Google Git
Sign in
apache / beam / a6cfaee1a74b5c48264c7cca7618c3e769c8eec6 / . / website / generated-content / documentation / ml / anomaly-detection / index.html
blob: 7db593afd63f9da4a956324487c1e14c17dbee08 [file] [log] [blame]
<!doctype html><html lang=en class=no-js><head><meta charset=utf-8><meta http-equiv=x-ua-compatible content="IE=edge"><meta name=viewport content="width=device-width,initial-scale=1"><title>Anomaly Detection</title><meta name=description content="Apache Beam is an open source, unified model and set of language-specific SDKs for defining and executing data processing workflows, and also data ingestion and integration flows, supporting Enterprise Integration Patterns (EIPs) and Domain Specific Languages (DSLs). Dataflow pipelines simplify the mechanics of large-scale batch and streaming data processing and can run on a number of runtimes like Apache Flink, Apache Spark, and Google Cloud Dataflow (a cloud service). Beam also brings DSL in different languages, allowing users to easily implement their data integration processes."><link href="https://fonts.googleapis.com/css?family=Roboto:100,300,400,500,700" rel=stylesheet><link rel=preload href=/scss/main.min.408fddfe3e8a45f87a5a8c9a839d77db667c1c534e5e5cd0d957ffc3dd6c14cf.css as=style><link href=/scss/main.min.408fddfe3e8a45f87a5a8c9a839d77db667c1c534e5e5cd0d957ffc3dd6c14cf.css rel=stylesheet integrity><script src=https://code.jquery.com/jquery-2.2.4.min.js></script><style>.body__contained img{max-width:100%}</style><script type=text/javascript src=/js/bootstrap.min.2979f9a6e32fc42c3e7406339ee9fe76b31d1b52059776a02b4a7fa6a4fd280a.js defer></script>
<script type=text/javascript src=/js/language-switch-v2.min.121952b7980b920320ab229551857669209945e39b05ba2b433a565385ca44c6.js defer></script>
<script type=text/javascript src=/js/fix-menu.min.039174b67107465f2090a493f91e126f7aa797f29420f9edab8a54d9dd4b3d2d.js defer></script>
<script type=text/javascript src=/js/section-nav.min.1405fd5e70fab5f6c54037c269b1d137487d8f3d1b3009032525f6db3fbce991.js defer></script>
<script type=text/javascript src=/js/page-nav.min.af231204c9c52c5089d53a4c02739eacbb7f939e3be1c6ffcc212e0ac4dbf879.js defer></script>
<script type=text/javascript src=/js/expandable-list.min.75a4526624a3b8898fe7fb9e3428c205b581f8b38c7926922467aef17eac69f2.js defer></script>
<script type=text/javascript src=/js/copy-to-clipboard.min.364c06423d7e8993fc42bb4abc38c03195bc8386db26d18774ce775d08d5b18d.js defer></script>
<script type=text/javascript src=/js/calendar.min.336664054fa0f52b08bbd4e3c59b5cb6d63dcfb2b4d602839746516b0817446b.js defer></script>
<script type=text/javascript src=/js/fix-playground-nested-scroll.min.0283f1037cb1b9d5074c6eaf041292b524a8148a7cdb803d5ccd6d1fc4eb3253.js defer></script>
<script type=text/javascript src=/js/anchor-content-jump-fix.min.22d3240f81632e4c11179b9d2aaf37a40da9414333c43aa97344e8b21a7df0e4.js defer></script>
<link rel=alternate type=application/rss+xml title="Apache Beam" href=/feed.xml><link rel=canonical href=/documentation/ml/anomaly-detection/ data-proofer-ignore><link rel="shortcut icon" type=image/x-icon href=/images/favicon.ico><link rel=stylesheet href=https://use.fontawesome.com/releases/v5.4.1/css/all.css integrity=sha384-5sAR7xN1Nv6T6+dT2mhtzEpVJvfS3NScPQTrOxhwjIuvcA67KV2R5Jz6kr4abQsz crossorigin=anonymous><link rel=stylesheet href=https://unpkg.com/swiper@8/swiper-bundle.min.css><script async src=https://platform.twitter.com/widgets.js></script>
<script>(function(e,t,n,s,o,i,a){e.GoogleAnalyticsObject=o,e[o]=e[o]||function(){(e[o].q=e[o].q||[]).push(arguments)},e[o].l=1*new Date,i=t.createElement(n),a=t.getElementsByTagName(n)[0],i.async=1,i.src=s,a.parentNode.insertBefore(i,a)})(window,document,"script","//www.google-analytics.com/analytics.js","ga"),ga("create","UA-73650088-1","auto"),ga("send","pageview")</script><script>(function(e,t,n,s,o,i){e.hj=e.hj||function(){(e.hj.q=e.hj.q||[]).push(arguments)},e._hjSettings={hjid:2182187,hjsv:6},o=t.getElementsByTagName("head")[0],i=t.createElement("script"),i.async=1,i.src=n+e._hjSettings.hjid+s+e._hjSettings.hjsv,o.appendChild(i)})(window,document,"https://static.hotjar.com/c/hotjar-",".js?sv=")</script></head><body class=body data-spy=scroll data-target=.page-nav data-offset=0><nav class="navigation-bar-mobile header navbar navbar-fixed-top"><div class=navbar-header><a href=/ class=navbar-brand><img alt=Brand style=height:46px;width:43px src=/images/beam_logo_navbar_mobile.png></a>
<a class=navbar-link href=/get-started/>Get Started</a>
<a class=navbar-link href=/documentation/>Documentation</a>
<button type=button class="navbar-toggle menu-open" aria-expanded=false aria-controls=navbar onclick=openMenu()>
<span class=sr-only>Toggle navigation</span>
<span class=icon-bar></span>
<span class=icon-bar></span>
<span class=icon-bar></span></button></div><div class="navbar-mask closed"></div><div id=navbar class="navbar-container closed"><button type=button class=navbar-toggle aria-expanded=false aria-controls=navbar id=closeMenu>
<span class=sr-only>Toggle navigation</span>
<span class=icon-bar></span>
<span class=icon-bar></span>
