niacin.augment package

Submodules

niacin.augment.randaugment module

Implementation of RandAugment algorithm

class niacin.augment.randaugment.RandAugment(transforms: List[Callable], m: int = 10, n: int = 1, shuffle: bool = True, seed: int = None)

Bases: object

Implements RandAugment algorithm (randaugment) as an iterator. RandAugment selects n functions at random from a sequence, and initializes them with magnitude m.

To use it, initialize this class with a list of transformation functions. Every time it is iterated over (e.g. in a for loop) it yields a random subset of those transformation functions.

The original paper claims that m is on a scale from 0-10, but its listed experiments regularly use magnitudes in the 20s and 30s. We take this to mean that “10” was a typo, and the scale was meant to extend to 100.

The defaults for m and n have been set to 1 and 10, respectively, for safety. Depending on your model size and task, you may achieve more accurate results with n ∈ {2, 3} and an m in the range [10, 20). By default, the transforms in a single sample will be returned in random order. If your transforms must occur in a logical sequence (e.g. swapping synonyms before removing random characters), set shuffle to False.

Parameters:

• transforms – sequence of transformation functions
• m – magnitude of transformation, on a scale of 0-100
• n – number of transforms to select
• shuffle – return transforms in random order
• seed – seed to use for the random number generator

m

Magnitude of transformation - should be a number between 0 and 100

n

Size of sample - should be less than total available transforms

Module contents

Augmentor classes. Currently includes:

• RandAugment