Configure Check Conditions#

The following guide includes different options for configuring a check’s condition(s):

Add Condition#

In order to add a condition to an existing check, we can use any of the pre-defined conditions for that check. The naming convention for the methods that add the condition is add_condition_....

If you want to create and add your custom condition logic for parsing the check’s result value, see Add a Custom Condition.

Add a condition to a new check#

from deepchecks.tabular.checks import DatasetsSizeComparison

check = DatasetsSizeComparison().add_condition_test_size_greater_or_equal(1000)
check

Out:

DatasetsSizeComparison
    Conditions:
            0: Test dataset size is greater or equal to 1000

Conditions are used mainly in the context of a Suite, and displayed in the Conditions Summary table. For example how to run in a suite you can look at Add a Custom Condition or if you would like to run the conditions outside of suite you can execute:

import pandas as pd

from deepchecks.tabular import Dataset

# Dummy data
train_dataset = Dataset(pd.DataFrame(data={'x': [1,2,3,4,5,6,7,8,9]}))
test_dataset = Dataset(pd.DataFrame(data={'x': [1,2,3]}))

condition_results = check.conditions_decision(check.run(train_dataset, test_dataset))
condition_results

Out:

[{'details': 'Test dataset contains 3 samples', 'category': <ConditionCategory.FAIL: 'FAIL'>, 'name': 'Test dataset size is greater or equal to 1000'}]

Add a condition to a check in a suite#

If we want to add a conditon to a check within an existing suite, we should first find the Check’s ID within the suite, and then add the condition to it, by running the relevant add_condition_ method on that check’s instance. See the next section to understand how to do so.

The condition will then be appended to the list of conditions on that check (or be the first one if no conditions are defined), and each condition will be evaluated separately when running the suite.

Remove / Edit a Condition#

Deepchecks provides different kinds of default suites, which come with pre-defined conditions. You may want to remove a condition in case it isn’t needed for you, or you may want to change the condition’s parameters (since conditions functions are immutable).

To remove a condition, start by printing the Suite and identifing the Check’s ID, and the Condition’s ID:

from deepchecks.tabular.suites import train_test_validation

suite = train_test_validation()
suite

Out:

Train Test Validation Suite: [
    0: DatasetsSizeComparison
            Conditions:
                    0: Test-Train size ratio is greater than 0.01
    1: NewLabelTrainTest
            Conditions:
                    0: Number of new label values is less or equal to 0
    2: CategoryMismatchTrainTest
            Conditions:
                    0: Ratio of samples with a new category is less or equal to 0%
    3: StringMismatchComparison
            Conditions:
                    0: No new variants allowed in test data
    4: DateTrainTestLeakageDuplicates
            Conditions:
                    0: Date leakage ratio is less or equal to 0%
    5: DateTrainTestLeakageOverlap
            Conditions:
                    0: Date leakage ratio is less or equal to 0%
    6: IndexTrainTestLeakage
            Conditions:
                    0: Ratio of leaking indices is less or equal to 0%
    7: IdentifierLeakage(ppscore_params={})
            Conditions:
                    0: Identifier columns PPS is less or equal to 0
    8: TrainTestSamplesMix
            Conditions:
                    0: Percentage of test data samples that appear in train data is less or equal to 10%
    9: FeatureLabelCorrelationChange(ppscore_params={})
            Conditions:
                    0: Train-Test features' Predictive Power Score difference is less than 0.2
                    1: Train features' Predictive Power Score is less than 0.7
    10: TrainTestFeatureDrift
            Conditions:
                    0: categorical drift score < 0.2 and numerical drift score < 0.1
    11: TrainTestLabelDrift
            Conditions:
                    0: categorical drift score < 0.2 and numerical drift score < 0.1 for label drift
    12: WholeDatasetDrift
            Conditions:
                    0: Drift value is less than 0.25
]

After we found the IDs we can remove the Condition:

# Access check by id
check = suite[9]
# Remove condition by id
check.remove_condition(0)

suite

Out:

Train Test Validation Suite: [
    0: DatasetsSizeComparison
            Conditions:
                    0: Test-Train size ratio is greater than 0.01
    1: NewLabelTrainTest
            Conditions:
                    0: Number of new label values is less or equal to 0
    2: CategoryMismatchTrainTest
            Conditions:
                    0: Ratio of samples with a new category is less or equal to 0%
    3: StringMismatchComparison
            Conditions:
                    0: No new variants allowed in test data
    4: DateTrainTestLeakageDuplicates
            Conditions:
                    0: Date leakage ratio is less or equal to 0%
    5: DateTrainTestLeakageOverlap
            Conditions:
                    0: Date leakage ratio is less or equal to 0%
    6: IndexTrainTestLeakage
            Conditions:
                    0: Ratio of leaking indices is less or equal to 0%
    7: IdentifierLeakage(ppscore_params={})
            Conditions:
                    0: Identifier columns PPS is less or equal to 0
    8: TrainTestSamplesMix
            Conditions:
                    0: Percentage of test data samples that appear in train data is less or equal to 10%
    9: FeatureLabelCorrelationChange(ppscore_params={})
            Conditions:
                    1: Train features' Predictive Power Score is less than 0.7
    10: TrainTestFeatureDrift
            Conditions:
                    0: categorical drift score < 0.2 and numerical drift score < 0.1
    11: TrainTestLabelDrift
            Conditions:
                    0: categorical drift score < 0.2 and numerical drift score < 0.1 for label drift
    12: WholeDatasetDrift
            Conditions:
                    0: Drift value is less than 0.25
]

Now if we want we can also re-add the Condition using the built-in methods on the check, with a different parameter.