<span class=icon-bar></span></button><ul class="nav navbar-nav"><li><div class=searchBar-mobile><script>(function(){var t,n="012923275103528129024:4emlchv9wzi",e=document.createElement("script");e.type="text/javascript",e.async=!0,e.src="https://cse.google.com/cse.js?cx="+n,t=document.getElementsByTagName("script")[0],t.parentNode.insertBefore(e,t)})()</script><gcse:search></gcse:search></div></li><li><a class=navbar-link href=/about>About</a></li><li><a class=navbar-link href=/get-started/>Get Started</a></li><li><span class=navbar-link>Documentation</span><ul><li><a href=/documentation/>General</a></li><li><a href=/documentation/sdks/java/>Languages</a></li><li><a href=/documentation/runners/capability-matrix/>Runners</a></li><li><a href=/documentation/io/connectors/>I/O Connectors</a></li></ul></li><li><a class=navbar-link href=/roadmap/>Roadmap</a></li><li><a class=navbar-link href=/community/>Community</a></li><li><a class=navbar-link href=/contribute/>Contribute</a></li><li><a class=navbar-link href=/blog/>Blog</a></li><li><a class=navbar-link href=/case-studies/>Case Studies</a></li></ul><ul class="nav navbar-nav navbar-right"><li><a href=https://github.com/apache/beam/edit/master/website/www/site/content/en/documentation/ml/anomaly-detection.md data-proofer-ignore><svg xmlns="http://www.w3.org/2000/svg" width="25" height="24" fill="none" viewBox="0 0 25 24"><path stroke="#ff6d00" stroke-linecap="round" stroke-linejoin="round" stroke-width="2.75" d="M4.543 20h4l10.5-10.5c.53-.53.828-1.25.828-2s-.298-1.47-.828-2-1.25-.828-2-.828-1.47.298-2 .828L4.543 16v4zm9.5-13.5 4 4"/></svg></a></li><li class=dropdown><a href=# class=dropdown-toggle id=apache-dropdown data-toggle=dropdown role=button aria-haspopup=true aria-expanded=false><img src=https://www.apache.org/foundation/press/kit/feather_small.png alt="Apache Logo" style=height:20px>
&nbsp;Apache
<span class=arrow-icon><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" fill="none" viewBox="0 0 20 20"><circle cx="10" cy="10" r="10" fill="#ff6d00"/><path stroke="#fff" stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M8.535 5.28l4.573 4.818-4.573 4.403"/></svg></span></a><ul class="dropdown-menu dropdown-menu-right"><li><a target=_blank href=https://www.apache.org/>ASF Homepage</a></li><li><a target=_blank href=https://www.apache.org/licenses/>License</a></li><li><a target=_blank href=https://www.apache.org/security/>Security</a></li><li><a target=_blank href=https://www.apache.org/foundation/thanks.html>Thanks</a></li><li><a target=_blank href=https://www.apache.org/foundation/sponsorship.html>Sponsorship</a></li><li><a target=_blank href=https://www.apache.org/foundation/policies/conduct>Code of Conduct</a></li></ul></li></ul></div></nav><nav class=navigation-bar-desktop><a href=/ class=navbar-logo><img src=/images/beam_logo_navbar.png alt="Beam Logo"></a><div class=navbar-bar-left><div class=navbar-links><a class=navbar-link href=/about>About</a>
<a class=navbar-link href=/get-started/>Get Started</a><li class="dropdown navbar-dropdown navbar-dropdown-documentation"><a href=# class="dropdown-toggle navbar-link" role=button aria-haspopup=true aria-expanded=false>Documentation
<span><svg xmlns="http://www.w3.org/2000/svg" width="12" height="11" fill="none" viewBox="0 0 12 11"><path stroke="#ff6d00" stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M10.666 4.535 5.847 9.108 1.444 4.535"/></svg></span></a><ul class=dropdown-menu><li><a class=navbar-dropdown-menu-link href=/documentation/>General</a></li><li><a class=navbar-dropdown-menu-link href=/documentation/sdks/java/>Languages</a></li><li><a class=navbar-dropdown-menu-link href=/documentation/runners/capability-matrix/>Runners</a></li><li><a class=navbar-dropdown-menu-link href=/documentation/io/connectors/>I/O Connectors</a></li></ul></li><a class=navbar-link href=/roadmap/>Roadmap</a>
<a class=navbar-link href=/community/>Community</a>
<a class=navbar-link href=/contribute/>Contribute</a>
<a class=navbar-link href=/blog/>Blog</a>
<a class=navbar-link href=/case-studies/>Case Studies</a></div><div id=iconsBar><a type=button onclick=showSearch()><svg xmlns="http://www.w3.org/2000/svg" width="25" height="24" fill="none" viewBox="0 0 25 24"><path stroke="#ff6d00" stroke-linecap="round" stroke-linejoin="round" stroke-width="2.75" d="M10.191 17c3.866.0 7-3.134 7-7s-3.134-7-7-7-7 3.134-7 7 3.134 7 7 7zm11 4-6-6"/></svg></a><a target=_blank href=https://github.com/apache/beam/edit/master/website/www/site/content/en/documentation/ml/anomaly-detection.md data-proofer-ignore><svg xmlns="http://www.w3.org/2000/svg" width="25" height="24" fill="none" viewBox="0 0 25 24"><path stroke="#ff6d00" stroke-linecap="round" stroke-linejoin="round" stroke-width="2.75" d="M4.543 20h4l10.5-10.5c.53-.53.828-1.25.828-2s-.298-1.47-.828-2-1.25-.828-2-.828-1.47.298-2 .828L4.543 16v4zm9.5-13.5 4 4"/></svg></a><li class="dropdown navbar-dropdown navbar-dropdown-apache"><a href=# class=dropdown-toggle role=button aria-haspopup=true aria-expanded=false><img src=https://www.apache.org/foundation/press/kit/feather_small.png alt="Apache Logo" style=height:20px>
&nbsp;Apache
