pyTORch augMENTOR
Image data augmentation with pytorch
Instalation:
Instaling current version with pip:
pip install tormentor
#pip3 install --user --upgrade git+https://github.com/anguelos/tormentorUse Cases:
Augment single sample:
python import torch, tormentor
img = torch.rand(3, 119,137)
mask = torch.ones([1, 119,137])
pc = (torch.rand(9),torch.rand(9))
aug = tormentor.Perspective()
new_img = aug(img)
aug(pc, img) # augment pointcloud and respective image
aug(pc, img, compute_img=False) # augment only pointcloud, img passed for dimensions
aug(pc, torch.empty([1, 320, 240]), compute_img=False) # augment only pointcloud, tensor passed for dimensions
aug(mask, is_mask=True) # augment maskAugment batch:
python import torch, tormentor
img = torch.rand(7,3, 119,137)
mask = torch.ones([7,1, 119,137])
pcl = [(torch.rand(9), torch.rand(9)) for _ in range(7)]
aug = tormentor.Rotate()
new_pcl, new_img = aug(pcl, img) # augment pointcloud and respective image
aug(pcl, img, compute_img=False) # augment only pointcloud, img passed for dimensions
aug(pcl, torch.empty([7, 1, 320, 240]), compute_img=False) # augment only pointcloud, tensor passed for dimensions
aug(mask, is_mask=True) # augment maskAugment MSCoco compliant Dataset:
import torchvision, tormentor
ds = torchvision.datasets.CocoDetection(root="./tmp_data/coco/val2017",
annFile="./tmp_data/coco/annotations/instances_val2017.json",
transform=torchvision.transforms.ToTensor());
aug_ds = tormentor.AugmentedCocoDs(ds, tormentor.Wrap(), device="cpu", add_mask=True)
inputs, target, validity = aug_ds[3] # accesing a single sample
aug_ds.show_augmentation(3)
If the device is a GTX 980 Ti time is 0.1 sec. for larger images, the GPU efficiency grows up to x10.
Change Augmentation Distributions:
import math, tormentor, torch, torchvision
tile = lambda x: torchvision.transforms.ToPILImage()(torchvision.utils.make_grid(x.cpu(), nrow=12))
generic_aug = tormentor.Rotate()
RotateABit = tormentor.Rotate.override_distributions(radians = tormentor.Uniform((0., math.pi / 8)))
custom_aug = RotateABit()
batch = torch.rand(24, 3, 32, 38, device="cuda")
tile(torch.cat([batch, generic_aug(batch), custom_aug(batch)], dim=0)).show()
Random Augmentation Type:
import math, tormentor, torch, torchvision
tile = lambda x: torchvision.transforms.ToPILImage()(torchvision.utils.make_grid(x.cpu(), nrow=12))
augmentation_types = [tormentor.Perspective, tormentor.Wrap, tormentor.PlasmaBrightness]
CustomAugmentation = tormentor.AugmentationChoice.create(augmentation_types)
aug = CustomAugmentation()
batch = torch.rand(24, 3, 64, 64, device="cuda")
tile(aug(batch)).show()
# checkup on determinism:
tile(aug(batch)).show()
Augmentation Cascade:
import math, tormentor, torch, torchvision
tile = lambda x: torchvision.transforms.ToPILImage()(torchvision.utils.make_grid(x.cpu(), nrow=12))
augmentation_types = [tormentor.Perspective, tormentor.PlasmaBrightness]
CustomAugmentation = tormentor.AugmentationCascade.create(augmentation_types)
aug = CustomAugmentation()
batch = torch.rand(24, 3, 64, 64, device="cuda")
tile(aug(batch)).show()
Create Custom Augmentation Class:
python import tormentor
class Lense(tormentor.SpatialImageAugmentation):
center_x = tormentor.Uniform((-.3, .3))
center_y = tormentor.Uniform((-.3, .3))
gamma = tormentor.Uniform((1., 1.))
def generate_batch_state(self, sampling_tensors):
batch_sz = sampling_tensors[0].size(0)
gamma = type(self).gamma(batch_sz, device=sampling_tensors[0].device).view(-1)
center_x = type(self).center_x(batch_sz, device=sampling_tensors[0].device).view(-1)
center_y = type(self).center_y(batch_sz, device=sampling_tensors[0].device).view(-1)
return center_x, center_y, gamma
@classmethod
def functional_sampling_field(cls, sampling_field, center_x, center_y, gamma):
field_x, field_y = sampling_field
center_x = center_x.unsqueeze(dim=1).unsqueeze(dim=1)
center_y = center_y.unsqueeze(dim=1).unsqueeze(dim=1)
gamma = gamma.unsqueeze(dim=1).unsqueeze(dim=1)
distance = ((center_x - field_x)**2 + (center_y - field_y)**2) ** .5
#distance = 1/(1+distance)
field_x, field_y = (field_x + field_x * distance ** gamma) , (field_y + field_y * distance ** gamma)
return field_x, field_y
Design Principles
- Simplify the definition of augmentations
- Every instance of every augmentation class is deterministic.
- Inputs and Outputs are pytorch tensors and pytorch is prefered for all computation.
- All data are by default 4D: [batch x channel x width x height].
- Single sample augmentation: batch-size must always be 1.
- Threadsafety: Every augmentation instance must be threadsafe.
- Input/Output is restricted to one or more channels of 2D images.
- Augmentations either preserve channels or the preserve pixels (space).
- The augmentation class has also its factory as a classmethod
- Restrict dependencies on torch and kornia (at least for the core packages).
Factory Dictates Constructor
In order to minimize the code needed to define an augmentation. The factory defines the random distributions from wich augmentation sample. The inherited constructor handles random seeds. The method forward_sample_img samples from the random distributions aug_parameters and employs them.
Internal Conventions
- Pointclouds are represented in image coordinates Sampling fields in normalised -1,1 coordinates
- By default we write code for batch processing
- Determinism is strictly handled by BaseAugmentation and all augment_*** methods.
- An augmentation must reside in a single device
- All randomness must be coming from pytorch
- Spatial augmentation samplingfields are normalised to -1, 1 so their effect magnitude is proporsional to image size (They are top down).
Cite
@misc{tormentor,
doi = {10.48550/ARXIV.2204.03776},
url = {https://arxiv.org/abs/2204.03776},
author = {Nicolaou, Anguelos and Christlein, Vincent and Riba, Edgar and Shi, Jian and Vogeler, Georg and Seuret, Mathias},
keywords = {Computer Vision and Pattern Recognition (cs.CV), Artificial Intelligence (cs.AI), FOS: Computer and information sciences, FOS: Computer and information sciences},
title = {TorMentor: Deterministic dynamic-path, data augmentations with fractals},
publisher = {arXiv},
year = {2022},
copyright = {Creative Commons Attribution 4.0 International}
}Accepted at ECV 2022 .
Download (pdf) .
