Blog
Open Source
Research

Petuum Community Events in April 2021 

5 min

April 5, 2021

nVidia GTC and PyTorch Ecosystem Day are just around the corner, and we’ll be presenting our latest open-source developments and real-world use cases! We have 6 events lined up across 3 venues, covering our latest tools for building Natural Language Processing pipelines, ways to conduct Learning training faster and cheaper, plus how we’re applying Time-Series modeling to advanced manufacturing and Automatic Process Control. Join us and strike up some conversations with our speakers!


Recap

CASL Forte presentation at Nvidia GTC 2020

Modularizing Natural Language Processing 

Presented by Zhengzhong (Hector) Liu and Zecong Hu

Recent success and growth in natural language processing and artificial intelligence have given the world many new applications, techniques, models, and architectures. We’ll show how appropriate abstraction and modularization can streamline both development and deployment of NLP technologies. We’ll provide a systematic overview of NLP abstractions and breakdown, the insights of machine learning integration, and the designs of NLP systems for fast module development. You’ll learn to use off-the-shelf tools to practice the modularized NLP and build practical applications.

Video


April 12

CASL AdaptDL presentation at Nvidia GTC 2021

S31561: Presenting AdaptDL: An Open-Source Resource-Adaptive Deep Learning Training and Scheduling Framework 

Presented by Aurick Qiao

AdaptDL is an open-source framework and scheduling algorithm that directly optimizes cluster-wide training performance and resource utilization. By elastically re-scaling jobs, co-adapting batch sizes and learning rates, and avoiding network interference, AdaptDL improves shared-cluster training compared with alternative schedulers. AdaptDL can automatically determine the optimal number of resources given a job’s need. It will efficiently add or remove resources dynamically to ensure the highest-level performance.

Session Details


April 12

Petuum Industrial AI Optimum for Oil and Gas at Nvidia GTC 2021

S31595: Unlock Greater Productivity and Operational Excellence Post-COVID with the Power of AI for Oil & Gas 

Presented by Dr. Roberto Linares and Henry Guo

Use AI software to manage a structural operating capacity shift at many sites from pre-Covid and post-Covid. Many traditional control systems that rely on linear models or complex rule-based systems were unable to handle this structural capacity change and the outputs were very sub-optimal. Our industrial AI software was able to assist our customers to operationalize production ready AI/ML for chemical processing.

Session Details


April 14

Petuum Industrial AI Optimum at Machine Learning in Oil and Gas 2021

Lead AI-Powered Autonomous Oil and Gas Manufacturing with Deep Learning Optimization 

Presented by Dr. Roberto Linares

Learn how Petuum aims to lead Industry 4.0 next generation manufacturing initiatives with revolutionary Deep Learning Optimization technology for refinery processes. Petuum applies non-linear deep learning AI software to Advanced Process Control (APC), Real-Time Optimization (RTO), and advanced soft sensors applications to help increase yields while decreasing energy consumption and costs.

Session Details


April 21

CASL AdaptDL presentation at PyTorch Ecosystem

AdaptDL for PyTorch 

Presented by Aurick Qiao

AdaptDL monitors training job performance in real-time, and elastically re-scales resources (GPUs, compute instances) while jobs are running. For each training job, AdaptDL automatically tunes the batch size, learning rate, and gradient accumulation. In the cloud (e.g. AWS), AdaptDL can auto-scale the number of provisioned Spot Instances. We’ve seen shared-cluster training jobs at Petuum and our partners complete 2–3x faster on average, with 3x cheaper cost in AWS using Spot Instances!

Blog


April 21

CASL Forte presentation at PyTorch Ecosystem

Forte for PyTorch 

Presented by Zhengzhong (Hector) Liu

Forte can stitch together different modules or tools to construct a composable NLP pipeline, broken down into tasks which can be solved by individual modules. For example, data readers that read from complex data formats (e.g., HTML, CSV), NLP processors (e.g., Named Entity Recognizer, Sentiment Analyzer), and other downstream consumers (e.g., visualization, serialization), etc.

Blog

Related

Cost-effective Hyper-parameter Tuning using NNI and AdaptDL

February 23, 2021

Introducing AdaptDL, an Open Source resource adaptive deep-learning framework

September 2, 2020