Download presentation
Presentation is loading. Please wait.
1
Machine Learning & Data Science Conference
1/26/2018 3:24 AM © 2015 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
2
Machine Learning & Data Science Conference
1/26/2018 3:24 AM © 2015 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
3
Why Do We Do… Deep learning GPU Cloud deep learning MXNet R
Multiple stages of nonlinear processing (training, evaluation) Powerful feature learning and representation in computer vision, speech and other tasks GPU Significant acceleration of training Cloud deep learning The gravity of data is moving to the cloud MXNet The power of an open platform Language bindings: R, Python, Scala, etc. R “Golden child of data science”
4
A Deep Learning Stack on a Cloud GPU VM
Deep Learning Framework CUDA, cuDNN MKL (or equiv.) OS (Windows or Ubuntu) NVIDIA K80 GPU CPU
5
But How About Data Wrangling?
Data Lake REPL Feature Engineering Data Pipelines Deep Learning with GPU Visualization Stats Spark Hadoop SQL NoSQL Operationalization Data Cleansing
6
Unified Data Science Workflow With Microsoft R Server
Deep Learning With GPU Data Lake Spark Data Pipelines SQL Visualization NoSQL Stats Feature Engineering Operationalization Deep Learning with GPU (training) HDInsight Spark (scoring) Azure Data Lake Store
7
Machine Learning & Data Science Conference
1/26/2018 3:24 AM Microsoft R Server What should we talk here? © 2015 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
8
1/26/2018 3:24 AM GPU on the Cloud © 2014 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
9
Building from Scratch Download & install CUDA, cuDNN and MKL
1/26/2018 3:24 AM Building from Scratch Download & install CUDA, cuDNN and MKL Configure and build MXNet Download and install Microsoft R Server (MRS) You are good to go in one hour Copy the instance for future use Check out our blog © 2014 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
10
DSVM for the Impatient…
1/26/2018 3:24 AM DSVM for the Impatient… Data Science Virtual Machine: built for data scientists Microsoft R Server Anaconda Python Jupyter with R and Python kernels Visual Studio Community Edition Power BI Desktop SQL Server 2016 Developer Edition Machine learning tools on DSVM: Vowpal Wabbit xgboost Mxnet and CNTK through deep learning toolkit © 2014 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
11
Demo: Starting MXNet on DSVM
12
Running MXNet on HDInsight Premium
1/26/2018 3:24 AM Running MXNet on HDInsight Premium © 2014 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
13
Scoring Deep Learning on HDInsight Premium
Microsoft R Server (MRS) HDInsight Spark MRS Edge Node HDInsight Spark Worker Nodes Ubuntu Linux Azure Data Lake Store Windows Azure Blob Storage
14
HDInsight Premium = Microsoft R Server + Spark
Machine Learning & Data Science Conference 1/26/2018 3:24 AM HDInsight Premium = Microsoft R Server + Spark Can we say HDI Premium = MRS + Spark? © 2015 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
15
Deep Learning Framework
Microsoft R Open Microsoft R Server Datasize In-memory In-Memory or Disk Based Speed of Analysis Single threaded Multi-threaded Multi-threaded, parallel processing 1:N servers Support Community Community + Commercial Analytic Breadth & Depth 7500+ innovative analytic packages 7500+ innovative packages + commercial parallel high-speed functions Licence Open Source Commercial license. Supported release with indemnity Copyright Microsoft Corporation. All rights reserved.
17
1/26/2018 3:24 AM Example © 2014 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
18
Training Dataset: CIFAR-10 Dataset
Learning Multiple Layers of Features from Tiny Images, Alex Krizhevsky, 2009.