<span class=arrow-icon><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" fill="none" viewBox="0 0 20 20"><circle cx="10" cy="10" r="10" fill="#ff6d00"/><path stroke="#fff" stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M8.535 5.28l4.573 4.818-4.573 4.403"/></svg></span></a><ul class=dropdown-menu><li><a class=navbar-dropdown-menu-link target=_blank href=https://www.apache.org/>ASF Homepage</a></li><li><a class=navbar-dropdown-menu-link target=_blank href=https://www.apache.org/licenses/>License</a></li><li><a class=navbar-dropdown-menu-link target=_blank href=https://www.apache.org/security/>Security</a></li><li><a class=navbar-dropdown-menu-link target=_blank href=https://www.apache.org/foundation/thanks.html>Thanks</a></li><li><a class=navbar-dropdown-menu-link target=_blank href=https://www.apache.org/foundation/sponsorship.html>Sponsorship</a></li><li><a class=navbar-dropdown-menu-link target=_blank href=https://www.apache.org/foundation/policies/conduct>Code of Conduct</a></li></ul></li></div><div class="searchBar disappear"><script>(function(){var t,n="012923275103528129024:4emlchv9wzi",e=document.createElement("script");e.type="text/javascript",e.async=!0,e.src="https://cse.google.com/cse.js?cx="+n,t=document.getElementsByTagName("script")[0],t.parentNode.insertBefore(e,t)})()</script><gcse:search></gcse:search>
<a type=button onclick=endSearch()><svg xmlns="http://www.w3.org/2000/svg" width="25" height="25" fill="none" viewBox="0 0 25 25"><path stroke="#ff6d00" stroke-linecap="round" stroke-linejoin="round" stroke-width="2.75" d="M21.122 20.827 4.727 4.432M21.122 4.43 4.727 20.827"/></svg></a></div></div></nav><div class=header-push></div><div class="top-banners swiper"><div class=swiper-wrapper><div class=swiper-slide><a href=https://tour.beam.apache.org><img class=banner-img-desktop src=/images/banners/tour-of-beam/tour-of-beam-desktop.png alt="Start Tour of Beam">
<img class=banner-img-mobile src=/images/banners/tour-of-beam/tour-of-beam-mobile.png alt="Start Tour of Beam"></a></div><div class=swiper-slide><a href=https://beam.apache.org/documentation/ml/overview/><img class=banner-img-desktop src=/images/banners/machine-learning/machine-learning-desktop.jpg alt="Machine Learning">
<img class=banner-img-mobile src=/images/banners/machine-learning/machine-learning-mobile.jpg alt="Machine Learning"></a></div></div><div class=swiper-pagination></div><div class=swiper-button-prev></div><div class=swiper-button-next></div></div><script src=/js/swiper-bundle.min.min.e0e8f81b0b15728d35ff73c07f42ddbb17a108d6f23df4953cb3e60df7ade675.js></script>
<script src=/js/sliders/top-banners.min.afa7d0a19acf7a3b28ca369490b3d401a619562a2a4c9612577be2f66a4b9855.js></script>
<script>function showSearch(){addPlaceholder();var e,t=document.querySelector(".searchBar");t.classList.remove("disappear"),e=document.querySelector("#iconsBar"),e.classList.add("disappear")}function addPlaceholder(){$("input:text").attr("placeholder","What are you looking for?")}function endSearch(){var e,t=document.querySelector(".searchBar");t.classList.add("disappear"),e=document.querySelector("#iconsBar"),e.classList.remove("disappear")}function blockScroll(){$("body").toggleClass("fixedPosition")}function openMenu(){addPlaceholder(),blockScroll()}</script><div class="clearfix container-main-content"><div class="section-nav closed" data-offset-top=90 data-offset-bottom=500><span class="section-nav-back glyphicon glyphicon-menu-left"></span><nav><ul class=section-nav-list data-section-nav><li><span class=section-nav-list-main-title>Documentation</span></li><li><a href=/documentation>Using the Documentation</a></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Concepts</span><ul class=section-nav-list><li><a href=/documentation/basics/>Basics of the Beam model</a></li><li><a href=/documentation/runtime/model/>How Beam executes a pipeline</a></li></ul></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Beam programming guide</span><ul class=section-nav-list><li><a href=/documentation/programming-guide/>Overview</a></li><li><a href=/documentation/programming-guide/#creating-a-pipeline>Pipelines</a></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>PCollections</span><ul class=section-nav-list><li><a href=/documentation/programming-guide/#pcollections>Creating a PCollection</a></li><li><a href=/documentation/programming-guide/#pcollection-characteristics>PCollection characteristics</a></li></ul></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Transforms</span><ul class=section-nav-list><li><a href=/documentation/programming-guide/#applying-transforms>Applying transforms</a></li><li><span class=section-nav-list-title>Core Beam transforms</span><ul class=section-nav-list><li><a href=/documentation/programming-guide/#pardo>ParDo</a></li><li><a href=/documentation/programming-guide/#groupbykey>GroupByKey</a></li><li><a href=/documentation/programming-guide/#cogroupbykey>CoGroupByKey</a></li><li><a href=/documentation/programming-guide/#combine>Combine</a></li><li><a href=/documentation/programming-guide/#flatten>Flatten</a></li><li><a href=/documentation/programming-guide/#partition>Partition</a></li></ul></li><li><a href=/documentation/programming-guide/#requirements-for-writing-user-code-for-beam-transforms>Requirements for user code</a></li><li><a href=/documentation/programming-guide/#side-inputs>Side inputs</a></li><li><a href=/documentation/programming-guide/#additional-outputs>Additional outputs</a></li><li><a href=/documentation/programming-guide/#composite-transforms>Composite