Download presentation
Presentation is loading. Please wait.
1
Increase the Rate of Experimentation with Azure Machine Learning
6/23/2018 9:40 PM BRK3319 Increase the Rate of Experimentation with Azure Machine Learning Sandhya Vankamamidi Senior PM Manager Ahmet Gyger Senior Program Manager © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
2
Microsoft AI Platform FAST & AGILE AI BUILT-IN ENTERPRISE PROVEN LOB
Cloud Apps + insights LOB CRM Graph Image FAST & AGILE AI BUILT-IN ENTERPRISE PROVEN Social IoT On-premises Edge
3
Azure Machine Learning
6/23/2018 9:40 PM Azure Machine Learning AZURE MACHINE LEARNING SERVICES TRAIN & DEPLOY OPTIONS Spark SQL Server Virtual machines GPUs Container services AZURE AI dev / data scientist workstation CPU/GPU ON-PREMISES Experimentation and Model Management SQL Server Machine Learning Server Notebooks IDEs Azure Machine Learning Workbench Visual Studio Code Tools for AI EDGE COMPUTING Azure IoT Edge © 2013 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.
4
Boston Hubway BikeShare
6/23/2018 9:40 PM Boston Hubway BikeShare © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
5
Demo: Setting up Azure ML
6/23/2018 9:40 PM Demo: Setting up Azure ML Sandhya Vankamamidi © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
6
Bring AI everywhere Build model Prepare data Operationalize
Benefit from the fastest AI developer cloud Prepare data Build model Operationalize
7
Data Preparation 6/23/2018 9:40 PM
© Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
8
Agile data preparation workflow
With Azure Machine Learning Read heterogenous data Sample Flatten Distill structure Acquire and rectangularize Enterprise data pipeline Schedule Deploy Scale Up/Out Secure Monitor Diagnose Profile Infer types and entities Clean Transform, extract, combine Augment Understand and make valuable Reshape Aggregate Featurize Compare Validate Prepare for consumption
9
data preparation demo Howz the weather?
6/23/2018 9:40 PM data preparation demo Howz the weather? © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
10
Ingest and sample Data store File Types Sampling Strategy
6/23/2018 9:40 PM Ingest and sample Data store File system - Azure Blob SQL Database File Types Delimited Files (CSV, TSV, TXT) Fixed Width Plain Text Excel Json Parquet Sampling Strategy Top N - Random N Full file Random % © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
11
PROSE: Program Synthesis using Example
6/23/2018 9:40 PM PROSE: Program Synthesis using Example Also powers Microsoft Excel Flash Fill How does it work ? Used in data preparation for Derive Column by example Combine Column by example Split Column by example Advanced mode for finer control Deep Dive: Split column by example A fixed delimiter, or Multiple, arbitrary delimiters appearing in particular contexts, or, Data patterns or certain entity types © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
12
Deep Dive: Split Column by Example
6/23/2018 9:40 PM Deep Dive: Split Column by Example Predictively splitting on multiple contextual delimiters Splitting IIS log Monday - Friday: 7:00 am - 6:00 pm,Saturday: 9:00 am - 5:00 pm,Sunday: Closed Timings_1 Timings_2 Timings_3 Timings_4 Timings_5 Timings_6 Timings_7 Timings_8 Timings_9 Monday Friday 7:00 am 6:00 pm Saturday 9:00 am 5:00 pm Sunday Closed [16/Oct/ :22: ] "GET /images/picture.gif HTTP/1.1" " "Mozilla/4.0 (compatible; MSIE 4)" "-" logtext_1 logtext_2 logtext_3 logtext_4 logtext_5 logtext_6 logtext_7 logtext_8 16/Oct/2016 16:22:33 -0200 GET images/picture.gif HTTP 1.1 © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
13
Other Transforms Column (duplicate, remove, keep) Convert Field,
6/23/2018 9:40 PM Other Transforms Column (duplicate, remove, keep) Convert Field, Append (rows, columns) Remove ( duplicate ) Adjust Precision Clustering, Trim String Join Summarize Summarize © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
14
data preparation demo Tripping IT!
6/23/2018 9:40 PM data preparation demo Tripping IT! © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
15
6/23/2018 9:40 PM Understanding Data 28 greats metrics to leverage in understanding the data Data Profile Min, Max, Count, Variance Mean Most common Number of missing values, NaNs, unique values, most common Quantile 50%/75% Median Kurtosis Skewness Unbiased standard error of the mean And more © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
16
Inspectors Column statistics : Numeric Histogram Value Counts Box Plot
6/23/2018 9:40 PM Inspectors Column statistics : Numeric Histogram Value Counts Box Plot Scatter Plot Time Series Map © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
17
Custom Scripts & Extensibility
6/23/2018 9:40 PM Custom Scripts & Extensibility Add Column Advanced Filter Transform Dataflow Transform Partition © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
18
Data preparation summary
Experience driven - Use ML to scale/fast-track interaction Generates code for multiple different run times - Initially Python + Spark (PySpark) Expression and escape hatch languages - Currently Python only, more later Extensible - Transforms
19
data preparation demo Scaling new Heights!
