flexcv.repeated

flexcv.repeated.RepeatedCV

Bases: CrossValidation

Class for repeated cross-validation. Inherits from CrossValidation.

Attributes:

Name	Type	Description
n_repeats	int	Number of repeated runs. (Default value = 3)
seeds	list[int]	List of seeds for the repeated runs. (Default value = None)
generator_seed	int	Seed to control generation of a list of seeds. (Default value = 42)
parent_run	run	A run to track the repeated meta data. (Default value = None)
children_runs	list[run]	A list of runs to track the single runs. (Default value = None)

Methods:

Name	Description
set_n_repeats	method that sets the number of repeated runs
set_seeds	method that sets the random seeds for the repeated runs
set_neptune	method that sets the neptune credentials for the repeated runs and initializes them
set_run	in contrast to CrossValidation, the set_run method takes in a parent run and a list of children runs
perform	method that performs repeated cross-validation

Example

from flexcv.synthesizer import generate_regression
from flexcv.models import LinearModel
from flexcv.model_mapping import ModelConfigDict, ModelMappingDict
from flexcv.repeated import RepeatedCV

# make sample data
X, y, group, random_slopes = generate_regression(10, 100, n_slopes=1, noise_level=9.1e-2, random_seed=42), random_seed=42

# create a model mapping
model_map = ModelMappingDict(
    {
        "LinearModel": ModelConfigDict(
            {
                "model": LinearModel,
                "requires_formula": True,
            }
        ),
    }
)

credentials = {your_neptune_credentials}

rcv = (
        RepeatedCV()
        .set_data(X, y, group, dataset_name="ExampleData")
        .set_models(model_map)
        .set_n_repeats(3)
        .set_neptune(credentials)
        .perform()
        .get_results()
)

rcv.summary.to_excel("repeated_cv.xlsx")  # save dataframe to excel file

Source code in flexcv/repeated.py

class RepeatedCV(CrossValidation):
    """Class for repeated cross-validation. Inherits from CrossValidation.

    Attributes:
        n_repeats (int): Number of repeated runs. (Default value = 3)
        seeds (list[int]): List of seeds for the repeated runs. (Default value = None)
        generator_seed (int): Seed to control generation of a list of seeds. (Default value = 42)
        parent_run (neptune.run): A run to track the repeated meta data. (Default value = None)
        children_runs list[neptune.run]: A list of runs to track the single runs. (Default value = None)

    Methods:
        set_n_repeats: method that sets the number of repeated runs
        set_seeds: method that sets the random seeds for the repeated runs
        set_neptune: method that sets the neptune credentials for the repeated runs and initializes them
        set_run: in contrast to CrossValidation, the set_run method takes in a parent run and a list of children runs
        perform: method that performs repeated cross-validation

    Example:
        ```python
        from flexcv.synthesizer import generate_regression
        from flexcv.models import LinearModel
        from flexcv.model_mapping import ModelConfigDict, ModelMappingDict
        from flexcv.repeated import RepeatedCV

        # make sample data
        X, y, group, random_slopes = generate_regression(10, 100, n_slopes=1, noise_level=9.1e-2, random_seed=42), random_seed=42

        # create a model mapping
        model_map = ModelMappingDict(
            {
                "LinearModel": ModelConfigDict(
                    {
                        "model": LinearModel,
                        "requires_formula": True,
                    }
                ),
            }
        )

        credentials = {your_neptune_credentials}

        rcv = (
                RepeatedCV()
                .set_data(X, y, group, dataset_name="ExampleData")
                .set_models(model_map)
                .set_n_repeats(3)
                .set_neptune(credentials)
                .perform()
                .get_results()
        )

        rcv.summary.to_excel("repeated_cv.xlsx")  # save dataframe to excel file
        ```
    """

    def __init__(self, **kwargs):
        super().__init__(**kwargs)
        self.n_repeats = 3
        self.seeds = None
        self.generator_seed = 42

    def set_n_repeats(self, n_repeats: int):
        """Set the number of repeats for repeated cross-validation.

