Product management and technology strategy leadership, building and bringing to market AI-based solutions.Product manager for Google's open-source ML framework - TensorFlow, providing powerful machine learning framework for solving challenging and high-impact problems. Focused on usability, adoption, and enterprise use-cases.Deep domain experience in healthcare, medical imaging, life sciences, and industrial imaging and analytics.View the profile
About the talk
To address the needs of AI-enabled businesses, Google recently introduced TensorFlow Enterprise. It incorporates enterprise-grade support, cloud scale performance, and Google Cloud-managed services. Learn how to get started and why the best way for businesses to experience TensorFlow is with TensorFlow Enterprise.
Speaker: Sandeep Gupta
Google Cloud Next ’20: OnAir → https://goo.gle/next2020
Subscribe to the GCP Channel → https://goo.gle/GCP
product: TensorFlow Enterprise Core; fullname: Sandeep Gupta;
event: Google Cloud Next 2020; re_ty: Publish;
When will the sun rise on gcp? Machine learning is transforming every aspect of our lives and solutions that drive business when you use protection. But if you are an Enterprise trying to use, there are some challenges you may face. Newtonsoft open source releases. Come out several times in this case be available, Second to make sure that you get the best support. Be design tensorflow. Enterprise is the GTA V optimized experience. It consists of three main pillars. Is performance and ability to handle
the price. Includes optimizations training on gcp. Second is manageability, your business problem. You do not want to spend your time and managing all the machines in your pasture to focus on the business logic of your problems. That matters to your business is it easier to build and run a meld with these managed services Play Jesus, come support. Let's start by looking at Clouds schedule. Forman's first. Enterprise be compiled and ship, a special breed of tensorflow, which is optimized for Google. Cloud infrastructure,
add ons with additional runtime optimization optimization which a specific for Google cloud machines and services. So this is a code snippet that you can use to train a model with potentially a large painting. He just said In Google Cloud Storage. Training data. As you see, this course, is no different from your typical points to your GCS bucket. A little Garden Supply. Enterprise best container to make this performant. Google Cloud Storage is ecp's sql-based Data Warehouse
with hundreds of millions of Euros. In many tables models, from a specific standard edition Transformations such as badging and shuffling, Optimized idle for many millions of efficient model training Whitton Supply, Enterprise improves the utilisation of accelerators such as gpus for the same accuracy as shown in the chart on the left. The purple dress. You see here is the GP utilization profile Whitton Supply, Enterprise 1.15. On one more example, you see that it is not indicate a short revolt and also attends of indicating
high at the equalisation compared to the orange glass is not happy. This better I do too but helps reduce the cost of train has achieved when you scared to use. For example, has been training the bird Lodge models for NLP problems. On a 64 GTO cluster. We see about a 30% speed up with these improvements Belton City of seen some optimizations for better scale and performance on gcp intensive. Do Enterprise and its services. We deliver tensorflow Enterprise
to a managed environment form of deep learning VM, images of a standard, Linux box and made available to you. What makes this convenient is that including device drivers, independent nations services on installing a combination of consecration And support for all Google compute engine security features, such as customers, manage, encryption keys, and V PC is built in right out of the box because these are your choices for deploying them and using them on the team
machine and accelerated Joyce. And you start the virtual machine was pretty considerate and you can start using tensorflow if you prefer, so you can go and start writing code in a notebook. Once you have a satisfactory model after running, some experiments with the full-scale or take advantage of distributed training for large-scale of data and models. Bctcs Exxon training is a managed service for serverless train training classes, and provisioning of compute resources, more importantly, it drives the same supplier
Enterprise, just trains with the full scale of data on an example of how to do this video model training code. In this example, using 10 workers for a distributor training plastered with 800 GB, use images as the default run time. Now that you've trained your model in full enterprise-scale, it is time to make it an end-to-end pipeline to continue running in production tax on pipelines, which are built with chance of Love, extended or TSX. Is hosted on Google kubernetes engine. So it can also Drive
the same tense of blow Enterprise container in confidence that your pipeline change, the same way as your training and hosted by Plains Enterprise add-ons. Let me show you two examples from the distributor training call, sing the right distribution strategy, based on the classic concentration that you provide using. The cloud is an important step in fine-tuning and optimizing your model. You can use the implementation of the Christian API, display API compatible with Christina, gcps Optimizer, search a
surprise, and make your journey from offline, experimentation to Cloudscape implementation easier. And let's talk about enterprise-grade support that comes with Enterprise. Open source. Tensorflow comes with a one-year of maintenance window back. Boarding important feature, enhancements from future releases. Based on testimony that you can play against based on your businesses. As of today, we support both sensible One. X and 2x Whitton Supply. Enterprise Long Term, support
example of what long-term support can mean for you in practice video. Recently launched their a 100 family of gpus bring significant performance. Improvements deep learning has support for 11:30 to text you Eastern Supply. Enterprise versions are the only official residence of the 1:15 and 2. One with this support belt before the slimes, in the main line, open source releases early user of these GPS on gcp with these pre-built images. Price 2.1, supports
Enterprise keep working together between Google cloud, and the dentist to cancel Flo on Google Cloud We're also looking at smoother integration between us and the training services, such as an experiment management for a seamless and joyful developer experience is this sentence of all the documentation and resources to thank you for listening.
Buy this talk
Buy this video
With ConferenceCast.tv, you get access to our library of the world's best conference talks.