transforms</a></li></ul></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Pipeline I/O</span><ul class=section-nav-list><li><a href=/documentation/programming-guide/#pipeline-io>Using I/O transforms</a></li><li><a href=/documentation/io/connectors/>I/O connectors</a></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>I/O connector guides</span><ul class=section-nav-list><li><a href=/documentation/io/built-in/parquet/>Apache Parquet I/O connector</a></li><li><a href=/documentation/io/built-in/hadoop/>Hadoop Input/Output Format IO</a></li><li><a href=/documentation/io/built-in/hcatalog/>HCatalog IO</a></li><li><a href=/documentation/io/built-in/google-bigquery/>Google BigQuery I/O connector</a></li><li><a href=/documentation/io/built-in/snowflake/>Snowflake I/O connector</a></li><li><a href=/documentation/io/built-in/cdap/>CDAP I/O connector</a></li><li><a href=/documentation/io/built-in/sparkreceiver/>Spark Receiver I/O connector</a></li><li><a href=/documentation/io/built-in/singlestore/>SingleStoreDB I/O connector</a></li><li><a href=/documentation/io/built-in/webapis/>Web APIs I/O connector</a></li></ul></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Developing new I/O connectors</span><ul class=section-nav-list><li><a href=/documentation/io/developing-io-overview/>Overview: Developing connectors</a></li><li><a href=/documentation/io/developing-io-java/>Developing connectors (Java)</a></li><li><a href=/documentation/io/developing-io-python/>Developing connectors (Python)</a></li><li><a href=/documentation/io/io-standards/>I/O Standards</a></li></ul></li><li><a href=/documentation/io/testing/>Testing I/O transforms</a></li></ul></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Schemas</span><ul class=section-nav-list><li><a href=/documentation/programming-guide/#what-is-a-schema>What is a schema</a></li><li><a href=/documentation/programming-guide/#schemas-for-pl-types>Schemas for programming language types</a></li><li><a href=/documentation/programming-guide/#schema-definition>Schema definition</a></li><li><a href=/documentation/programming-guide/#logical-types>Logical types</a></li><li><a href=/documentation/programming-guide/#creating-schemas>Creating schemas</a></li><li><a href=/documentation/programming-guide/#using-schemas>Using schemas</a></li></ul></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Data encoding and type safety</span><ul class=section-nav-list><li><a href=/documentation/programming-guide/#data-encoding-and-type-safety>Data encoding basics</a></li><li><a href=/documentation/programming-guide/#specifying-coders>Specifying coders</a></li><li><a href=/documentation/programming-guide/#default-coders-and-the-coderregistry>Default coders and the CoderRegistry</a></li></ul></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Windowing</span><ul class=section-nav-list><li><a href=/documentation/programming-guide/#windowing>Windowing basics</a></li><li><a href=/documentation/programming-guide/#provided-windowing-functions>Provided windowing functions</a></li><li><a href=/documentation/programming-guide/#setting-your-pcollections-windowing-function>Setting your PCollection’s windowing function</a></li><li><a href=/documentation/programming-guide/#watermarks-and-late-data>Watermarks and late data</a></li><li><a href=/documentation/programming-guide/#adding-timestamps-to-a-pcollections-elements>Adding timestamps to a PCollection’s elements</a></li></ul></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Triggers</span><ul class=section-nav-list><li><a href=/documentation/programming-guide/#triggers>Trigger basics</a></li><li><a href=/documentation/programming-guide/#event-time-triggers>Event time triggers and the default trigger</a></li><li><a href=/documentation/programming-guide/#processing-time-triggers>Processing time triggers</a></li><li><a href=/documentation/programming-guide/#data-driven-triggers>Data-driven triggers</a></li><li><a href=/documentation/programming-guide/#setting-a-trigger>Setting a trigger</a></li><li><a href=/documentation/programming-guide/#composite-triggers>Composite triggers</a></li></ul></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Metrics</span><ul class=section-nav-list><li><a href=/documentation/programming-guide/#metrics>Metrics basics</a></li><li><a href=/documentation/programming-guide/#types-of-metrics>Types of metrics</a></li><li><a href=/documentation/programming-guide/#querying-metrics>Querying metrics</a></li><li><a href=/documentation/programming-guide/#using-metrics>Using metrics in pipeline</a></li><li><a href=/documentation/programming-guide/#export-metrics>Export metrics</a></li></ul></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>State and Timers</span><ul class=section-nav-list><li><a href=/documentation/programming-guide/#types-of-state>Types of state</a></li><li><a href=/documentation/programming-guide/#deferred-state-reads>Deferred state reads</a></li><li><a href=/documentation/programming-guide/#timers>Timers</a></li><li><a href=/documentation/programming-guide/#garbage-collecting-state>Garbage collecting state</a></li><li><a href=/documentation/programming-guide/#state-timers-examples>State and timers examples</a></li></ul></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Splittable DoFns</span><ul