Note
Click here to download the full example code
Build-in Optuna Optimizers#
The custom optuna example shows how to implement a specific optuna optimizer with full control over all aspects. This is still the recommended way to do things, as you often will have specific requirements for your objective function.
However, there are still a number of problems that can be solved by a relative generic GridSearch or GridSearchCV. Therefore, we provide Optuna equivalents for these usecases to make use of the advanced samplers optuna provides.
Note
We still recommend to read through the custom optuna example before using the specific implementations demonstrated here.
OptunaSearch - GridSearch on Steroids#
The OptunaSearch
class can be used in all cases where you would use GridSearch
.
The following is equivalent to the GridSearch example (Grid Search optimal Algorithm Parameter).
from pathlib import Path
import pandas as pd
from examples.algorithms.algorithms_qrs_detection_final import QRSDetector
from examples.datasets.datasets_final_ecg import ECGExampleData
from tpcp import Parameter, Pipeline, cf
try:
HERE = Path(__file__).parent
except NameError:
HERE = Path(".").resolve()
data_path = HERE.parent.parent / "example_data/ecg_mit_bih_arrhythmia/data"
example_data = ECGExampleData(data_path)
class MyPipeline(Pipeline[ECGExampleData]):
algorithm: Parameter[QRSDetector]
r_peak_positions_: pd.Series
def __init__(self, algorithm: QRSDetector = cf(QRSDetector())):
self.algorithm = algorithm
def run(self, datapoint: ECGExampleData):
# Note: We need to clone the algorithm instance, to make sure we don't leak any data between runs.
algo = self.algorithm.clone()
algo.detect(datapoint.data["ecg"], datapoint.sampling_rate_hz)
self.r_peak_positions_ = algo.r_peak_positions_
return self
pipe = MyPipeline()
Optuna Study#
To use optuna we need to create an optuna study, or rather a function that returns one, that can be used by
OptunaSearch
to create it.
We will set this up identical to the custom optuna example.
Note
We use a in-memory study here, if you want to use multiprocessing or ensure that your search can be continued, use a different study backend.
from optuna import Trial, create_study, samplers
def get_study():
# We use a simple RandomSampler, but every optuna sampler will work
sampler = samplers.RandomSampler(seed=42)
return create_study(direction="maximize", sampler=sampler)
Search Space#
In contrast to GridSearch
where we define a fix parameter grid, in optuna we define a search space.
Which value sin this search space will actually be evaluated depends on the chosen sampler.
This also needs to be a function that takes the current trial object as input.
def create_search_space(trial: Trial):
trial.suggest_float("algorithm__min_r_peak_height_over_baseline", 0.1, 2, step=0.1)
trial.suggest_float("algorithm__high_pass_filter_cutoff_hz", 0.1, 2, step=0.1)
Score#
We use the same scoring function as in the GridSearch
example:
from examples.algorithms.algorithms_qrs_detection_final import match_events_with_reference, precision_recall_f1_score
def score(pipeline: MyPipeline, datapoint: ECGExampleData):
# We use the `safe_run` wrapper instead of just run. This is always a good idea.
# We don't need to clone the pipeline here, as OptunaSearch will already clone the pipeline internally.
pipeline = pipeline.safe_run(datapoint)
tolerance_s = 0.02 # We just use 20 ms for this example
matches = match_events_with_reference(
pipeline.r_peak_positions_.to_numpy(),
datapoint.r_peak_positions_.to_numpy(),
tolerance=tolerance_s * datapoint.sampling_rate_hz,
)
precision, recall, f1_score = precision_recall_f1_score(matches)
return {"precision": precision, "recall": recall, "f1_score": f1_score}
Running the search#
Now we can run the search.
Note, that because our scoring function returns a dictionary, we need to specify the key we want to optimize by
passing it to score_name
.
In this case, we want to maximize the f1 score.
from tpcp.optimize.optuna import OptunaSearch
opti = OptunaSearch(pipe, get_study, create_search_space, scoring=score, n_trials=10, score_name="f1_score")
opti = opti.optimize(example_data)
Inspecting the results#
The results are very similar to the output of GridSearch
.
Besides the main results, we provide the results for each single datapoint and the respective grouplabel for the
datapoints.
results = pd.DataFrame(opti.search_results_)
results
We can also get the best para combi and an instance of the pipeline initialized with the best parameter combination.
print("Best Para Combi:", opti.best_params_)
print("Best score:", opti.best_score_)
print("Paras of optimized Pipeline:", opti.optimized_pipeline_.get_params())
Best Para Combi: {'algorithm__min_r_peak_height_over_baseline': 0.4, 'algorithm__high_pass_filter_cutoff_hz': 0.4}
Best score: 0.858757056619628
Paras of optimized Pipeline: {'algorithm__high_pass_filter_cutoff_hz': 0.4, 'algorithm__max_heart_rate_bpm': 200.0, 'algorithm__min_r_peak_height_over_baseline': 0.4, 'algorithm': QRSDetector(high_pass_filter_cutoff_hz=0.4, max_heart_rate_bpm=200.0, min_r_peak_height_over_baseline=0.4)}
Total running time of the script: ( 0 minutes 9.315 seconds)
Estimated memory usage: 8 MB