Creating your own split logic

If the previous built-in options are not what you are looking for, there is also the option of implementing your own!

For this, ZenML provides the BaseSplit interface that you can subclass in a standard object-oriented manner to define your own custom split logic.

from zenml.core.steps.split.base_split_step import BaseSplit

class MyCustomSplit(BaseSplit):

# your custom split logic follows

There are two main abstract methods that you have to implement to be able to use your custom split with ZenML’s own split component, partition_fn and get_split_names. The former returns your custom partition function along with its keyword arguments for use in ZenML’s split component. To be eligible in use in a Split Step, the partition function needs to adhere to the following design contract:

  1. The signature is of the following type:

def my_partition(element, n, **kwargs) -> int,

where n is the number of splits.

  1. The partition_fn only returns signed integers i less than n, i.e. 0 ≤ i ≤ n - 1.

Then, the class method partition_fn returns the tuple (my_partition, kwargs) consisting of your custom partition function and its keyword arguments.

The second method get_split_names needs to return a list of data split names used in your ZenML pipeline. These can be different depending on your target workload.

A quick example

Now that the theoretical flow is in place, we can quickly give an example by building a partition function that splits data into train and eval sets based on whether an integer feature in the data is odd or even.

from zenml.core.steps.split.base_split_step import BaseSplit
from zenml.core.steps.split.utils import get_categorical_value

def OddEvenPartitionFn(element, num_partitions, int_feature):

    # get integer value of int_feature from the element
    int_value = get_categorical_value(element, int_feature)
    return int_value % 2

class OddEvenSplit(BaseSplit):
    def __init__(self, 
                 int_feature: Text,
        self.int_feature = int_feature
    def partition_fn(self):
        return OddEvenPartitionFn, {"int_feature": self.int_feature}
    def get_split_names(self):
        return ["train", "eval"]

Note that we could not have done the same logic with a categorical split, since the split above does not assume any knowledge about which integers are present beforehand. For a categorical split to work as built into ZenML, prior knowledge about the categorical domain of the feature has to be present.