Note
Click here to download the full example code
Custom Optuna Optimizer#
Warning
This example shows more advanced features of tpcp
when using
Optuna for hyperparameter optimization.
To make this example understandable, you should make yourself familiar with Optuna first and understand
how it works, before trying to go through this example.
Note
This example uses the dataclass
version of CustomOptunaOptimize
.
To learn more about dataclass interfaces, checkout this example: Dataclass and Attrs support.
When working with dataclasses, be aware that the order of your parameters when inheriting from an other
dataclass can not be controlled.
Therefore, we heavily recommend passing the parameters as keyword arguments.
The most popular method of (hyper-)parameter optimization is GridSearch (or GridSearchCV for optimizable pipelines). These methods perform an exhaustive search of the parameter space by simply testing every option. Considering that training and testing an algorithm can be very costly, exhaustive gridsearch takes a long time and is sometimes not feasible at all due to the required computational load.
For these cases various alternatives exist like RandomizedSearchCV
,
HalvingGridSearchCV
, or advanced blackbox optimizer like
TPESampler
.
tpcp
does not implement all of these methods explicitly, as it would simply be too much work.
However, we try to make it relatively simple to bring such methods into the tpcp
ecosystem by providing an interface
for Optuna.
Optuna is a state-of-the-art hyperparameter optimization framework that allows to implement any of the methods
mentioned above (and more), and allows to easily create custom samplers and pruners.
However, Optuna uses a (very elegant) functional interface that does not play well with the sklearn-inspired
interface of tpcp
.
Therefore, we provide the CustomOptunaOptimize
class which you can subclass to create
your own Optuna based optimizer.
Note
There is no need to create a custom subclass if you only want to run the hyperparameter optimization and
not nest the optimization into other tpcp
methods like cross_validate
.
For these cases, you can simply use Optuna with its default interface and just call the respective tpcp
methods in the objective function.
In this example, we are going to create an optimized gridsearch using custom pruning that terminates trials early if we already realise at the first couple of participants that the parameter combination will not work well.
Keep in mind that this example should merely demonstrate the possibility to integrate Optuna with tpcp
.
You are very much encouraged to read through the
Optuna documentation and create your
own project-specific optimizers.
Note
As some usecases are pretty common, we also provide explicit versions of Optuna optimize subclasses that can be used without implementing your own subclass. Check out the example about built-in optuna optimizers for more information.
The Prerequisites#
First, we need a dataset and a pipeline we want to optimize.
For this example we are using the QRSDetector
pipeline (the non-trainable version) and the ECGExampleData
dataset.
Check out the other examples to learn more about them.
We will simply copy the code over and create an instance of both objects to be used later.
Note
We make pretty extensive use of Python’s optional typing features (in particular generics) in this example.
This can be a little overwhelming, and you might not need that in your implementation.
So whenever, you see TpcpClass[SomeClassName]
and you don’t understand what it means, you can safely
ignore it.
But just for your understanding, if you see for example Pipeline[ECGExampleData]
you should mentally
read it as “A pipeline that requires a Dataset
of type ECGExampleData
internally.
Whenever you encounter a variable ending with a T
(e.g. PipelineT
), these are TypeVar types to type
generics.
You should read that as “Some subclass of Pipeline
, but we don’t know which yet”.
from pathlib import Path
from typing import Any, Callable, Optional, Sequence, Tuple, Union
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"
# The dataset
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
# The pipeline
pipe = MyPipeline()
What We Want To Do#
In the GridSearch Example, we already performed a gridsearch using the
tpcp-GridSearch
class.
Here, we want to do something similar, but improve the gridsearch in two key aspects:
Instead of doing an exhaustive gridsearch, we use one of Optuna’s advanced samplers
When we encounter a parameter combination that doesn’t work, we want to stop testing as early as possible to not waste any time on bad parameter combinations.
We will start by implementing the first aspect and will then make some modifications to enable the second.
The first thing we need for any gridsearch is a score function that tells us how good our parameter combination works. In the GridSearch Example we used a scorer that returns accuracy, precision and f1-score. We will use basically the same function here, but only return the f1-score, as this is the parameter we want to optimize. We could still calculate and return multiple other scores, but this would complicate the implementation of our Optimizer and hence, is kept as exercise for the reader ;) .
from examples.algorithms.algorithms_qrs_detection_final import match_events_with_reference, precision_recall_f1_score
def f1_score(pipeline: MyPipeline, datapoint: ECGExampleData) -> float:
# 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 GridSearch will already clone the pipeline internally and `run`
# will clone it again.
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,
)
*_, f1_score_ = precision_recall_f1_score(matches)
return f1_score_
The Custom Optimizer#
Optimizers in tpcp
are nothing magical – they are simply pipelines that take a different pipeline as input parameter
and have an action method called optimize
that takes in a dataset and then optimizes some parameters of the passed
pipeline using this data.