        Args:
            n_repeats(int): Number of repeated runs.

        Returns:
            (RepeatedCV): self

        """
        self.n_repeats = n_repeats
        return self

    def set_neptune(self, neptune_credentials):
        """Set your neptune credentials and initialize the repeated runs automatically.
        If you do not want to pass your credentials explicitly, use set_run and call init_repeated_runs from outside the class yourself.

        Args:
            neptune_credentials (dict): Your API token and project name.

        Returns:
            (RepeatedCV): self

        """
        self.neptune_credentials = neptune_credentials
        parent_run, children_runs = init_repeated_runs(
            self.n_repeats, neptune_credentials
        )
        self.parent_run = parent_run
        self.children_runs = children_runs
        return self

    def set_seeds(self, seeds=None, generator_seed=42):
        """Set the seeds for the repeated runs.
        If no seeds are passed, a list of seeds is generated randomly.
        The random seed for the random seed generation can be set via the generator_seed argument.
        Therefore, the random seed for the repeated runs is set to the same value for each repeated run and is deterministic.

        Args:
            seeds (list[int]): The list of seeds to use in the repeats. (Default value = None)
            generator_seed (int): Seed to control generation of a list of seeds. (Default value = 42)

        Returns:
            (RepeatedCV): self
        """
        self.generator_seed = generator_seed
        if seeds is None:
            np.random.seed(generator_seed)
            self.seeds = np.random.randint(42000, size=self.n_repeats).tolist()
            return self

        self.seeds = seeds
        return self

    def set_run(self, parent_run=None, children_run=None):
        """Use this method if you want to pass your own parent run and children runs.

        Args:
            parent_run (neptune.run): A run to track the repeated meta data. (Default value = None)
            children_run list[neptune.run]: A list of runs to track the single runs. (Default value = None)

        Returns:
            (RepeatedCV): self

        """
        self.parent_run = parent_run
        self.children_run = children_run
        return self

    def _perform_repeats(self):
        """Performs repeated cross-validation.

        Returns:
            (pd.DataFrame): DataFrame with aggregated metrics.

        """
        add_module_handlers(logger)

        if hasattr(self, "parent_run") and self.parent_run is not None:
            repeated_run = self.parent_run
        else:
            logger.info("No parent run found. Initializing dummy run.")
            repeated_run = Run()
            self.parent_run = repeated_run

        if not hasattr(self, "children_runs") or self.children_runs is None:
            logger.info("No children runs found. Initializing dummy runs.")
            self.children_runs = [Run() for _ in range(self.n_repeats)]

        repeated_id = repeated_run["sys/id"].fetch()

        if self.seeds is None:
            logger.info(
                f"No seeds found. Initializing seeds with genearator seed {self.generator_seed}."
            )
            self.set_seeds()

        run_ids = []
        run_results = []
        for seed, inner_run in zip(self.seeds, self.children_runs):
            # handle neptune
            inner_id = inner_run["sys/id"].fetch()

            # if inner_id is not a string, we are not using neptune and need to generate a run id
            if not isinstance(inner_id, str):
                i = -1
                inner_id = f"run_{i+1}"

            inner_run["HostRun"] = repeated_id
            inner_run["seed"] = seed

            # instantiate a new CrossValidation instance for each run
            cv_in = CrossValidation()

            # assign all necessary attributes of self to the inner cv instance
            for key, value in self.config.items():
                cv_in.config[key] = value

            # set inner cv config
            cv_in.config["random_seed"] = seed
            cv_in.config["run"] = inner_run

            # perform single run and get the results
            results = cv_in.perform().get_results()

            # append the run id and the run metric to the lists
            run_ids.append(inner_id)
            run_results.append(results)
            inner_run.stop()
            plt.close()
        # run_dfs have the same column and index names and we
        df = aggregate_(run_results)

