August 2021 Summaries

Juan Roberto Honorato is an AI Tech Lead at Anastasia.ai, where he helps develop an ML driven platform to deliver business solutions for its customers. The company's mission is to democratize AI for every business by providing scalable and cost-effective technologies. Recently, the team has been using Ray, a distributed computing framework, which led to astonishing results in demand prediction problems. Compared to their AWS Batch implementation, their Ray implementation was 9x faster and reduced costs by 87%. The blog post explores how Anastasia's platform works, their initial attempts at solving the problem using Python's built-in parallel modules, and how they realized the need for horizontal scaling with training and hyperparameter tuning. It also discusses how Ray helps them scale easily and how Anyscale further optimizes the system.

The CartPole environment is a classic reinforcement learning problem where an agent must balance a pole on top of a moving cart to maximize its total reward over time. The agent learns by interacting with an environment and receiving rewards or penalties for its actions, with the ultimate goal of balancing the pole successfully. Through various code examples and video demonstrations, this tutorial provides an introduction to reinforcement learning concepts, such as agents, environments, observations, actions, rewards, and policies. The tutorial also explores how a neural network can represent a policy in deep reinforcement learning, and how RLlib's Proximal Policy Optimization (PPO) algorithm can be used to train the agent. Additionally, the tutorial discusses hyperparameter tuning using Ray Tune, which allows users to find optimal hyperparameters for solving the CartPole environment in the fewest timesteps possible.

FLAML is a lightweight Python library from Microsoft Research that provides efficient and economical machine learning model selection using cutting-edge algorithms designed to be resource-efficient and easily parallelizable. FLAML can utilize Ray Tune for distributed hyperparameter tuning, scaling up its AutoML methods across a cluster. The library addresses the need for economical AutoML methods by leveraging insights about the structure of the search space to choose search orders optimized for both good performance and low cost. This allows for efficient exploration of the search space while minimizing computational resources. FLAML's CFO and BlendSearch methods demonstrate this approach, offering advantages in finding good configurations quickly and navigating ones with good performance while concentrating on low evaluation time. The library is integrated with Ray Tune, enabling parallelization of hyperparameter search and leveraging cutting-edge optimization algorithms at scale. FLAML has been shown to achieve comparable or better performance than state-of-the-art AutoML solutions using significantly fewer computational resources, demonstrating its potential for real-world applications.

The new version 1.6 of Ray includes several key updates, including the introduction of Ray Datasets for large-scale data loading, Runtime Environments for better dependency packaging, support for GCP TPU VMs in the Autoscaler, and the addition of Ray Lightning as a plugin for PyTorch Lightning to enable parallel training. These features aim to address interoperability issues, provide high-level abstractions, and simplify the process of running large-scale data processing applications on Ray. The updates are designed to make it easier for users to work with Ray and take advantage of its distributed computing capabilities.

TL;DR: Use PyTorch Lightning with Ray to enable multi-node training and automatic cluster configuration with minimal code changes. PyTorch Lightning abstracts away engineering code, making deep learning experiments easier to reproduce and improving developer productivity. However, parallelizing training across multiple GPUs requires significant expertise and infrastructure setup. Ray Lightning simplifies this process by providing a simple plugin for PyTorch Lightning that can scale out training with minimal code changes, works with Jupyter Notebooks, seamlessly creates multi-node clusters on AWS/Azure/GCP, integrates with Ray Tune, and is fully open source and free to use. With Ray Lightning, scaling up PyTorch Lightning training becomes much easier and more flexible, allowing users to run their training jobs programmatically and automatically scale instances up and down as they train.

The Ikigai Labs platform provides an AI-charged spreadsheet that offers a collaborative space for data integration, processing, visualization, and interactive dashboards. The platform aims to achieve three missions simultaneously: mission-critical data pipelines, highly interactive data pipelines, and instantly browsable data pipelines. To overcome technical challenges, the team used Ray Core, which provided flexibility in resolving issues such as scalability and interactivity. Ray Serve was also adopted to serve custom Python scripts with real-time interaction, while overcoming challenges related to dependency management and task submission overhead. The platform implemented a version-aware deployment mechanism to avoid race conditions and utilized Ray Client for eliminating the task submission step. Additionally, Ray Workflow is being integrated to add fault-tolerance to remote tasks and optimize resource management.

Ray is a fast, simple distributed execution framework that makes it easy to scale your applications by parallelizing and distributing existing Python code with minimal changes. It allows you to leverage state-of-the-art machine learning libraries while running on multiple CPU cores or clusters.

LightGBM-Ray is a new framework that integrates LightGBM with the distributed computing platform Ray, allowing users to easily scale their LightGBM training and prediction workloads on large clusters or cloud providers. The framework provides seamless integration with Ray Tune for hyperparameter search, multi-node and multi-GPU training support, and built-in support for categorical variables. It offers several advantages over XGBoost-Ray, including faster training times, better accuracy in certain situations, and optimized performance for larger datasets. LightGBM-Ray does not change the underlying LightGBM code but rather leverages Ray to manage data sharding and actors, ensuring fault-tolerant distributed training and prediction. While it currently relies on XGBoost as a hard dependency, efforts are underway to remove this requirement in future releases.