# Day 1: MATLAB Platform and Data Analytics ## Introduction and Welcome - Andy Grace Session Q\&A slots: *Enter*: first name, last name, comapny: comment/question Text: (508) 433-3693 R2017b Pre-Release coming week of June 12, 2017 ## MATLAB Keynote - Roy Lurie * MATLAB is designed for engineers and scientists, not a general purpose programming system * MATLAB has apps for interactive exploration, and automated code generation for deployment on production systems * MATLAB makes data science easy and accessible for Engineers and Scientists * MATLAB Strategy Highlights * Live Editor * Good for algorithm exploration and publishing * Not just a place code runs, but a place to tell the story of what is happening * App Designer * Catching up to GUIDE * No comment about sluggish performance for even modest sized apps * Engineering Data Analytics * JSON * Tall arrays * Big data algorithms * Spark (and Hadoop) integration * Deep Learning * Existing deep learning tools are for programmers * MATLAB makes deep learning easy and accessible for Engineers * net = alexnet; * An already trained network readily available for download * R2017b is a major release for Deep Learning * World-Class Prediction Performance * CUA compiled with GPU coder * Much faster than TensorFlow and Caffe * New Algorithms * Image segmentation analysis is easier to do now * MATLAB makes Internet of Things easy and accessible fo Engineers and Scientists * New application called ThingSpeak * Online offerings * MatWorks cloud ecosystem is complete in 2017 * MATLAB Drive * Akin to Microsoft Live, DropBox, Google Drive * MATLAB Online * Available to all named users * Mobile editor ## MATLAB Platform Experience - Loren Dean * MATLAB Platform * 2015 - Architecture * 2016 - Design Points * 2017 - Workflows * Individual Workflow * Personal code * Single discpline * Single desktop or small cluster * Team Workflow * Multidisciplinary work * Engineering domain * Job function (IT, DevOps, ...) * Servers and databases * CI Server * Cloud * Large cluster * Databases * Data marts * Production system * Value Drivers in MATLAB similar to model-based design in Simulink * MathWorks Cloud Framework * Questions revolving around Amazon Web Services (AWS) / Microsoft Azure * Cloud Center is used to spin up a cluser on AWS ### MATLAB Code Development Tools - John Glass * Live Editor * Creating Live Scripts * Equation Editor (R2016b) * Interactive Plots (R2016b/R2017a) * Editable Figures (R2016b/R2017a) * Functions in Scripts (R2016b) * Also available in plain-text editor * Function Assistance (R2017b) * Evaluation is faster, typing is smoother in R2017b * Auto and graphical merge ### MATLAB Language - Dave Foti * Simplifying common programmint tasks * String Arrays (R2016b) * Library of new functions for string manipulation (R2016b) * Use of strings with double quotes (R2017a) * New cell array display (R2017b) * No file left behind (R2017b) * New help tip for adding a path to a file that isn't yet visible * Simplifying class and function authoring * Object property validation (R2017a) * Declaration properties * Function input arguments * Define multiple functions with different argument signatures * Defining code modules * Define interfaces * Not worry about helper functions being on the path * Defining dependencies * Connecting with other langauges * Interfaces * Use libraries from inside MATLAB * Engine * Embed MATLAB as a computation engine * C++ * Building out C++ Engine API in R2017b * HTTP and RESTful web services * jsonencode and jsondecode (R2016b) ### MATLAB Performance, Code Compatibility, and Code Testing - Vadim Teverovsky * Performance Update * R2017b Examples: * Support for AVX2 instructions * Improved performance of small arrays * Imrpoved try/catch performance * Faster MATLAB Startup after reboot on a representative machine * 39 sec (R2016a) -> 24.5 sec (R2016b) -> 14.5 sec (R2017a) * Upgrading to a New Release * Read release notes for potential issues * R2017a * Grouped by category * Can filter by incompatibility * Find these issues in your code (MAB 2016: The most time-consuming step) * R2017b * MATLAB Code Compatibility Report * Detect potential incompatibilities * Figure out how to update your code * Update your code * Test updated code * MATLAB Code Testing * Unit Testing * Unit Testing allows you to lock down the behavior of your code * Scales from simple to large-scale testing * Framework * Easy creation and execution of your tests * Write tests as scripts, functions, or classes * Industry-standard xUnit style * R2017b * Integration into Live Script * Performance Testing * Mocking Framework * R2017a * Isolate a