The CustomOptunaOptimize
class already implements most of that for us and simply
requires us to implement our objective function for the optimization like we would need for Optuna anyway.
Here, we define the objective function within the create_objective
method of our custom optimizer and return it.
The objective we define here is slightly different from the pure Optuna objective function, as it also takes a
Pipeline
and a Dataset
as input in addition to the trial-object.
The content of our objective function is very similar to our score function, but we do not expect just a single datapoint, but an entire dataset. Also, we need to handle getting and applying our parameters within the objective function.
Because we define the function nested within another method, we have access to all class parameters. Hence, if we want to add certain configurations to our objective, we can add parameters to the Optimizer itself and then access it in the objective function.
For the Optimizer we want to build we primarily need two custom pieces of configuration:
The score function we want to use. We want to make that configurable and not hard-code “f1-score” into our optimizer.
The search space for the parameter search. In Optuna the search space is defined by calls to methods on a
optuna.trial.Trial
object. Therefore, we take in a callable that gets the trial object passed and returns the selected parameters. You will see how this works later on.
With these two pieces of configuration in place our objective needs to simply do four things:
First, we need to call the search space function to get the parameters.
Then, apply these parameters to our pipeline.
Afterwards, we need to calculate how good the pipeline with the new parameters works for each of the datapoints within our test dataset.
Finally, we return the aggregated score.
To avoid writing our own for-loop (for now) for the third step, we use Scorer
with our
custom score function.
Scorer
handles looping and aggregating results over multiple datapoints.
With that, our implementation looks as follows:
from dataclasses import dataclass
from optuna import Trial
from tpcp.optimize.optuna import CustomOptunaOptimize
from tpcp.types import DatasetT, PipelineT
from tpcp.validate import Scorer
@dataclass(repr=False)
class OptunaSearch(CustomOptunaOptimize.as_dataclass()[PipelineT, DatasetT]):
# We need to provide default values in Python <3.10, as we can not use the keyword-only syntax for dataclasses.
create_search_space: Optional[Callable[[Trial], None]] = None
score_function: Optional[Callable[[PipelineT, DatasetT], float]] = None
def create_objective(self) -> Callable[[Trial, PipelineT, DatasetT], Union[float, Sequence[float]]]:
# Here we define our objective function
def objective(trial: Trial, pipeline: PipelineT, dataset: DatasetT) -> float:
# First we need to select parameters for the current trial
if self.create_search_space is None:
raise ValueError("No valid search space parameter.")
self.create_search_space(trial)
# Then we apply these parameters to the pipeline
pipeline = pipeline.set_params(**self.sanitize_params(trial.params))
# We wrap the score function with a scorer to avoid writing our own for-loop to aggregate the results.
if self.score_function is None:
raise ValueError("No valid score function.")
scorer = Scorer(self.score_function)
# In the end, we calculate the results per datapoint.
# Note that we could expose the `error_score` parameter on an optimizer level.
# But let's keep it simple for now.
average_score, single_scores = scorer(pipeline, dataset)
# As a bonus, we use the custom params option of optuna to store the individual scores per datapoint and the
# respective data labels
trial.set_user_attr("single_scores", single_scores)
trial.set_user_attr("data_labels", dataset.group_labels)
return average_score
return objective
Note
This implementation is nearly identical to the OptunaSearch
class.
If you really just need a Grid Search equivalent with optuna as backend, you should use this class.
Otherwise, the custom class build in this example is a good starting point for further experimentation.
Running the optimization#
To run the optimization, we need to create a new Optuna study, a custom sampler and the function that defines our search space:
Instead of creating the study directly, we create a function that returns the parameters we want to pass to
create_study
.
This way, the OptunaSearch class can control the study creation and can ensure that we have independent studies
for each run.
This method gets a random seed as input.
We use that to control the random sampler of Optuna.
This way, we ensure that a new random seed is used for each process, in case we use multiprocessing.
If we passed a fixed seed to the sampler, we would get the same results for each process.
We use a simple in-memory study with the direction “maximize”, as we want to optimize for the highest f1-score However, we wrap it by a callable to ensure that we get a new and independent study everytime our Optuna optimizer is called.
from optuna import samplers
def get_study_params(seed):
# We use a simple RandomSampler, but every optuna sampler will work
sampler = samplers.RandomSampler(seed=seed)
return dict(sampler=sampler, direction="maximize")
The search space function requires a little more explanation:
In Optuna, we can use the suggest_...
methods on a trial to get a new value within a given range.