class=section-nav-list><li><a href=/documentation/programming-guide/#sdf-basics>Basics</a></li><li><a href=/documentation/programming-guide/#sizing-and-progress>Sizing and progress</a></li><li><a href=/documentation/programming-guide/#user-initiated-checkpoint>User-initiated checkpoint</a></li><li><a href=/documentation/programming-guide/#runner-initiated-split>Runner initiated split</a></li><li><a href=/documentation/programming-guide/#watermark-estimation>Watermark estimation</a></li><li><a href=/documentation/programming-guide/#truncating-during-drain>Truncating during drain</a></li><li><a href=/documentation/programming-guide/#bundle-finalization>Bundle finalization</a></li></ul></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Multi-language Pipelines</span><ul class=section-nav-list><li><a href=/documentation/programming-guide/#create-x-lang-transforms>Creating cross-language transforms</a></li><li><a href=/documentation/programming-guide/#use-x-lang-transforms>Using cross-language transforms</a></li><li><a href=/documentation/programming-guide/#x-lang-transform-runner-support>Runner Support</a></li></ul></li><li><a href=/documentation/programming-guide/#batched-dofns>Batched DoFns</a></li><li><a href=/documentation/programming-guide/#transform-service>Transform service</a></li></ul></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Pipeline development lifecycle</span><ul class=section-nav-list><li><a href=/documentation/pipelines/design-your-pipeline/>Design Your Pipeline</a></li><li><a href=/documentation/pipelines/create-your-pipeline/>Create Your Pipeline</a></li><li><a href=/documentation/pipelines/test-your-pipeline/>Test Your Pipeline</a></li></ul></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Common pipeline patterns</span><ul class=section-nav-list><li><a href=/documentation/patterns/overview/>Overview</a></li><li><a href=/documentation/patterns/file-processing/>File processing</a></li><li><a href=/documentation/patterns/side-inputs/>Side inputs</a></li><li><a href=/documentation/patterns/pipeline-options/>Pipeline options</a></li><li><a href=/documentation/patterns/custom-io/>Custom I/O</a></li><li><a href=/documentation/patterns/custom-windows/>Custom windows</a></li><li><a href=/documentation/patterns/bigqueryio/>BigQueryIO</a></li><li><a href=/documentation/patterns/ai-platform/>AI Platform</a></li><li><a href=/documentation/patterns/schema/>Schema</a></li><li><a href=/documentation/patterns/bqml/>BigQuery ML</a></li><li><a href=/documentation/patterns/grouping-elements-for-efficient-external-service-calls/>Grouping elements for efficient external service calls</a></li></ul></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>AI/ML pipelines</span><ul class=section-nav-list><li><a href=/documentation/ml/overview/>Get started with AI/ML</a></li><li><a href=/documentation/ml/about-ml/>About Beam ML</a></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Prediction and inference</span><ul class=section-nav-list><li><a href=/documentation/ml/inference-overview/>Overview</a></li><li><a href=/documentation/ml/multi-model-pipelines/>Build a pipeline with multiple models</a></li><li><a href=/documentation/ml/tensorrt-runinference>Build a custom model handler with TensorRT</a></li><li><a href=/documentation/ml/large-language-modeling>Use LLM inference</a></li><li><a href=/documentation/ml/multi-language-inference/>Build a multi-language inference pipeline</a></li><li><a href=/documentation/ml/side-input-updates/>Update your model in production</a></li></ul></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Data processing</span><ul class=section-nav-list><li><a href=/documentation/ml/preprocess-data/>Preprocess data</a></li><li><a href=/documentation/ml/data-processing/>Explore your data</a></li></ul></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Workflow orchestration</span><ul class=section-nav-list><li><a href=/documentation/ml/orchestration/>Use ML-OPS workflow orchestrators</a></li></ul></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Model training</span><ul class=section-nav-list><li><a href=/documentation/ml/per-entity-training>Per-entity training</a></li><li><a href=/documentation/ml/online-clustering/>Online clustering</a></li><li><a href=/documentation/ml/model-evaluation/>ML model evaluation</a></li></ul></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Use cases</span><ul class=section-nav-list><li><a href=/documentation/ml/anomaly-detection/>Build an anomaly detection pipeline</a></li></ul></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Reference</span><ul class=section-nav-list><li><a href=/documentation/ml/runinference-metrics/>RunInference metrics</a></li><li><a href=/documentation/ml/model-evaluation/>Model validation</a></li></ul></li></ul></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Runtime systems</span><ul class=section-nav-list><li><a href=/documentation/runtime/environments/>Container environments</a></li><li><a href=/documentation/runtime/resource-hints/>Resource hints</a></li><li><a href=/documentation/runtime/sdk-harness-config/>SDK Harness Configuration</a></li></ul></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Transform catalog</span><ul class=section-nav-list><li class=section-nav-item--collapsible><span