portion of a system to test by imitating behavior of dependencies * App Testing Framework * App Designer apps only ### Scaling MATLAB Applications from Desktop to HPC - Jos Martin * Context for Scalability * Problem Sizes Grow * What to do? * More Machines * More system complexity * Scaling Data * Big data, GPU, Deep Learning * Datatypes * tall / Disks / Pre-processing, statistics, machine learning * distributed / Cluster / Sparse and dense numerics * gpuArray * GPU-Accelerated Deep Learning * R2017b - CPU Coder * Set training options to specify multi-gpu/parallel * Scaling Compute * Standard method: parfor & parfeval * R2017a - Interactive Parallel Programming * Problem: * You have a long running paralle program * Solution: DataQueue * Options * Parallel Computing Toolbox * MATLAB Parallel Cloud * Cloud Center * Virtual Private Cloud Support * GPU Clusters (P2) * Custom Cloud * MATLAB Distributed Computing Cluster * Scaling Simulink Workflows * Running many simulations * Traditional method: parfor + sim * New parsim in R2017a * Setup simulation inputs * Run simulation in parallel using parsim * R2017b * Live streaming of simulation results in Simulation Data Inspector * Simulation Manager ### Deploying MATLAB Code in Embedded Systems - Fred Smith * MATLAB Coder - The Most Interesting App in the World * Generates embeddable C code * ANSI C * Trend * Performance * GPU * SIMD * Multi-core * Algorithms * Machine Learning * Computer Vision * Integration * OpenCV * AUTOSAR * Your code * R2017a * Algorithms: Dynamic memory allocation * Performance: Loop invariant code motion * Performance; Fast FFT * R2017b * Performace: CPU Coder * Utilizes code for fast deep learning performance ### Deploying MATLAB Code for Production Use - Jim Stewart * Application deployment * Share code with other MATLAB users * Create your toolboxes and share with others using Toolbox Packaging * Package toolboxes as a single installer file * Automatic inclusion of dependencies * Share code with non-MATLAB users * Package code as * Standaline executables * Excel add-ins * Big Data applications using MATLAB compiler * Integrate with * Applications written in other languages * MATLAB Compiler and MATLAB Compiler SDK * Deploy App Designer Web Apps on Intranet * Beta program in R2017a/R2017b * Provide a web-app server that can be installed on internal network * Deploy Big Data Applications * Apache Spark (R2016b) * Deploy into production systems with MATLAB Production Server * Deploy updates without server restarts * Downloadable Performance report (R2017a) * IT admins manage software using Dashboard (R2017a) * Operationalize your MATLAB analytics * Availability of Cloud Formation Scripts for AWS and Azure (R2017a) * MATLAB client for MATLAB Production Server * Algorithm and version management ### Harware and Analytics in the IoT Workflow - Rob Purser * Pull it all together * IoT Challenges: * New technologies * Sensor analytics, cloud, machine learning, and more * Multidisciplinary teams * Engineers, solution architects, IT, DevOps, and more * Software in multiple places * Embedded, in the cloud, on desktops, in browser * Steps to pulling it all together * Start with an idea...Explore using MATLAB * Pysics-based models * Collect test data * Data acquisition and instrumentation * Cmaeras * Refine your idea with your team * Transform your idea to algorithm * Prototype with hardware * Code generation from MATLAB and Simulink * Validate with multiple IoT devices * Analytic IoT Platform * ThingSpeak * TMW's IoT analytic platform (R2015a) * Collect * Analyze * Act * Develop production online analytics * Can deploy algorithm to ThingSpeak, instead of running locally * Go to Production or Scale Up * Tools * Predictive Maintenance Toolbox * Sensor Fusion and Tracking Toolbox (R2017b) ## MATLAB Data Analytics - Mary Ann Freeman * Predictive Maintenance * Modeling * First principles modeling * Physical models * Kalman filters * Data-drive modeling * Curve fitting * Machine learning * Generate failure data using Simulink * Features * Access data with readtable, datastore (R2016b) * Preprocessing with timetable, tall (R2017a) * Work with Simulink logged output (R2017a) * Explore graphically (R2017a) * Machine learning models (R2017b) * Deep Learning * Deep learning is a neural network that learns directly from the data * Deep learning model can extract features automatically to classify objects * Deep learning is transforming machine learning * Data * Computer vision, text, and speech * Recognition fields revolutionized * Key projects * Tensorflow, caffe, et al * Issue * Difficult to learn and apply * Deep learning