Data Augmentation¶
d3rlpy provides data augmentation techniques tightly integrated with reinforcement learning algorithms.
- Kostrikov et al., Image Augmentation Is All You Need: Regularizing Deep Reinforcement Learning from Pixels.
- Laskin et al., Reinforcement Learning with Augmented Data.
Efficient data augmentation potentially boosts algorithm performance significantly.
from d3rlpy.algos import DiscreteCQL
# choose data augmentation types
cql = DiscreteCQL(augmentation=['random_shift', 'intensity'],
n_augmentations=2)
You can also tune data augmentation parameters by yourself.
from d3rlpy.augmentation.image import RandomShift
random_shift = RandomShift(shift_size=10)
cql = DiscreteCQL(augmentation=[random_shift, 'intensity'],
n_augmentations=2)
Image Observation¶
d3rlpy.augmentation.image.RandomShift |
Random shift augmentation. |
d3rlpy.augmentation.image.Cutout |
Cutout augmentation. |
d3rlpy.augmentation.image.HorizontalFlip |
Horizontal flip augmentation. |
d3rlpy.augmentation.image.VerticalFlip |
Vertical flip augmentation. |
d3rlpy.augmentation.image.RandomRotation |
Random rotation augmentation. |
d3rlpy.augmentation.image.Intensity |
Intensity augmentation. |
d3rlpy.augmentation.image.ColorJitter |
Color Jitter augmentation. |
Vector Observation¶
d3rlpy.augmentation.vector.SingleAmplitudeScaling |
Single Amplitude Scaling augmentation. |
d3rlpy.augmentation.vector.MultipleAmplitudeScaling |
Multiple Amplitude Scaling augmentation. |