# Re-add the condition with new parameter
check.add_condition_feature_pps_difference_less_than(0.01)

suite

Out:

Train Test Validation Suite: [
    0: DatasetsSizeComparison
            Conditions:
                    0: Test-Train size ratio is greater than 0.01
    1: NewLabelTrainTest
            Conditions:
                    0: Number of new label values is less or equal to 0
    2: CategoryMismatchTrainTest
            Conditions:
                    0: Ratio of samples with a new category is less or equal to 0%
    3: StringMismatchComparison
            Conditions:
                    0: No new variants allowed in test data
    4: DateTrainTestLeakageDuplicates
            Conditions:
                    0: Date leakage ratio is less or equal to 0%
    5: DateTrainTestLeakageOverlap
            Conditions:
                    0: Date leakage ratio is less or equal to 0%
    6: IndexTrainTestLeakage
            Conditions:
                    0: Ratio of leaking indices is less or equal to 0%
    7: IdentifierLeakage(ppscore_params={})
            Conditions:
                    0: Identifier columns PPS is less or equal to 0
    8: TrainTestSamplesMix
            Conditions:
                    0: Percentage of test data samples that appear in train data is less or equal to 10%
    9: FeatureLabelCorrelationChange(ppscore_params={})
            Conditions:
                    1: Train features' Predictive Power Score is less than 0.7
                    2: Train-Test features' Predictive Power Score difference is less than 0.01
    10: TrainTestFeatureDrift
            Conditions:
                    0: categorical drift score < 0.2 and numerical drift score < 0.1
    11: TrainTestLabelDrift
            Conditions:
                    0: categorical drift score < 0.2 and numerical drift score < 0.1 for label drift
    12: WholeDatasetDrift
            Conditions:
                    0: Drift value is less than 0.25
]

Add a Custom Condition#

In order to write conditions we first have to know what value a given check produces.

Let’s look at the check DatasetsSizeComparison and see it’s return value in order to write a condition for it.

import pandas as pd

from deepchecks.tabular import Dataset
from deepchecks.tabular.checks import DatasetsSizeComparison

# We'll use dummy data for the purpose of this demonstration
train_dataset = Dataset(pd.DataFrame(data={'x': [1,2,3,4,5,6,7,8,9]}))
test_dataset = Dataset(pd.DataFrame(data={'x': [1,2,3]}))

result = DatasetsSizeComparison().run(train_dataset, test_dataset)
result.value

Out:

{'Train': 9, 'Test': 3}

Now we know what the return value looks like. Let’s add a new condition that validates that the ratio between the train and test datasets size is inside a given range. To create condition we need to use the add_condition method of the check which accepts a condition name and a function. This function receives the value of the CheckResult that we saw above and should return either a boolean or a ConditionResult containing a boolean and optional extra info that will be displayed in the Conditions Summary table.

Note: When implementing a condition in a custom check, you may want to add a method ``add_condition_x()`` to allow any consumer of your check to apply the condition (possibly with given parameters). For examples look at implemented Checks’ source code

from deepchecks.core import ConditionResult

# Our parameters for the condition
low_threshold = 0.4
high_threshold = 0.6

# Create the condition function
def custom_condition(value: dict, low=low_threshold, high=high_threshold):
    ratio = value['Test'] / value['Train']
    if low <= ratio <= high:
        return ConditionResult(ConditionCategory.PASS)
    else:
        # Note: if you doesn't care about the extra info, you can return directly a boolean
        return ConditionResult(ConditionCategory.FAIL, f'Test-Train ratio is {ratio:.2}')

# Create the condition name
condition_name = f'Test-Train ratio is between {low_threshold} to {high_threshold}'

# Create check instance with the condition
check = DatasetsSizeComparison().add_condition(condition_name, custom_condition)

Now we will use a Suite to demonstrate the action of the condition, since the suite runs the condition for us automatically and prints out a Conditions Summary table (for all the conditions defined on the checks within the suite):

from deepchecks.tabular import Suite

# Using suite to run check & condition
suite = Suite('Suite for Condition',
    check
)

suite.run(train_dataset, test_dataset)

Out:

Suite for Condition:
|     | 0/1 [00:00<?, ? Check/s]
Suite for Condition


Set Custom Condition Category#

When writing your own condition logic, you can decide to mark a condition result as either fail or warn, by passing the category to the ConditionResult object. For example we can even write condition which sets the category based on severity of the result:

from deepchecks.core import ConditionCategory, ConditionResult

# Our parameters for the condition
low_threshold = 0.3
high_threshold = 0.7

# Create the condition function for check `DatasetsSizeComparison`
def custom_condition(value: dict):
    ratio = value['Test'] / value['Train']
    if low_threshold <= ratio <= high_threshold:
        return ConditionResult(ConditionCategory.PASS)
    elif ratio < low_threshold:
        return ConditionResult(ConditionCategory.FAIL, f'Test-Train ratio is {ratio:.2}', ConditionCategory.FAIL)
    else:
        return ConditionResult(ConditionCategory.FAIL, f'Test-Train ratio is {ratio:.2}', ConditionCategory.WARN)

Total running time of the script: ( 0 minutes 0.148 seconds)

Gallery generated by Sphinx-Gallery