enablers * Massive sets of labeled data * Acceleration with GPUs * Topics to come * Data handling: time-stamped & big data * Deep learning & machine learning * App building & web apps ### Time-Stampled Data Handling with Timetables - Nick Haddad * timetable overview (R2016b) * Used in accessing and exploring & analysis * `table2timetable` * timetable flows seemlessly in functions such as * `rmmissing` * `filloutliers` * `smoothdata` * `synchronize` is used to synchronize data * Import timetable into Simulation Data Inspector (R2017b) * Signal Analyzer App (R2017a) * Comparision with timeseries * timeseries * timetable * timetable Roadmap * MATLAB end-toend workflows * Adoption across products * Collections * Events * Time-based preprocessing * Visulization ### Big Data Handling with Tall Arrays - Penny Anderson * Tall Arrays through R2017b * Overviews * Gives capability to work with data that doesn't otherwise fit in the computing machine's memory * Introduced in R2016b * Ideal for tabular/columnar data * Overall data size is too big to fit into memory * Functions specific to Tall Arrays * Access data with `datastore` and `tall` * Calculat statistics using `gather` * Roadmap for Tall Arrays * `synchronize` will work with Tall Arrays * ThingSpeak * Access: Write your own datastore (R2017b) * Machine Learning with Tall Arrays * R2016b * kmeans * fitlm * fitglm * fitcdiscr * cvpartition * R2017a * fitclinear * fitcnb ### Machine Learning & Deep Learning with MATLAB - Jon Cherrie * Deep Learning with MATLAB * Pre-trained network: `net = alexnet;` * Inspect the network: `net.Layers;` * Topics * Easy to develop systems with machine learning & deep learning * Developing predictive models * Apps * Ground Truth Labeler (R2017a - Automated Driving System Toolbox) * Classification Learner (R2015a - Statistics and Machine Learning Toolbox) * Regression Learner (R2017a - Statics and Machine Learning Toolbox) * Model Optimization * Bayesian Optimization now built into many functions natively * Access to the Latest Research * Reference models * AlexNet, VGG-16, VGG-19 * TensorFlow importer in the works * Caffe model importer * Deployment * Accessiclbe to non-experts * Tutorials - Deep Learning in 11-lines of MATLAB Code * High perfromance embedded implementation * Code generation workflow * Train your model * Implement a wrapper function * Generate code * Support for predictive models * R2017b * fitcdiscr * fitcknn * fitrsvm * fitrensemble * fitrgp ### Text Analytics: Deriving Information from Text Data - Ben Hinkle * New product * Text Analytics Toolbox (R2017b) * `extractFileText` * `eraseTags` * `removeWords` * `bagOfWords` * `predict` * Applications * Document classification * Field reports * Bug reports * Sentiment analysis * Survey data * Trial notes * Social media * Predictive maintenance * Equipment log notes * Combine notes with numerica data ### Deep Learning Deployment on FPGAs, CPUs, and GPUs - Girish Venkataramani * Deep learning workflow * Training * Definte network * Injest data * Deployment * Challenges * Fitting on embedded * Research DNNs are general * Application-specific DNN can be tailored to be much smaller * Re-architect network > weight compression > quantization to FP16, INT8 * Scale performance on the platform of choice * NVIDIA GPU * R2017a * 14x faster than Caffe * 4x faster than TensorFlow * A deep learning (DL) coder infrastructure * GPU Coder > CUDA code (R2017b) ### App Building Advancements and Web Apps - Chris Portal * Feedback (2014-2016) * Components * Parity with GUIDe * More Components + Richer Configurations * Layout Management * App Sharing * Run apps on Mobile and Browser platforms * App Designer * Resources for GUIDE to App Designer migration * Developments * Components * R2017a * Graphics Pan & Zoom * R2017b * 3D Plotting (App Designer) * UIMENU (App Designer) * UITREE (MATLAB Interface) * Graphics * R2017a * Heatmap * R2017b * Tall plot * Moving to the Web * End goal is targeting web-based dashboards and data-centric apps * Extending MATLAB platform across MATLAB Online & MATLAB Drive (R2017b) ### Migrating GUIDE Work to App Designer - Jyoti Gandhe * Resources being developed to automate current workflow * `fig2mlapp()` * Migration workflow * Migrate Components & Callbacks (will be automated) * Lay out components * Configure component properties * Tag components with names * Wire callbacks to callback functions * Copy callback functions over * Update Code (will remain a manual process) * Use App Designer style code * Reconcile component API differences * Run migrated app to validate/troubleshoot * Roadmap * Migration tool will be available as an