Description

A data augmentation library for Deep Learning that supports images, segmentation masks, labels, and keypoints. Furthermore, SOLT is fast and has OpenCV in its backend. Full auto-generated docs and examples are available at https://oulu-imeds.github.io/solt/.

Features

• Support of Images, masks, and keypoints for all the transforms (including multiple items at the time)
• Fast and PyTorch-integrated
• Convenient and flexible serialization API
• Excellent documentation
• Easy to extend
• 100% Code coverage

Examples

Images: Images + Keypoints: Medical Images + Binary Masks: Medical Images + Multiclass Masks

E.g. the last row is generated using the following transforms stream.

stream = solt.Stream([
    slt.Rotate(angle_range=(-20, 20), p=1, padding='r'),
    slt.Crop((256, 256)),
    solt.SelectiveStream([
        slt.GammaCorrection(gamma_range=0.5, p=1),
        slt.Noise(gain_range=0.1, p=1),
        slt.Blur()    
    ], n=3)
])

img_aug, mask_aug = stream({'image': img, 'mask': mask})

If you want to visualize the results, you need to modify the execution of the transforms:

img_aug, mask_aug = stream({'image': img, 'mask': mask}, return_torch=False).data

Installation

The most recent version is available in pip:

pip install solt

You can fetch the most fresh changes from this repository:

pip install git+https://github.com/MIPT-Oulu/solt

Benchmark

We propose a fair benchmark based on the refactored version of the one proposed by albumentations team. Still, here, we also convert the results into a PyTorch tensor and do the ImageNet normalization. The following numbers support a realistic and honest comparison between the libraries (number of images per second, the higher - the better):

	albumentations 0.4.3	torchvision (Pillow-SIMD backend) 0.5.0	augmentor 0.2.8	solt 0.1.9
HorizontalFlip	2253	2549	2561	3530
VerticalFlip	2380	2557	2572	3740
RotateAny	1479	1389	670	2070
Crop224	2566	1966	1981	4281
Crop128	5467	5738	5720	7186
Crop64	9285	9112	9049	10345
Crop32	11979	10550	10607	12348
Pad300	1642	109	-	2631
VHFlipRotateCrop	1574	1334	616	1889
HFlipCrop	2391	1943	1917	3572

Python and library versions: Python 3.7.0 (default, Oct 9 2018, 10:31:47) [GCC 7.3.0], numpy 1.18.1, pillow-simd 7.0.0.post3, opencv-python 4.2.0.32, scikit-image 0.16.2, scipy 1.4.1.

The code was run on AMD Threadripper 1900. Please find the details about the benchmark here.

How to contribute

Follow the guidelines described here.

Author

Aleksei Tiulpin, Research Unit of Health Sciences and Technology Faculty of Medicine University of Oulu, Finland.

How to cite

If you use SOLT and cite it in your research, please, don't hesitate to send an email to Aleksei Tiulpin. All the papers that use SOLT are listed here.

@misc{solt2019,
  author       = {Aleksei Tiulpin},
  title        = {SOLT: Streaming over Lightweight Transformations},
  month        = jul,
  year         = 2019,
  version      = {v0.1.9},
  doi          = {10.5281/zenodo.3702819},
  url          = {https://doi.org/10.5281/zenodo.3702819}
}