Deploying models is becoming easier every day, especially thanks to excellent tutorials like Transformers-Deploy. It talks about how to convert and optimize a Huggingface model and deploy it on the Nvidia Triton inference engine. Nvidia Triton is an exceptionally fast and solid tool and should be very high on the list when searching for ways to deploy a model. Our developers know this, of course, so ClearML Serving uses Nvidia Triton on the backend if a model needs GPU acceleration.

It has been months ago when Toloka and ClearML met together to create this joint project. Our goal was to showcase to other ML practitioners how to first gather data and then version and manage data before it is fed to an ML model. We believe that following those best practices will help others build better and more robust AI solutions. If you are curious, have a look at the project we have created together.

What if I want to serve a Huggingface model on ClearML? Where do I start? In general, machine learning engineers know by now that a good model serving engine is invaluable when serving models in production. These days, NVIDIA’s Triton inference engine is a popular option to do so, but it is lacking in some respects.

The popular object detection model and framework made by ultralytics now has ClearML built-in. It’s now easier than ever to train a YOLOv5 model and have the ClearML experiment manager track it automatically. But that’s not all, you can easily specifiy a ClearML dataset version ID as the data input and it will automatically be used to train your model on. Follow us along in this blogpost, where we talk about the possibilities and guide you through the process of implementing them.

Fully automatic retraining loop using ClearML Data Right, so you want to create a fully automatic retraining loop, that you can set up once and then pretty much forget about. Where do we even start?!

With the rise of drone usage, it’s easier to take aerial footage than ever before. The resulting data can trigger quick, effective action; removing guesswork and increasing aerial awareness, which can have profound implications on growing profits and trimming expenses. And as drone use rises, so does the usage of AI, to navigate, detect, identify, and track meaningful artifacts and objects.

Sometimes the need for processing power you or your team requires is very high one day and very low another. Especially in machine learning environments, this is a common problem. One day a team might be training their models and the need for compute will be sky high, but other days they’ll be doing research and figuring out how to solve a specific problem, with only the need for a web browser and some coffee.

The algorithm team at WSC Sports faced a challenge. How could our computer vision model, that is working in a dynamic environment, maintain high quality results? Especially as in our case, new data may appear daily and be visually different from the already trained data. Bit of a head-scratcher right? Well, we’ve developed a system that is doing just that and showing exceptional results!

This is part 3 of our 3-part Hyperparameter Optimization series, if you haven’t read the previous 2 parts where we explain ClearML’s approach towards HPO, you can find them here and here. In this blog post, we will focus on applying everything we learned to a “real world” use case.

If someone had told my 15-years-ago self that I’d become a DevOps engineer, I’d have scratched my head and asked them to repeat that. Back then, of course, applications were either maintained on a dedicated server or (sigh!) installed on end-user machines with little control or flexibility. Today, these paradigms are essentially obsolete; cloud computing is ubiquitous and successful.