Catalyst 21.09 — top3 features
Hi guys, nice project! I am Sergey, creator of the Catalyst framework, and today I would like to give a short dive into the novel features that we recently released, thanks to our contributors.
SelfSupervised extensions
Self-supervised learning is a hot topic nowadays, and as a framework for accelerated deep learning R&D, we have no option to miss it! Efficient representations are the core of deep learning, so they can boost all its domains simultaneously. As an R&D community, we would like to give every practitioner simple yet powerful abstractions to experiment with their ideas. A short introduction to our vision, you could find under the PyTorch Community Voices talk.
To do so, we have extended our core library Runers with a SelfSupervised one. An overview of our abstractions you could find in our previous post:
Long story short, now you could run your self-supervised experiments with only a few lines of code:
More examples you could find under our self-supervised folder. We are also working on a large-scale self-supervised training blog post, so stay tuned!
Metrics, criterions, loggers
Besides SelfSupervisedRunner abstraction, to get examples done, we also implement a wide variety of self-supervised criterions from different papers:
Moreover, we have also implemented an R-squared metric and callback, so give it a try your next regression time.
Finally, thanks to our contributors, we have also integrated the Comet.ml logger, so you can log everything with cosmic speed:
CI/CD resurrection
Last but not least, thanks to our DevOps skills, we have also re-invent our CI/CD pipelines for docker support. Now you could start working with the Catalyst even faster:
Those were the top3 features of our last Catalyst release. For more detailed history, check out our changelog or Github releases page.
If you have found these features helpful for you — share the post, star the project, and join our slack channel for community discussions and calls.
And big thanks to all our friends and contributors for this release!