        # log the repeated run results to neptune
        repeated_run["summary"].upload(File.as_html(df))
        repeated_run[
            "sys/description"
        ] = f"Host run for repeated runs with {self.n_repeats} repeats. run_ids: {run_ids}"
        repeated_run["RelatedRuns"] = ", ".join(run_ids)
        repeated_run["seeds"] = self.seeds
        repeated_run["mapping"] = self.config["mapping"]
        repeated_run.stop()
        return df

    def perform(self):
        """Wrapper method to perform repeated cross-validation. Overwrites the perform method of CrossValidation.

        Returns:
            (RepeatedCV): self

        """
        if self.seeds is None:
            self.set_seeds()
        summary_df = self._perform_repeats()
        self._summary_df = summary_df
        return self

    def get_results(self):
        """Returns the results of repeated cross-validation.

        Returns:
            (RepeatedResult): RepeatedResult object with summary property that returns a DataFrame with aggregated metrics.
        """
        return RepeatedResult(self._summary_df)

flexcv.repeated.RepeatedCV.get_results()

Returns the results of repeated cross-validation.

Returns:

Type	Description
RepeatedResult	RepeatedResult object with summary property that returns a DataFrame with aggregated metrics.

Source code in flexcv/repeated.py

def get_results(self):
    """Returns the results of repeated cross-validation.

    Returns:
        (RepeatedResult): RepeatedResult object with summary property that returns a DataFrame with aggregated metrics.
    """
    return RepeatedResult(self._summary_df)

flexcv.repeated.RepeatedCV.perform()

Wrapper method to perform repeated cross-validation. Overwrites the perform method of CrossValidation.

Returns:

Type	Description
RepeatedCV	self

Source code in flexcv/repeated.py

def perform(self):
    """Wrapper method to perform repeated cross-validation. Overwrites the perform method of CrossValidation.

    Returns:
        (RepeatedCV): self

    """
    if self.seeds is None:
        self.set_seeds()
    summary_df = self._perform_repeats()
    self._summary_df = summary_df
    return self

flexcv.repeated.RepeatedCV.set_n_repeats(n_repeats)

Set the number of repeats for repeated cross-validation.

Parameters:

Name	Type	Description	Default
n_repeats(int)		Number of repeated runs.	required

Returns:

Type	Description
RepeatedCV	self

Source code in flexcv/repeated.py

def set_n_repeats(self, n_repeats: int):
    """Set the number of repeats for repeated cross-validation.

    Args:
        n_repeats(int): Number of repeated runs.

    Returns:
        (RepeatedCV): self

    """
    self.n_repeats = n_repeats
    return self

flexcv.repeated.RepeatedCV.set_neptune(neptune_credentials)

Set your neptune credentials and initialize the repeated runs automatically. If you do not want to pass your credentials explicitly, use set_run and call init_repeated_runs from outside the class yourself.

Parameters:

Name	Type	Description	Default
neptune_credentials	dict	Your API token and project name.	required

Returns:

Type	Description
RepeatedCV	self

Source code in flexcv/repeated.py

def set_neptune(self, neptune_credentials):
    """Set your neptune credentials and initialize the repeated runs automatically.
    If you do not want to pass your credentials explicitly, use set_run and call init_repeated_runs from outside the class yourself.

    Args:
        neptune_credentials (dict): Your API token and project name.

    Returns:
        (RepeatedCV): self

    """
    self.neptune_credentials = neptune_credentials
    parent_run, children_runs = init_repeated_runs(
        self.n_repeats, neptune_credentials
    )
    self.parent_run = parent_run
    self.children_runs = children_runs
    return self

flexcv.repeated.RepeatedCV.set_run(parent_run=None, children_run=None)

Use this method if you want to pass your own parent run and children runs.