This uses our sampler in the background to suggest a new value that makes sense based on the trials that are already
completed.
The selected parameters are stored in the trial object so that we can access them after the function was called.
We use the names of the parameters we want to modify in our pipeline (using the __
for nested values).
This makes applying the parameters to the pipeline later on easy.
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)
Finally, we are ready to run the pipeline.
We create a new instance and set the stopping criteria (in this case 10 random trials).
Then we can use the familiar Optimize
interface to run everything.
opti = OptunaSearch(
pipe,
get_study_params,
create_search_space=create_search_space,
score_function=f1_score,
n_trials=10,
random_seed=42,
)
opti = opti.optimize(example_data)
print(
f"The best performance was achieved with the parameters {opti.best_params_} and an f1-score of {opti.best_score_}."
)
The best performance was achieved with the parameters {'algorithm__min_r_peak_height_over_baseline': 0.4, 'algorithm__high_pass_filter_cutoff_hz': 0.4} and an f1-score of 0.858757056619628.
We can use opti.search_results_
to get a full overview over all results.
These parameters and parameter names are slightly modified compared to the normal Optuna output, to make it similar
to the output of GridSearch
.
pd.DataFrame(opti.search_results_)
If you need even more insides, you can access the study object directly.
<optuna.study.study.Study object at 0x7fb04c458a90>
And like with all Optimizers, we can access the optimized_pipeline_
and call run
directly on the Optimizer.
opti.optimized_pipeline_
MyPipeline(algorithm=QRSDetector(high_pass_filter_cutoff_hz=0.4, max_heart_rate_bpm=200.0, min_r_peak_height_over_baseline=0.4))
out_pipe = opti.run(example_data[0])
out_pipe.r_peak_positions_
0 77
1 370
2 663
3 947
4 1231
...
2268 648978
2269 649232
2270 649485
2271 649734
2272 649991
Length: 2273, dtype: int64
With this we created a simple random search optimizer (or grid search, or whatever sampler we want to use) using
Optuna.
By using CustomOptunaOptimize
we get compatibility with the tpcp
optimizer interface.
This means we could throw this optimizer into cross_validate
and things would just work.
A step further: Custom Pruning#
Simply to demonstrate the power of having full access to all Optuna features, we will implement a custom pruner that stops testing a trial when one datapoint scores below a certain threshold. The idea is that when we iterate through the datapoints and process them one by one, and find one datapoint where the performance is really bad, we already know that this will not be our best choice/a choice we want to use. Hence, there is no need to compute scores for the remaining datapoints.
In Optuna we can implement this using a custom Pruner and the callback feature of the
Scorer
class.
The Pruner will be called everytime we report a new result and will tell us if we should stop evaluating the trial.
Note
This a unusual usage of pruning. Usually, pruning is used to stop after a certain number of training epochs of an ML classifier and not to stop half-way through evaluating your dataset. But it works and is practical.
To create our custom pruner we need a new class, sub-classing BasePruner
.
Then we implement a prune
method that simply checks if the current intermediate value is below a certain threshold.
If yes we return True, telling optuna that the trial can be pruned.
from optuna.pruners import BasePruner
from optuna.study.study import Study
from optuna.trial import FrozenTrial
class MinDatapointPerformancePruner(BasePruner):
def __init__(self, min_performance: float):
self.min_performance = min_performance
def prune(self, _: Study, trial: FrozenTrial) -> bool:
step = trial.last_step
if step is not None:
score = trial.intermediate_values[step]
if score < self.min_performance:
return True
return False
Afterwards, we need to modify our optimizer to work with the pruner.
We need to report each calculated value from each datapoint to Optuna as soon as it was calculated and not wait
until we ran through the entire dataset.
We can do that by passing a callback function to the Scorer
.
This callback will be called after each datapoint is evaluated and allows us to access the most recent score.
We define this callback within the objective function to have access to the trial object of the outer scope.
Using the trial
object, we can report the most recent score to Optuna using trial.report
.
This will call the pruner and allows us to check afterwards, if the trial should be pruned.