class=section-nav-list-title>Python</span><ul class=section-nav-list><li><a href=/documentation/transforms/python/overview/>Overview</a></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Element-wise</span><ul class=section-nav-list><li class=section-nav-item--collapsible><span class=section-nav-list-title>Enrichment</span><ul class=section-nav-list><li><a href=/documentation/transforms/python/elementwise/enrichment/>Overview</a></li><li><a href=/documentation/transforms/python/elementwise/enrichment-bigtable/>Bigtable example</a></li><li><a href=/documentation/transforms/python/elementwise/enrichment-vertexai/>Vertex AI Feature Store examples</a></li></ul></li><li><a href=/documentation/transforms/python/elementwise/filter/>Filter</a></li><li><a href=/documentation/transforms/python/elementwise/flatmap/>FlatMap</a></li><li><a href=/documentation/transforms/python/elementwise/keys/>Keys</a></li><li><a href=/documentation/transforms/python/elementwise/kvswap/>KvSwap</a></li><li><a href=/documentation/transforms/python/elementwise/map/>Map</a></li><li><a href=/documentation/transforms/python/elementwise/mltransform/>MLTransform</a></li><li><a href=/documentation/transforms/python/elementwise/pardo/>ParDo</a></li><li><a href=/documentation/transforms/python/elementwise/partition/>Partition</a></li><li><a href=/documentation/transforms/python/elementwise/regex/>Regex</a></li><li><a href=/documentation/transforms/python/elementwise/reify/>Reify</a></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>RunInference</span><ul class=section-nav-list><li><a href=/documentation/transforms/python/elementwise/runinference/>Overview</a></li><li><a href=/documentation/transforms/python/elementwise/runinference-pytorch/>PyTorch examples</a></li><li><a href=/documentation/transforms/python/elementwise/runinference-sklearn/>Sklearn examples</a></li></ul></li><li><a href=/documentation/transforms/python/elementwise/tostring/>ToString</a></li><li><a href=/documentation/transforms/python/elementwise/values/>Values</a></li><li><a href=/documentation/transforms/python/elementwise/withtimestamps/>WithTimestamps</a></li></ul></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Aggregation</span><ul class=section-nav-list><li><a href=/documentation/transforms/python/aggregation/approximatequantiles/>ApproximateQuantiles</a></li><li><a href=/documentation/transforms/python/aggregation/approximateunique/>ApproximateUnique</a></li><li><a href=/documentation/transforms/python/aggregation/cogroupbykey/>CoGroupByKey</a></li><li><a href=/documentation/transforms/python/aggregation/combineglobally/>CombineGlobally</a></li><li><a href=/documentation/transforms/python/aggregation/combineperkey/>CombinePerKey</a></li><li><a href=/documentation/transforms/python/aggregation/combinevalues/>CombineValues</a></li><li><a href=/documentation/transforms/python/aggregation/count/>Count</a></li><li><a href=/documentation/transforms/python/aggregation/distinct/>Distinct</a></li><li><a href=/documentation/transforms/python/aggregation/groupby/>GroupBy</a></li><li><a href=/documentation/transforms/python/aggregation/groupbykey/>GroupByKey</a></li><li><a href=/documentation/transforms/python/aggregation/groupintobatches/>GroupIntoBatches</a></li><li><a href=/documentation/transforms/python/aggregation/latest/>Latest</a></li><li><a href=/documentation/transforms/python/aggregation/max/>Max</a></li><li><a href=/documentation/transforms/python/aggregation/mean/>Mean</a></li><li><a href=/documentation/transforms/python/aggregation/min/>Min</a></li><li><a href=/documentation/transforms/python/aggregation/sample/>Sample</a></li><li><a href=/documentation/transforms/python/aggregation/sum/>Sum</a></li><li><a href=/documentation/transforms/python/aggregation/top/>Top</a></li><li><a href=/documentation/transforms/python/aggregation/tolist/>ToList</a></li></ul></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Other</span><ul class=section-nav-list><li><a href=/documentation/transforms/python/other/create/>Create</a></li><li><a href=/documentation/transforms/python/other/flatten/>Flatten</a></li><li><a href=/documentation/transforms/python/other/reshuffle/>Reshuffle</a></li><li><a href=/documentation/transforms/python/other/windowinto/>WindowInto</a></li></ul></li></ul></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Java</span><ul class=section-nav-list><li><a href=/documentation/transforms/java/overview/>Overview</a></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Element-wise</span><ul class=section-nav-list><li><a href=/documentation/transforms/java/elementwise/filter/>Filter</a></li><li><a href=/documentation/transforms/java/elementwise/flatmapelements/>FlatMapElements</a></li><li><a href=/documentation/transforms/java/elementwise/keys/>Keys</a></li><li><a href=/documentation/transforms/java/elementwise/kvswap/>KvSwap</a></li><li><a href=/documentation/transforms/java/elementwise/mapelements/>MapElements</a></li><li><a href=/documentation/transforms/java/elementwise/pardo/>ParDo</a></li><li><a href=/documentation/transforms/java/elementwise/partition/>Partition</a></li><li><a href=/documentation/transforms/java/elementwise/regex/>Regex</a></li><li><a href=/documentation/transforms/java/elementwise/reify/>Reify</a></li><li><a