6/23/2018 9:40 PM data preparation demo Scaling new Heights! © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
20
Switching to scale up and scale out
6/23/2018 9:40 PM Switching to scale up and scale out Defining compute targets az ml computetarget attach -a <HDI instance> -u <username> -w <password> -t cluster -n <Name of compute target> Preparing Compute Targets before first use az ml experiment prepare -c myHDI Look st Aml_config/myHDI.compute © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
21
COLLABORATION & GIT 6/23/2018 9:40 PM
© Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
22
6/23/2018 9:40 PM Collaboration and Git All projects have a local Git repository, with master and run history branches Projects can be associated with a cloud Git VSTS repository - *which should have an empty master* Reproducibility = a git commit at each experiment submission Because each experiment submission is a commit, one can restore a project at the state it was at a given time. Collaboration Each user needs to have access (permission) to both the cloud git repository and the Azure Machine Learning project VSTS account AML experimentation account VSTS project Workspace Git repository AML project User needs permissions in both places © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
23
ExPERIMENTATION 6/23/2018 9:40 PM
© Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
24
Visual Studio Code Tools for AI
6/23/2018 9:40 PM Visual Studio Code Tools for AI Access AzureML functionalities from VS Code. Access, explore, and install samples. View list of jobs. View job details. Preview and download artifacts from an experiment. © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
25
Command-line Interface
6/23/2018 9:40 PM Command-line Interface Most of Azure Machine Learning functionalities are available through the CLI. Empower both the IT Admin and the Data Scientist. Integrated with Azure CLI. © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
26
6/23/2018 9:40 PM Run History & Metrics Azure Machine Learning tracks experiment using a “run history” service. Using Azureml_logger library allows to track evolution of metrics across runs. Can compare runs to one another. View panel can be customized. Job panel to track progress near real-time. # # Pattern to invoke Azure ML Logger to record metrics. # Import Azure ML Logger library from azureml.logging import get_azureml_logger # Create a new instance of the logger run_logger = get_azureml_logger() # log a value (associated to a given experiment and project) run_logger.log("key", value) # log an array of values (associated to a given run) run_logger.log("Actual", [testlabel[i] for i in range(len(testlabel))[0::100]]) © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
27
6/23/2018 9:40 PM Transparent compute Experiment are organized in a way that they can be executed on many compute target. Compute target (where the experiment is run) Local, Local Docker, Remote Docker Virtual Machine, HD Insight. IP Address, Username, Password. Run configuration (configuration on the compute target) Local setting (runtime, environment variables, …) Conda dependencies (libraries and version for experiment) © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
28
6/23/2018 9:40 PM Notebook integration Fully-interactive, in-place, Jupyter Notebook experience, complete with markdown and code cells as well as their rendered output. Fully support run history, custom metrics, and versioning. Can also be access through localhost web-browser. © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
29
Pre-trained DNN models
6/23/2018 9:40 PM Microsoft ML Spark Solve machine learning problems dealing with big data. Provide cutting edge machine learning algorithms on Spark: deep learning, computer vision, text analytics. Run on every platform and language supported by Spark. Follow the SparkML pipeline model for composability. MML- Spark consists of Estimators and Transforms that can be combined with existing SparkML components into pipelines. Use SparkML DataFrames as a common format. Consistently handle different datatypes – text, categorical, images – for different algorithms. Eliminate low-level type coercion, encoding or vector assembly. PySpark wrappers Wrapper generation Pre-trained DNN models Scala API CNTK Java Bindings OpenCV Java Bindings Spark core © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
30
Promote & download # setup environment az ml env setup
6/23/2018 9:40 PM Promote & download # setup environment az ml env setup -l <location> -n <environment name> # set environment az ml env set -g <resource-group> # Create modelmanagement account az ml account modelmanagement create --name <modelmanagement name> --resource-group <resource-group> --location <region> # deploy model az ml service create realtime -m <path to .link file> -f <scoring script> -r <runtime (spark-py|python)> –n <webservice name> Promote: transform an artifact to an asset (accessible to Azure ML). Promoting is a way to create an association between an asset and a version. Promote a trained model, so you can reference it in your scoring script. This will generate a link file that can be referenced in your scoring code and when creating the real-time service. An asset can be downloaded locally to be then used for the creating of the image used to serve your ‘Dockerized’ web-service. © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
31
Summary Open platform Intelligent data preparation
Agile experimentation Data science provenance Operationalization at scale Collaboration and sharing
32
6/23/2018 9:40 PM Sessions to watch for Lots of great sessions! Just search for ‘Machine Learning’ BRK Operationalize your models with Azure Machine Learning Hyatt Plaza International G © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
33
Getting started Where do you want to go today? Sign up at azure.com/ml
34
Getting started Documentation: https://aka.ms/azureml-docs
Where do you want to go today? Documentation: Download Azure ML workbench: (MSI for Windows) (DMG for MacOS) Microsoft Machine Learning Spark Visual Studio Code Tools for AI
35
6/23/2018 9:40 PM Q & A © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
36
Please evaluate this session
Tech Ready 15 6/23/2018 Please evaluate this session From your Please expand notes window at bottom of slide and read. Then Delete this text box. PC or tablet: visit MyIgnite Phone: download and use the Microsoft Ignite mobile app Your input is important! © 2012 Microsoft Corporation. All rights reserved. Microsoft, Windows, Windows Vista 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.
38
Use ML to pick a winner from list
6/23/2018 9:40 PM User types a pattern Send us feedback Expressed as DSL? Use ML to pick a winner from list Generate range of programs which fit example 1 winner Done No Yes © Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
Similar presentations