19
Test Subject
20
Test Subject + CIFAR-10 Test Batch
21
Workflow Overview Microsoft R Server NC-24 GPU VM MXNet
Azure Data Lake Store NC-24 GPU VM MXNet HDInsight Spark MXNet Script Action Training Data Train Input Data Score Output Data Model Export
22
convolution -10 10
23
convolution -1 10
24
Convolutional Neural Networks
25
Microsoft Research: ResNet
26
Demo: MXNet Build on Azure NC24 GPU VM Ubuntu 16.04
Machine Learning & Data Science Conference 1/26/2018 3:24 AM Demo: MXNet Build on Azure NC24 GPU VM Ubuntu 16.04 © 2015 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
27
Demo: MXNet Scoring on HDInsight Premium
28
1/26/2018 3:24 AM Results © 2014 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
29
Test Subject Tagged
30
Machine Learning & Data Science Conference
1/26/2018 3:24 AM Active Learning Train Score Enhance Do we need this slide? What will you talk about? We are seriously over budget in terms of # of slides Assess © 2015 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
31
Summary GPU and big data unlocked deep learning’s potential
1/26/2018 3:24 AM Summary GPU and big data unlocked deep learning’s potential Most of the data will live in the cloud We showed how cloud deep learning can be integrated into data science workflow using Microsoft R Server Spark, Azure GPU and Data Science Virtual Machine MXNet © 2014 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
32
Deep Learning Applications
1/26/2018 3:24 AM Deep Learning Applications Computer vision Do image analysis, OCR and more with Microsoft cognitive services computer vision API Surveillance, medical images financial transactions Speech and natural language processing Do speech and speaker recognition, language understanding and text analytics with Microsoft cognitive services language and speech APIs Machine translation (example: real-time translation on Skype) Reinforcement Learning Is reinforcement learning is applicable? © 2014 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
33
Additional Resources Read our blog: (http://aka.ms/mxnet-mrs)
1/26/2018 3:24 AM Additional Resources Read our blog: ( Sign up for Azure GPU ( Spin up DSVM ( MSR Deep Residual Learning Talk (YouTube video) Deng Li, Dong Yu, “Deep Learning: Methods and Applications”. Brandon Rohrer, “Deep Learning Demystified” (YouTube video) Acknowledgements Patrick Buehler, Vivek Gupta, Huseyin Yildiz, Karan Batta, Paul Shealy, Gopi Kumar, Mario Inchiosa, Yunshan Zhu, Jianhui Wu (Microsoft); Qiang Kou (Indiana U.); Tianqi Chen (U. Washington). © 2014 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
34
1/26/2018 3:24 AM © 2015 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
35
Final Image Tagging Deliverable
Training: Azure NC-24 GPU VM MXNet Build Ubuntu Linux Microsoft R Server CIFAR-10 Train Data Synthetic Tile Creation Train Set Validation Set Trained MXNet Model Export Test Batch ADLS HDI Spark Cluster with MXNet Running from MRS through rxExec MRS rxExec Parallel Tiling & Scoring Mona Lisa Mosaic Distributed Calls to rxExec Bounding Box Generation Tagged High-Resolution Mona Lisa Mosaic … HDI Spark MRS Edge Node HDI Spark MRS Edge Node Spark Worker Nodes
36
Overview of the talk just for us
Why Examples itemization at construction site, grocery store, night club female / male ratio, medical imaging, transaction data fraud LIDAR / infrared / metadata enhancement NLP Generative vs Discriminative models Negatives – real intelligence Reenforcememnt Learning Talk Overview Tools Overview Architecture Overview MXNet transparency of code (compare to CNTK?) use cases: language bindings, Kaggle popularity Azure VMs Linux Mxnet build – why? How? Example CIFAR training (model resume, more pre-init with different training set) Windows DSVM pre-built MXNet with examples: show NLP Python example here, go over other examples HDI Spark Why (ADLS goes here)? How (show installation script)? CIFAR scoring example from the blog: training on Azure GPU VM, ADLS upload scoring on HDI Spark with ADLS through MRS Conclusion
37
Preferred text layout (no bullets)
1/26/2018 3:24 AM Preferred text layout (no bullets) Main topic 1: size 40pt Size 20pt for the subtopics Main topic 2: size 40pt Main topic 3: size 40pt © 2014 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