href=/documentation/transforms/java/elementwise/tostring/>ToString</a></li><li><a href=/documentation/transforms/java/elementwise/values/>Values</a></li><li><a href=/documentation/transforms/java/elementwise/withkeys/>WithKeys</a></li><li><a href=/documentation/transforms/java/elementwise/withtimestamps/>WithTimestamps</a></li></ul></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Aggregation</span><ul class=section-nav-list><li><a href=/documentation/transforms/java/aggregation/approximatequantiles/>ApproximateQuantiles</a></li><li><a href=/documentation/transforms/java/aggregation/approximateunique/>ApproximateUnique</a></li><li><a href=/documentation/transforms/java/aggregation/cogroupbykey/>CoGroupByKey</a></li><li><a href=/documentation/transforms/java/aggregation/combine/>Combine</a></li><li><a href=/documentation/transforms/java/aggregation/combinewithcontext/>CombineWithContext</a></li><li><a href=/documentation/transforms/java/aggregation/count/>Count</a></li><li><a href=/documentation/transforms/java/aggregation/distinct/>Distinct</a></li><li><a href=/documentation/transforms/java/aggregation/groupbykey/>GroupByKey</a></li><li><a href=/documentation/transforms/java/aggregation/groupintobatches/>GroupIntoBatches</a></li><li><a href=/documentation/transforms/java/aggregation/hllcount/>HllCount</a></li><li><a href=/documentation/transforms/java/aggregation/latest/>Latest</a></li><li><a href=/documentation/transforms/java/aggregation/max/>Max</a></li><li><a href=/documentation/transforms/java/aggregation/mean/>Mean</a></li><li><a href=/documentation/transforms/java/aggregation/min/>Min</a></li><li><a href=/documentation/transforms/java/aggregation/sample/>Sample</a></li><li><a href=/documentation/transforms/java/aggregation/sum/>Sum</a></li><li><a href=/documentation/transforms/java/aggregation/top/>Top</a></li></ul></li><li class=section-nav-item--collapsible><span class=section-nav-list-title>Other</span><ul class=section-nav-list><li><a href=/documentation/transforms/java/other/create/>Create</a></li><li><a href=/documentation/transforms/java/other/flatten/>Flatten</a></li><li><a href=/documentation/transforms/java/other/passert/>PAssert</a></li><li><a href=/documentation/transforms/java/other/view/>View</a></li><li><a href=/documentation/transforms/java/other/window/>Window</a></li></ul></li></ul></li></ul></li><li><a href=/documentation/glossary/>Glossary</a></li><li><a href=https://cwiki.apache.org/confluence/display/BEAM/Apache+Beam>Beam Wiki <img src=/images/external-link-icon.png width=14 height=14 alt="External link."></a></li></ul></nav></div><nav class="page-nav clearfix" data-offset-top=90 data-offset-bottom=500><nav id=TableOfContents><ul><li><a href=#dataset-for-anomaly-detection>Dataset for Anomaly Detection</a></li><li><a href=#anomaly-detection-algorithm>Anomaly Detection Algorithm</a></li><li><a href=#ingestion-to-pubsub>Ingestion to Pub/Sub</a><ul><li><a href=#run-the-pipeline>Run the Pipeline</a></li></ul></li><li><a href=#anomaly-detection-on-streaming-data>Anomaly Detection on Streaming Data</a><ul><li><a href=#run-the-pipeline-1>Run the Pipeline</a></li><li><a href=#get-embedding-from-a-language-model>Get Embedding from a Language Model</a></li><li><a href=#get-predictions>Get Predictions</a></li></ul></li></ul></nav></nav><div class="body__contained body__section-nav arrow-list arrow-list--no-mt"><h1 id=anomaly-detection-example>Anomaly Detection Example</h1><p>The anomaly detection example demonstrates how to set up an anomaly detection pipeline that reads text from Pub/Sub in real time, and then detects anomalies using a trained HDBSCAN clustering model.</p><h2 id=dataset-for-anomaly-detection>Dataset for Anomaly Detection</h2><p>This example uses a dataset called <a href=https://huggingface.co/datasets/emotion>emotion</a> that contains 20,000 English Twitter messages with 6 basic emotions: anger, fear, joy, love, sadness, and surprise. The dataset has three splits: train (for training), validation, and test (for performance evaluation). Because it contains the text and the category (class) of the dataset, it is a supervised dataset. You can use the <a href=https://huggingface.co/docs/datasets/index>Hugging Face datasets page</a> to access this dataset.</p><p>The following text shows examples from the train split of the dataset:</p><table><thead><tr><th style=text-align:left>Text</th><th style=text-align:center>Type of emotion</th></tr></thead><tbody><tr><td style=text-align:left>im grabbing a minute to post i feel greedy wrong</td><td style=text-align:center>Anger</td></tr><tr><td style=text-align:left>i am ever feeling nostalgic about the fireplace i will know that it is still on the property</td><td style=text-align:center>Love</td></tr><tr><td style=text-align:left>ive been taking or milligrams or times recommended amount and ive fallen asleep a lot faster but i also feel like so funny</td><td style=text-align:center>Fear</td></tr><tr><td style=text-align:left>on a boat trip to denmark</td><td style=text-align:center>Joy</td></tr><tr><td style=text-align:left>i feel you know basically like a fake in the realm of science fiction</td><td style=text-align:center>Sadness</td></tr><tr><td style=text-align:left>i began having them several times a week feeling tortured by the hallucinations moving people and figures sounds and vibrations</td><td style=text-align:center>Fear</td></tr></tbody></table><h2 id=anomaly-detection-algorithm>Anomaly Detection Algorithm</h2><p><a href=https://hdbscan.readthedocs.io/en/latest/how_hdbscan_works.html>HDBSCAN</a> is a clustering algorithm that extends DBSCAN by converting it into a hierarchical clustering algorithm and then extracting a flat clustering based in the stability of clusters. When trained, the model predicts <code>-1</code> if a new data point is an outlier, otherwise it predicts one of the existing clusters.