Parameters:

Name	Type	Description	Default
parent_run	run	A run to track the repeated meta data. (Default value = None)	None
children_run	list[run]	A list of runs to track the single runs. (Default value = None)	None

Returns:

Type	Description
RepeatedCV	self

Source code in flexcv/repeated.py

def set_run(self, parent_run=None, children_run=None):
    """Use this method if you want to pass your own parent run and children runs.

    Args:
        parent_run (neptune.run): A run to track the repeated meta data. (Default value = None)
        children_run list[neptune.run]: A list of runs to track the single runs. (Default value = None)

    Returns:
        (RepeatedCV): self

    """
    self.parent_run = parent_run
    self.children_run = children_run
    return self

flexcv.repeated.RepeatedCV.set_seeds(seeds=None, generator_seed=42)

Set the seeds for the repeated runs. If no seeds are passed, a list of seeds is generated randomly. The random seed for the random seed generation can be set via the generator_seed argument. Therefore, the random seed for the repeated runs is set to the same value for each repeated run and is deterministic.

Parameters:

Name	Type	Description	Default
seeds	list[int]	The list of seeds to use in the repeats. (Default value = None)	None
generator_seed	int	Seed to control generation of a list of seeds. (Default value = 42)	42

Returns:

Type	Description
RepeatedCV	self

Source code in flexcv/repeated.py

def set_seeds(self, seeds=None, generator_seed=42):
    """Set the seeds for the repeated runs.
    If no seeds are passed, a list of seeds is generated randomly.
    The random seed for the random seed generation can be set via the generator_seed argument.
    Therefore, the random seed for the repeated runs is set to the same value for each repeated run and is deterministic.

    Args:
        seeds (list[int]): The list of seeds to use in the repeats. (Default value = None)
        generator_seed (int): Seed to control generation of a list of seeds. (Default value = 42)

    Returns:
        (RepeatedCV): self
    """
    self.generator_seed = generator_seed
    if seeds is None:
        np.random.seed(generator_seed)
        self.seeds = np.random.randint(42000, size=self.n_repeats).tolist()
        return self

    self.seeds = seeds
    return self

flexcv.repeated.RepeatedResult

Class for results of repeated cross-validation. Implements a summary property that returns a DataFrame with aggregated metrics.

Source code in flexcv/repeated.py

class RepeatedResult:
    """Class for results of repeated cross-validation.
    Implements a summary property that returns a DataFrame with aggregated metrics.
    """

    def __init__(self, df):
        """Constructor method for RepeatedResult class.

        Args:
            df(pd.DataFrame): DataFrame with aggregated metrics.
        """
        self._summary_df = df

    @property
    def summary(self):
        """Summary property that returns a DataFrame with aggregated metrics.

        Returns:
            (pd.DataFrame): DataFrame with aggregated metrics.

        """
        return self._summary_df

flexcv.repeated.RepeatedResult.summary property

Summary property that returns a DataFrame with aggregated metrics.

Returns:

Type	Description
DataFrame	DataFrame with aggregated metrics.

flexcv.repeated.RepeatedResult.__init__(df)

Constructor method for RepeatedResult class.

Parameters:

Name	Type	Description	Default
df(pd.DataFrame)		DataFrame with aggregated metrics.	required

Source code in flexcv/repeated.py

def __init__(self, df):
    """Constructor method for RepeatedResult class.

    Args:
        df(pd.DataFrame): DataFrame with aggregated metrics.
    """
    self._summary_df = df

flexcv.repeated.aggregate_(repeated_runs)

Aggregate the results of repeated runs into a single DataFrame. Therefore, the nested dict structure of the results is flattened. First, the model results are averaged over folds of individual runs. Second, the repeated (individual) runs are averaged.

Parameters:

Name	Type	Description	Default
repeated_runs		list: List of results of repeated runs.	required

Returns:

Type	Description
DataFrame	Summary statistics of repeated runs.

