RandAugment¶

class mmpretrain.datasets.transforms.RandAugment(policies, num_policies, magnitude_level, magnitude_std=0.0, total_level=10, hparams={'pad_val': 128})[源代码]¶

Random augmentation.

This data augmentation is proposed in RandAugment: Practical automated data augmentation with a reduced search space.

参数:

policies (str | list[dict]) –
The policies of random augmentation. If string, use preset policies collection like “timm_increasing”. If list, each item is one specific augmentation policy dict. The policy dict shall should have these keys:
- type (str), The type of augmentation.
- magnitude_range (Sequence[number], optional): For those augmentation have magnitude, you need to specify the magnitude level mapping range. For example, assume total_level is 10, magnitude_level=3 specify magnitude is 3 if magnitude_range=(0, 10) while specify magnitude is 7 if magnitude_range=(10, 0).
- other keyword arguments of the augmentation.
num_policies (int) – Number of policies to select from policies each time.
magnitude_level (int | float) – Magnitude level for all the augmentation selected.
magnitude_std (Number | str) –
Deviation of magnitude noise applied.
- If positive number, the magnitude obeys normal distribution \(\mathcal{N}(magnitude_level, magnitude_std)\).
- If 0 or negative number, magnitude remains unchanged.
- If str “inf”, the magnitude obeys uniform distribution \(Uniform(min, magnitude)\).
total_level (int | float) – Total level for the magnitude. Defaults to 10.
hparams (dict) – Configs of hyperparameters. Hyperparameters will be used in policies that require these arguments if these arguments are not set in policy dicts. Defaults to dict(pad_val=128).

Available preset policies

"timm_increasing": The _RAND_INCREASING_TRANSFORMS policy from timm

示例

To use “timm-increasing” policies collection, select two policies every time, and magnitude_level of every policy is 6 (total is 10 by default)

>>> import numpy as np
>>> from mmpretrain.datasets import RandAugment
>>> transform = RandAugment(
...     policies='timm_increasing',
...     num_policies=2,
...     magnitude_level=6,
... )
>>> data = {'img': np.random.randint(0, 256, (224, 224, 3))}
>>> results = transform(data)
>>> print(results['img'].shape)
(224, 224, 3)

If you want the magnitude_level randomly changes every time, you can use magnitude_std to specify the random distribution. For example, a normal distribution \(\mathcal{N}(6, 0.5)\).

>>> transform = RandAugment(
...     policies='timm_increasing',
...     num_policies=2,
...     magnitude_level=6,
...     magnitude_std=0.5,
... )

You can also use your own policies:

>>> policies = [
...     dict(type='AutoContrast'),
...     dict(type='Rotate', magnitude_range=(0, 30)),
...     dict(type='ColorTransform', magnitude_range=(0, 0.9)),
... ]
>>> transform = RandAugment(
...     policies=policies,
...     num_policies=2,
...     magnitude_level=6
... )

备注

magnitude_std will introduce some randomness to policy, modified by https://github.com/rwightman/pytorch-image-models.

When magnitude_std=0, we calculate the magnitude as follows:

\[\text{magnitude} = \frac{\text{magnitude_level}} {\text{totallevel}} \times (\text{val2} - \text{val1}) + \text{val1}\]

transform(results)[源代码]¶: Randomly choose a sub-policy to apply.