libb.downcast

downcast(df, rtol=1e-05, atol=1e-08, numpy_dtypes_only=False)[source]

Downcast DataFrame to minimum viable type for each column.

Parameters:

  • df (DataFrame) – DataFrame to downcast.

  • rtol (float, default 1e-05) – Relative tolerance for numeric comparison.

  • atol (float, default 1e-08) – Absolute tolerance for numeric comparison.

  • numpy_dtypes_only (bool, default False) – Restrict the result to numpy dtypes, skipping pandas extension dtypes.

Returns:

Each column narrowed to the smallest type whose values stay within tolerance of the original.

Return type:

DataFrame

Notes

  • Tolerance follows the numpy.allclose convention: a candidate type is accepted when |a - b| <= atol + rtol * |b| for every value.

  • Tightening the tolerance widens the result, since fewer narrow types can represent the values: at rtol=atol=1e-10 the float column stays float64, and only at the looser default does it fit float32.

Examples

>>> from numpy import linspace, random
>>> from pandas import DataFrame
>>> data = {
... "integers": linspace(1, 100, 100),
... "floats": linspace(1, 1000, 100).round(2),
... "booleans": random.choice([1, 0], 100),
... "categories": random.choice(["foo", "bar", "baz"], 100)}
>>> df = DataFrame(data)
>>> tight = downcast(df, rtol=1e-10, atol=1e-10)
>>> tight.dtypes['integers'], tight.dtypes['floats']
(dtype('uint8'), dtype('float64'))
>>> loose = downcast(df, rtol=1e-05, atol=1e-08)
>>> loose.dtypes['integers'], loose.dtypes['floats']
(dtype('uint8'), dtype('float32'))