38
Bullet points layout with subtitle Subtitle
1/26/2018 3:24 AM Bullet points layout with subtitle Subtitle Example of a bulleted slide with a subhead Set the slide title to “Sentence case” Set subheads to “Sentence case” Hyperlink style © 2014 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
39
Video
40
Demo Speaker name
41
1/26/2018 3:24 AM Photo layout 1 © 2014 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
42
Microsoft brand guidelines
1/26/2018 3:24 AM Microsoft brand guidelines Looking for more slide resources? Brand guidelines for PowerPoint templates is a separate slide deck that provides an overview of the Microsoft brand, guidelines, resources, tips and much more. A few of the slides are shown at right. Download from: ations/Pages/StoryBoard.aspx?section=Elements1 © 2014 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
43
Deep Learning in Microsoft R Server Using MXNet on High-Performance GPUs in the Public Cloud
1/26/2018 3:24 AM © 2014 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
44
1/26/2018 3:24 AM Section title © 2014 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
45
1/26/2018 3:24 AM Section title © 2014 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
46
1/26/2018 3:24 AM Section title © 2014 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
47
1/26/2018 3:24 AM Software code slide This slide layout uses Consolas, a monotype font which is ideal for showing software code. © 2014 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
48
Machine Learning & Data Science Conference
1/26/2018 3:24 AM © 2015 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
49
Adjusting list levels Main topic 1: size 40pt Main topic 2: size 40pt
1/26/2018 3:24 AM Adjusting list levels Main topic 1: size 40pt Size 20pt for the ssubtopics Size 20pt for the subtopics Main topic 2: size 40pt Main topic 3: size 40pt Use the “Decrease List Level” and “Increase List Level” tools on the Home Menu to change text levels. Try this: Place your cursor in any row of text to the left that says “Size 20pt for subtopics” Next click the Home tab, and then on the “Decrease List level” tool. Notice how the line moves up one level. Now try placing your cursor in one of the “Main topic…” lines of text. Click the “Increase List Level” tool and see how the text is pushed in one level Use these 2 tools to adjust your text levels as you work © 2014 Microsoft Corporation. All rights reserved. Microsoft, Windows, and other product names are or may be registered trademarks and/or trademarks in the U.S. and/or other countries. The information herein is for informational purposes only and represents the current view of Microsoft Corporation as of the date of this presentation. Because Microsoft must respond to changing market conditions, it should not be interpreted to be a commitment on the part of Microsoft, and Microsoft cannot guarantee the accuracy of any information provided after the date of this presentation. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
50
Accent colors 1-6 – (6 Theme Colors to the far right)
1/26/2018 3:24 AM Slide palette info The PowerPoint palette for this template has been built for you and is shown below. Avoid using too many colors in your presentation. Accent colors 1-6 – (6 Theme Colors to the far right) Accent 1 Accent 2 Accent 3 Accent 4 Accent 5 Accent 6 Use Accents 4-6 sparingly – only when more colors are necessary. Use Accent 1 as the main accent color. Use Accent 2 and Accent 3 when additional colors are needed. © 2014 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
51
Technologies and Tools Covered in This Talk
HDInsight Spark + Azure Data Lake Store Microsoft R Server
52
High-Level Process Overview
Training Data Training: Azure NC-24 GPU VM MXNet Build Ubuntu Linux Microsoft R Server Azure Data Lake Store Scoring Data Scoring: Azure HDInsight Apache Spark MXNet Script Action Ubuntu Linux Microsoft R Server Scored Data
53
Talk Overview GPU on the Cloud Running MXNet on HDInsight Premium
1/26/2018 3:24 AM Talk Overview GPU on the Cloud Running MXNet on HDInsight Premium An Ex © 2014 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
Similar presentations
