Thanks for tuning in to Google I/O. View all sessions on demandWatch on demand

tfds.even_splits

Generates a list of non-overlapping sub-splits of same size.

tfds.even_splits(
    split: str, n: int, *, drop_remainder: bool = False
) -> List[tfds.typing.SplitArg]

Example:

split0, split1, split2 = tfds.even_splits('train', n=3, drop_remainder=True)

# Load 1/3 of the train split.
ds = tfds.load('my_dataset', split=split0)

tfds.even_splits supports arbitrary sub-splits inputs, including other tfds.even_splits outputs.

split Split (e.g. 'train', 'test[75%:]',...)
n Number of sub-splits to create
drop_remainder Drop examples if the number of examples in the datasets is not evenly divisible by n. If False, examples are distributed evenly across subsplits, starting by the first. For example, if there is 11 examples with n=3, splits will contain [4, 4, 3] examples respectivelly.

The list of subsplits. Those splits can be combined together (with +) or with other subsplits (e.g. split + 'test[75%:]').