R2017b add-on ## Breakout Sessions ### MATLAB Testing: Frameworks and Workflows - Andy Campbell * TMW Developer Blog: blogs.mathworks.com * Visit their tech showcase for examples and demo of workflow * Unit test framework * Script-based/ function-based, class-based * TMW encourags exploration of script-based tests * Performance test framework * App test framework * Mocking framework #### Continuous integration - Andy Campbell * Most users in the session are using Jenkins * Go to the blog about continuous integration * Integrations * TAPPlugin (Original format) * XMLPlugin (Junit format) * TAPPlugin (Version 13) * TestReportPlugin (html) * CodeCoveragePlugin (Cobertura output) * Support package for Jenkins * Streamlining online, cloud, etc * Simulink test support * Simulink Coverage #### Typical workflow - Stephen Carter * Write > Run > Analyze * Write * Class-based, function-based, script-based * Live script compatibility (R2017b+) * Artifacts * FigureDiagnostic (R2017a) * ScreenshotDiagnostic (R2017a) * `testCase.onFailure(@() doSomething())` convenience (R2017b) * Run * Formal interface (TestRunner) * Analyze * `results = runner.run(suite); results(1).Details.DiagnosticRecord` (R2016a) * HTML test report (R2017b) #### Extensions & Plugins- David Hrushka * Extensions * Plugin output streams * Constraints & Tolerances * Fixtures * Diagnostics * Test Suite Hierarchy * Data Drive Testing #### Mocking Framework - David Hrushka * Shipping since R2017ag * `testCase.createMock(?Pet);` #### App Designer Test Tools - Steve McClure * Driving your app * Only works with human interaction, not programmatically modifying controls * Programmatic gestures * Press * Choose * Drag * Type * Guiding principles * Support testing interactions for what App-authors can do * Boost your code coverage * Realistic Gestures * Eg, respecting z-ordering * Isolate the App Under Test * App is locked out during tests * Interface Philosophy * Leveare a small set of functions with common and intuitive APIs * Emphasize the App-*User* rather than the App-*Author* #### The MATLAB Performance Testing Framework - Andy Campbell * R2016a * `runperf` * Fixed time experiment * Variable sample time experiment using classical statistics * R2017a * Sample summary * Leverages the new `samplefun` function * R2017b * Multiple measurement points * `testCase.startMeasuring;` * `testCase.stopMeasureing;` ### Upgrading to the Current Release: MATLAB, Simulink, and Code Generation - Narfi Stefansson/Richard Adame * Release Compatibility * Good internal MathWorks processes * Understand and reduce customer impact * Make chages deliberately * Tooling * Understand what changed in a release * Find the affected artifacts (code/models/files/...) * Documentation * Document product changes * Provide information on how to migrate affected artifacts * Test frameworks * Know that all changes have been addressed * Documentation * Release notes can now filter by range of releases * Release notes provide summary * Explain Differences between functions that are no longer recommended * Release Notes are concise * Migration documentation should describe: * Motivation for change * Functionality differences * How to update code/model/... * Should be accessible after original functionality has been removed * White-paper from TMW with all they know about upgrading * Upgrading MATLAB Code * Finding specific places where code changes are needed * Code Compatilibity Report (R2017b) * Analyzes your MATLAB codebase * Performs \~300 checks to locate incompatibilities * Creates report with recommendations * Also has a command-line version for integration with scripts * Simulink Upgrade Advisor * R2016b has a command-line API for automation of upgrade * R2017a can upgrade all models in a Simulink Project * R2017b has better handling of library blocks * Simulink Project Provides Encapsulation * Re-Using Generated Code * Production code integration across releases is a barrier to upgrading * Typical upgrade approaches * Manual integration: Difficult * Duplicate utilities * Non-functional changes * Large number of utilities * Generate new code for all components: costly re-verification * Automatically reuse previously verified code in new releases * Avoids re-verification of components * System-level integration * R2016b * Reuse existing shared code * Avoid generating duplicate utilities * Reduce integration merging * Automatically integrate model code * Import code from previous releases back to R2010a * Generate efficient integrated code * Cross-release code generation * Presentation - Day 3, Thursday Morning, 8:30am * Code Generation and Targeting Session