a given machine learning algorithm whose tuner is not part of the library of implemented tuners, in such case, you can opt for creating your own custom tuner as shown in the example below where we implement a tuner for gaussian process classifier.

from optuna.trial import Trial
from dataclasses import dataclass
from metatune.baseline import BaseTuner
from sklearn.gaussian_process import GaussianProcessClassifier
from sklearn.gaussian_process.kernels import RBF, WhiteKernel, ExpSineSquared
from typing import Callable, Dict, Iterable, Any, Optional
from types import MappingProxyType
from metatune import MetaTune

@dataclass
class CustomGPCTuner(BaseTuner):
    model_class: Callable = GaussianProcessClassifier
    kernel_space: Iterable[object] = (RBF(), WhiteKernel(), ExpSineSquared())
    max_iter_predict_space: Dict[str, Any] = MappingProxyType({
        "low":10, "high":300, "step":1, "log":True
    })

    def sample_params(self, trial: Optional[Trial]=None) -> Dict[str, Any]:
        super().sample_params(trial)
                
        params = {}
        params["kernel"] = trial.suggest_categorical(f"{self.__class__.__name__}_kernel", self.kernel_space)
        params["max_iter_predict"] = trial.suggest_int(
            f"{self.__class__.__name__}_max_iter_predict", **dict(self.max_iter_predict_space))
        
        return params

    def sample_model(self, trial: Optional[Trial]=None) -> Any:
        super().sample_model(trial)
        params = self.sample_params(trial)
        model = super().evaluate_sampled_model("classification", self.model_class, params)
        self.model = model

        return model

A thing to note is that every custom tuner must exhibit data class behaviour and hence must be decorated with the @dataclass decorator. The tuner class must also extend the BaseTuner class and must have the model_class attribute which corresponds to a callable of the model class to be tuned.

Since a dataclass require that its default class attributes are immutable, we cannot directly use a dictionary to assign values of an int or float sample space, this is because dictionaries are mutable data structures, so in this case we use the MappingProxyType class to wrap any dictionary used to define a search space, so as to exhibit immutable characteristics, which can then be used to initialise an integer or float search space.

The sample_model(...) method of your custom tuner must first call the super().sample_model(...) method from the BaseTuner class. Your custom sample_model(...) method collects all the sampled parameters via your custom sample_params(...) method and passes them as argument, alongside the task ("regression" or "classification") and the model_class into the super().evaluate_sampled_model(...) method. The super().evaluate_sampled_model(...) method first verifies that the keys of the params dictionary returned by the sample_params(...) method all correspond to argument names used to initialised the model_class object, it then initialises an object of the model_class and fits a small set of data with it to check for bad parameter combinations that have been sampled. If the combination of sampled parameters is bad / errotic, it triggers a pruning exception to prune the trial and moves on to the next trial.

You can set the defined custom tuner as only tuner in the search space like so:

tuner = MetaTune(task="classification", single_tuner=CustomGPCTuner())

Or

tuner = MetaTune(task="classification", custom_tuners=[CustomGPCTuner()], custom_only=True)

You can also decide to add it to the already existing search space as a new tuner like so:

tuner = MetaTune(task="classification", custom_tuners=[CustomGPCTuner()], custom_only=False)