Note:

The summary statistics are returned as a DataFrame with the following structure:
- index: [aggregate]_[metric_name]
- columns: [model_name]
- values: [metric_value]

Source code in flexcv/repeated.py

def aggregate_(repeated_runs) -> pd.DataFrame:
    """Aggregate the results of repeated runs into a single DataFrame.
    Therefore, the nested dict structure of the results is flattened.
    First, the model results are averaged over folds of individual runs.
    Second, the repeated (individual) runs are averaged.

    Args:
        repeated_runs: list: List of results of repeated runs.

    Returns:
        (pd.DataFrame): Summary statistics of repeated runs.

    Note:

        The summary statistics are returned as a DataFrame with the following structure:
        - index: [aggregate]_[metric_name]
        - columns: [model_name]
        - values: [metric_value]

    """

    model_keys = list(repeated_runs[0].keys())

    # we want the metrics keys
    # index [0] gives us first run
    # index [model_keys[0]] -> gives us the dict for the first model
    # index ["fold_by_metrics"] -> gives us a dict[metrics: list]
    # .keys() -> gives us the keys of the metrics dict
    result_keys = repeated_runs[0][model_keys[0]]["folds_by_metrics"].keys()
    results = []
    for result_key in result_keys:
        tmp_df = pd.DataFrame(
            [
                pd.Series(
                    [
                        try_mean(run[model_key]["folds_by_metrics"][result_key])
                        for model_key in model_keys
                    ],
                    index=model_keys,
                )
                for run in repeated_runs
            ]
        ).agg(["mean", "std"])
        tmp_df.index = [f"{result_key}_{index}" for index in tmp_df.index]
        results.append(tmp_df)
    return pd.concat(results)

flexcv.repeated.init_repeated_runs(n_repeats, neptune_credentials)

Initialize a repeated run and n_repeats children runs. The children runs are linked to the parent run via the "HostRun" key. The children runs are initialized with the same credentials as the parent run.

Parameters:

Name	Type	Description	Default
n_repeats	int	Number of repeated runs.	required
use_neptune	bool	Whether to use neptune or not.	required
neptune_credentials	dict	Credentials for neptune.	required

Returns:

Type	Description
tuple[Run, list[Run]]	parent_run (Run), children_runs (list of n_repeats runs)

Source code in flexcv/repeated.py

def init_repeated_runs(n_repeats, neptune_credentials) -> tuple:
    """Initialize a repeated run and n_repeats children runs.
    The children runs are linked to the parent run via the "HostRun" key.
    The children runs are initialized with the same credentials as the parent run.

    Args:
        n_repeats (int): Number of repeated runs.
        use_neptune (bool): Whether to use neptune or not.
        neptune_credentials (dict): Credentials for neptune.

    Returns:
        (tuple[Run, list[Run]]): parent_run (Run), children_runs (list of n_repeats runs)

    """
    parent_run = neptune.init_run(**neptune_credentials)
    children_runs = [neptune.init_run(**neptune_credentials) for _ in range(n_repeats)]
    return parent_run, children_runs

flexcv.repeated.try_mean(x)

Mean function that handles NaN values for cases where the input contains string "NaN". If all elements in x are equal to "NaN", -99 is returned. If x contains "NaN" and other values, -999 is returned.

Parameters:

Name	Type	Description	Default
x		(array-like): Input array.	required

Returns:

Type	Description
float	Mean of the input array.

Source code in flexcv/repeated.py

def try_mean(x):
    """Mean function that handles NaN values for cases where the input contains string "NaN".
    If all elements in x are equal to "NaN", -99 is returned.
    If x contains "NaN" and other values, -999 is returned.

    Args:
        x: (array-like): Input array.

    Returns:
        (float): Mean of the input array.

    """
    from numpy.core._exceptions import UFuncTypeError

    try:
        return np.mean(x)
    except (ValueError, UFuncTypeError):
        # entered if x contains str "NaN"
        # check if all elements in x are equal to "NaN"
        if np.all([element == "NaN" for element in x]):
            return -99
        else:
            return -999