We’re happy to share an update on Labelformat - A tool for converting computer vision label formats. This project started internally at Lightly as we had to deal with lots of different labeling formats for our projects. We didn’t find anything working out of the box without having to sign up for a service. We just wanted some easy-to-use tool that could read and write from different formats.
The open-source package is under the MIT license. Our hope is that others might contribute additional formats to the repo to solve this issue. Please take a look. If you like what we do and think a format is missing, create an issue. If you feel motivated enough and have the time, we would be super grateful for a PR. And if you don’t have the time to create an issue or PR but still want to support us, you can also leave a star :)
Features
- Support for common dataset label formats such as YOLO, COCO, Kitti (more coming soon)
- Support for common tool formats such as Labelbox (more coming soon)
- Minimal dependencies, targets Python 3.7 or higher
- Memory conscious - datasets are processed file-by-file instead of loading everything in memory (when possible)
- Typed
- Tested with round trip tests to ensure consistency
- MIT license
- Runs on Windows, Linux, macOS
Github link:
I would not recommend it unless you only focus on smaller models and small experiments. Biggest advantage is the huge amount of memory available. But the bottleneck is memory bandwidth.
We did some tests out of fun (as there were not many benchmarks available). You can find the results here:
https://www.lightly.ai/post/apple-m1-and-m2-performance-for-training-ssl-models
Support got better but back when we did the tests there was still no proper half precision support and also torch.compile wouldn’t work. There is hope that the software support will catch up. I’m curious to see other results. We definitely need more benchmarks :)