Skip to content

Latest commit

 

History

History
113 lines (94 loc) · 5.69 KB

README.md

File metadata and controls

113 lines (94 loc) · 5.69 KB
Raw

slide

PyPI version CI codecov License DOI

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: Cats Images + Keypoints: Cats Medical Images + Binary Masks: Brain MRI Medical Images + Multiclass Masks Knee MRI

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}
}