We need intelligence. More specifically, our planet’s total application and data services layer needs an increasing amount of Artificial Intelligence (AI) and so, logically, we also need a strong handle on Machine Learning (ML) and the data wrangling, data management and data orchestration functions needed to feed the ML pipeline with goodness.
That’s where MLOps comes in.
It would (arguably) be easy to interpret MLOps the wrong way around i.e. this is not Machine Learning intelligence to drive IT Ops (operations) so that sysadmins, database administrators (DBAs), penetration testers and the rest of the team get better AI functions, it is in fact that process in reverse.
In real terms, MLOps is operations ‘for, of and dedicated to’ the Machine Leaning function i.e. it is Ops procedures and workflows designed to provide a higher quality data stream for the ML function to ingest and process. so that the resulting AI engine build is more performant and, well, smarter basically.
Ultimately of course, MLOps does provide Ops with ML, so the first definition above is also true, but only if the second notion is carried out first and foremost. Tautological twists notwithstanding then, what’s happening in this space now?
Open source MLOps
ClearML is an open source, MLOps platform that helps data science, MLOps and DevOps teams develop, orchestrate and automate ML workflows. All of that happens, at scale, as you might expect.
When the organisation talks about MLOps (which as we now know is the process of applying good operations support to data management functions that feed an ML pipeline and so inform the higher AI engine it might serve with intelligence and machine know-how), the company talks about it in the context of being driven by a unified end-to-end MLOps suite.
Late last year we saw ClearML partner with AI specialist Ultralytics to integrate ClearML’s open source toolkit into Ultralytics’ colourfully named YOLOv5 stack. But why?
Because the firms want to enable what they call a more ‘streamlined’ approach to ML that enables ‘experiment tracking’ to a higher level.
What is ML experiment tracking?
A nicely defined by NeptuneAI here, ML experiment tracking is the process of saving all experiment-related information that you care about for every experiment you run.
NeptuneAI clarifies further and says that this is, “Metadata you care about [and it]will strongly depend [and differ] on your project, but it may include: scripts used for running the experiment; environment configuration files; versions of the data used for training and evaluation; parameter configurations; evaluation metrics; model weights; performance visualisations; and example predictions on the validation set (common in computer vision).
Back with ClearML, the company says that this means users will be able to track every YOLOv5 training run in ClearML Experiment; version and connect datasets to models (ClearML Data); automate execution on local and remote cloud machines; get the very best mAP using ClearML Hyperparameter Optimization; deploy YOLOv5 models with state-of-the-art Triton serving engine (ClearML Serving); and build model performance dashboards in Grafana.
“We’re excited about our integration with YOLOv5,” said Moses Guttmann, co-founder and CEO of ClearML. “This integration makes it even simpler to train a YOLOv5 model and use ClearML Experiment to track it automatically. Users can easily leverage ClearML Data to create a data lineage and connect a YOLOv5 Model to the underlying data that was used to train it, resulting in enhanced user understanding of a model’s behaviour.”
ML model training in the cloud
Guttmann noted that ClearML enables users to run and automate model training in the cloud as well as deploy models to production environments.
“ClearML brings awesome new capabilities to YOLOv5 that I’m very excited about,” said Glenn Jocher, founder and CEO of Ultralytics. “Users can track experiments, visualise results interactively and organise and queue YOLOv5 runs remotely, all in one place for free. The remote training capabilities are especially powerful, and not something that has been available before with YOLOv5. This is just one more step in Ultralytics’ quest to make AI easy and to help everyone build and deploy amazing solutions with YOLOv5!”
The YOLO algorithm broke traditional computer vision boundaries by creating the world’s first fast and accurate object detection solution, replacing older two-stage AI solutions and even older classical methods. Its makers claim that now, YOLOv5 is the world’s most popular open-source object detection AI.
Launched by Ultralytics in 2020, it has arguably moved to the forefront of the AI space. A favourite tool for vision tasks, YOLOv5 by Ultralytics now has 300 contributors, 33,000 GitHub stars and hundreds of thousands of active monthly users and organisations around the world who create real-world AI products and services of their own.
What is streamlined ML again?
So once again, what exactly is streamlined ML again? The team behind ClearML say streamlined ML is what’s happening here i.e. a platform that it is purpose-built for the MLOps industry, enabling MLOps teams to build, execute, manage, monitor and analyse the entire MLOps process from a single fully integrated platform – all with just two lines of code.