August 2020 Summaries

The upcoming Ray Summit will focus on scalable machine learning, scalable Python, and making it accessible to everyone. The conference aims to highlight the challenges of developing scalable applications, particularly in machine learning and deep learning, and how Ray is addressing these challenges by providing a platform for distributed computing and state management. Speakers from leading companies and open-source communities, including financial services, retail, manufacturing, media, and advertising, will share their experiences with Ray. The conference also features presentations on machine learning and cloud-native services, as well as discussions on the future of distributed computing, reinforcement learning, and MLOps. Keynote speakers include experts from leading companies such as Uber, Autodesk, and Microsoft, who will discuss their use cases for scalable machine learning and RL applications.

PyTorch Lightning is a popular framework in machine learning that offers several benefits, including reduced boilerplate code and improved scalability. To perform hyperparameter tuning, which is crucial for model selection, researchers can use Ray Tune, an industry-standard tool that provides various features such as the ability to run multiple training runs at scale and analyze results with hyperparameter analysis tools. By adding a callback to communicate with Ray Tune, users can make their PyTorch Lightning models configurable and find the best combination of hyperparameters for their model by running Ray Tune's search algorithm on a defined search space. The process is straightforward, requiring only a few lines of code, and does not require changing anything in the LightningModule. Ray Tune supports various methods for defining the search space, and users can visualize the training progress using TensorBoard. By leveraging Ray Tune with PyTorch Lightning, researchers can easily perform hyperparameter tuning and improve their model's accuracy.

Ray Summit, a free online event, has concluded its fourth iteration, Ray Summit Connect, which took place on August 12, 2020. The event featured talks and presentations by experts in the field of machine learning, including Simon Mo discussing Ray Serve, a scalable and programmable ML serving framework, and Qingqing Mao explaining how to build scalable natural language processing pipelines with Ray. Additionally, Ding Ding presented Project Zouwu, a scalable AutoML solution for telco time series analysis using Ray and Analytics Zoo. A Q&A session was also held, followed by a panel discussion moderated by Dean Wampler. The event is now available in video format for those who missed it, and the organizers invite attendees to join them at their next event, Ray Summit, scheduled to take place soon.

Model servers are crucial components of software infrastructure for productionizing machine learning. They can use a trained model artifact produced using various popular tools and support multiple programming languages. Key features companies should look for in a model server include user interfaces, low-latency capabilities, microservice architecture, deployment options, and advanced monitoring and observability capabilities. Model servers should also integrate with model management tools and support efficient batch and online scoring of large datasets.