We then write some debug information and end the trial by raising a TrialPruned
exception.
from optuna import TrialPruned
@dataclass(repr=False)
class OptunaSearchEarlyStopping(CustomOptunaOptimize.as_dataclass()[PipelineT, DatasetT]):
# We need to provide default values in Python <3.10, as we can not use the keyword-only syntax for dataclasses.
create_search_space: Optional[Callable[[Trial], None]] = None
score_function: Optional[Callable[[PipelineT, DatasetT], float]] = None
def create_objective(self) -> Callable[[Trial, PipelineT, DatasetT], Union[float, Sequence[float]]]:
def objective(trial: Trial, pipeline: PipelineT, dataset: DatasetT) -> float:
# First, we need to select parameters for the current trial
if self.create_search_space is None:
raise ValueError("No valid search space parameter.")
self.create_search_space(trial)
# Then, we apply these parameters to the pipeline
# Note, we use `get_trial_params` instead of getting the paras directly, as this method will transform
# the literal eval transform, if specified in the params.
pipeline = pipeline.set_params(**self.sanitize_params(trial.params))
def single_score_callback(*, step: int, dataset: DatasetT, scores: Tuple[float, ...], **_: Any):
# We need to report the new score value.
# This will call the pruner internally and then tell us if we should stop
trial.report(float(scores[step]), step)
if trial.should_prune():
# Apparently, our last value was bad, and we should abort.
# However, before we do so, we will save the scores so far as debug information
trial.set_user_attr("single_scores", scores)
trial.set_user_attr("data_labels", dataset[: step + 1].group_labels)
# And, finally, we abort the trial
raise TrialPruned(
f"Pruned at datapoint {step} ({dataset[step].group_labels[0]}) with value " f"{scores[step]}."
)
# We wrap the score function with a Scorer object to avoid writing our own for-loop to aggregate the
# results. We pass our callback and `trial` which is passed as a generic kwarg to scorer and hence can be
# accessed from within our callback.
if self.score_function is None:
raise ValueError("No valid score function.")
scorer = Scorer(self.score_function, single_score_callback=single_score_callback)
# Calculate the results per datapoint.
average_score, single_scores = scorer(pipeline, dataset)
# As a bonus, we use the custom params option of Optuna to store the individual scores per datapoint and the
# respective data labels.
trial.set_user_attr("single_scores", single_scores)
trial.set_user_attr("data_labels", dataset.group_labels)
return average_score
return objective
Running the new Optimizer stays the same (we even reuse the search space). We only need to add an instance of our pruner to the study.
def get_study_params(seed):
sampler = samplers.RandomSampler(seed=seed)
return dict(direction="maximize", sampler=sampler, pruner=MinDatapointPerformancePruner(0.3))
opti_early_stop = OptunaSearchEarlyStopping(
pipe,
get_study_params,
create_search_space=create_search_space,
score_function=f1_score,
n_trials=10,
random_seed=42,
)
opti_early_stop.optimize(example_data)
OptunaSearchEarlyStopping(callbacks=None, create_search_space=<function create_search_space at 0x7fb06daf2700>, eval_str_paras=(), gc_after_trial=False, get_study_params=<function get_study_params at 0x7fb04c4f51f0>, n_jobs=1, n_trials=10, pipeline=MyPipeline(algorithm=QRSDetector(high_pass_filter_cutoff_hz=1, max_heart_rate_bpm=200.0, min_r_peak_height_over_baseline=1.0)), random_seed=42, return_optimized=True, score_function=<function f1_score at 0x7fb064f47ee0>, show_progress_bar=False, timeout=None)
And then we can inspect the output.
Compared to our previous run, we can see that many trials report NaN as score and “PRUNED” in the state
column.
For each of these values we saved some time.
For the other trials, we get the same results as earlier.
pd.DataFrame(opti_early_stop.search_results_)
Summary#
The tpcp <-> Optuna interface is a little bit more low-level than many other tpcp features. Therefore, here is a short summary of the steps you need:
Create a custom optimizer than inherits from
CustomOptunaOptimize
Overwrite the
create_objective
method so that it returns a Callable.The returned callable should expect a
Trial
, aPipeline
, and aDataset
object as input. Otherwise, it is identical to the objective function you would write in “plain” Optuna, and hence, should only return a single cost value for the optimization.If your objective function requires parameter, add them as class attributes via the init.
(optional) If you want to report additional values from your optimization, you can do that via the
set_user_attr
parameter of theTrial
object.(optional) Early stopping and other Pruners can be implemented identical to Optuna. Using the callback option of
Scorer
you can even hook into the datapoint iteration to trigger early stopping during the iteration over the dataset.
Next steps#
Building a custom optimizer is a little more involved than just using GridSearch
. However, it allows great
flexibility with relatively small overhead compared to a pure implementation in Optuna.
In this example we created an objective function that only makes sense for pipelines that don’t have an internal
optimization.
However, instead of just a simple search, you could also create a cross-validation-based search by using
cross_validate
within your objective to split the passed data into multiple train-test sets
and optimize hyperparameters similar to GridSearchCV
.
Total running time of the script: ( 0 minutes 18.181 seconds)
Estimated memory usage: 13 MB