[AI-1190] Implement and test augmentations

[ChristofferEdlund]

Problem

We want to introduce the possibility to use Albumentation transforms with darwin-py torch datasets.

Solution

Introducing an AlbumentationsTransform class in torch.transforms that can be be used in the following manner:

from darwin.torch.dataset import (
    ClassificationDataset,
    InstanceSegmentationDataset,
    ObjectDetectionDataset,
    SemanticSegmentationDataset,
)

from darwin.torch.transforms import AlbumentationsTransform

dataset_path = Path(r"/path/to/local/darwin/dataset")
transform = AlbumentationsTransform.from_path('/tmp/transform.json')

inst_dataset = InstanceSegmentationDataset(dataset_path=dataset_path, transform=transform)

One can initilize the AlbumentationsTransform in three ways:

• AlbumentationsTransform(albumentation_transform) <- needs an albumentations transformation
• AlbumentationsTransform.frompath(path) <- a path pointing to a .json or .yaml file defining the transformation
• AlbumentationsTransform.fromdict(dict) <- a dictionary defining the transform

To read more about the dictionary and file formats supported, we refer to the albumentations documentation.

Further.

• Instance segmentation dataset output is changed for bounding boxes to be coco format (X, Y, W, H) to be consistent with darwin-json annotations and the ObjectDetecion torch dataset. [BREAKING CHANGE]
• Clamping is introduced to ObjectDetection bboxes that is outside of image for more robust data loading.

Changelog

• Introduced albumentation transform support for darwin torch datasts
• [BREAKING CHANGE] darwin.torch.dataset.InstanceSegmentationDataset has bbox coordinates changes from pascal_voc to coco format (X, Y, H, W)
• darwin.torch.dataset.ObjectDetectionDataset clamps bbox coordinates out-of-bound.

[linear]

AI-1190 Implement and test augmentations

This task is about implementing and benchmarking augmentations for model training. We will first test it for object detection on HF models and benchmark against not using augmentations.

If this improve model performance, let's generate a more general solution where any integration can import and use the augmentations.

[owencjones]

Updated title so that it's easier for me on deployment 😉

[almazan]

All changes look good to me.

Edit: noticed something in a second pass. See below.

[almazan]

Some comments need to be addressed before approval

[almazan]

This is not clamping bboxes anymore, since now indexes 2 and 3 are height and width instead of x2 and y2. Boxes now could go outside of the image.