</p><h2 id=ingestion-to-pubsub>Ingestion to Pub/Sub</h2><p>Ingest the data into <a href=https://cloud.google.com/pubsub/docs/overview>Pub/Sub</a> so that while clustering, the model can read the tweets from Pub/Sub. Pub/Sub is a messaging service for exchanging event data among applications and services. Streaming analytics and data integration pipelines use Pub/Sub to ingest and distribute data.</p><p>You can see the full example code for ingesting data into Pub/Sub in <a href=https://github.com/apache/beam/tree/master/sdks/python/apache_beam/examples/inference/anomaly_detection/write_data_to_pubsub_pipeline/>GitHub</a></p><p>The file structure for the ingestion pipeline is shown in the following diagram:</p><pre><code>write_data_to_pubsub_pipeline/
├── pipeline/
│ ├── __init__.py
│ ├── options.py
│ └── utils.py
├── __init__.py
├── config.py
├── main.py
└── setup.py
</code></pre><p><code>pipeline/utils.py</code> contains the code for loading the emotion dataset and two <code>beam.DoFn</code> that are used for data transformation.</p><p><code>pipeline/options.py</code> contains the pipeline options to configure the Dataflow pipeline.</p><p><code>config.py</code> defines variables that are used multiple times, like Google Cloud PROJECT_ID and NUM_WORKERS.</p><p><code>setup.py</code> defines the packages and requirements for the pipeline to run.</p><p><code>main.py</code> contains the pipeline code and additional functions used for running the pipeline.</p><h3 id=run-the-pipeline>Run the Pipeline</h3><p>To run the pipeline, install the required packages.For this example, you need access to a Google Cloud project, and you need to configure the Google Cloud variables, like <code>PROJECT_ID</code>, <code>REGION</code>, <code>PubSub TOPIC_ID</code>, and others in the <code>config.py</code> file.</p><ol><li>Locally on your machine: <code>python main.py</code></li><li>On GCP for Dataflow: <code>python main.py --mode cloud</code></li></ol><p>The <code>write_data_to_pubsub_pipeline</code> contains four different transforms:</p><ol><li>Load the emotion dataset using Hugging Face datasets (for simplicity, we take samples from three classes instead of six).</li><li>Associate each piece of text with a unique identifier (UID).</li><li>Convert the text into the format that Pub/Sub expects.</li><li>Write the formatted message to Pub/Sub.</li></ol><h2 id=anomaly-detection-on-streaming-data>Anomaly Detection on Streaming Data</h2><p>After ingesting the data to Pub/Sub, run the anomaly detection pipeline. This pipeline reads the streaming message from Pub/Sub, converts the text to an embedding using a language model, and feeds the embedding to an already trained clustering model to predict whether the message is an anomaly. One prerequisite for this pipeline is to have an HDBSCAN clustering model trained on the training split of the dataset.</p><p>You can find the full example code for anomaly detection in <a href=https://github.com/apache/beam/tree/master/sdks/python/apache_beam/examples/inference/anomaly_detection/anomaly_detection_pipeline/>GitHub</a></p><p>The following diagram shows the file structure for the anomaly_detection pipeline:</p><pre><code>anomaly_detection_pipeline/
├── pipeline/
│ ├── __init__.py
│ ├── options.py
│ └── transformations.py
├── __init__.py
├── config.py
├── main.py
└── setup.py
</code></pre><p><code>pipeline/transformations.py</code> contains the code for different <code>beam.DoFn</code> and additional functions that are used in pipeline.</p><p><code>pipeline/options.py</code> contains the pipeline options to configure the Dataflow pipeline.</p><p><code>config.py</code> defines variables that are used multiple times, like the Google Cloud PROJECT_ID and NUM_WORKERS.</p><p><code>setup.py</code> defines the packages and requirements for the pipeline to run.</p><p><code>main.py</code> contains the pipeline code and additional functions used to run the pipeline.</p><h3 id=run-the-pipeline-1>Run the Pipeline</h3><p>Install the required packages and push the data to Pub/Sub. For this example, you need access to a Google Cloud project, and you need to configure the Google Cloud variables, like <code>PROJECT_ID</code>, <code>REGION</code>, <code>PubSub SUBSCRIPTION_ID</code>, and others in the <code>config.py</code> file.</p><ol><li>Locally on your machine: <code>python main.py</code></li><li>On GCP for Dataflow: <code>python main.py --mode cloud</code></li></ol><p>The pipeline includes the following steps:</p><ol><li>Read the message from Pub/Sub.</li><li>Convert the Pub/Sub message into a <code>PCollection</code> of dictionaries where the key is the UID and the value is the Twitter text.</li><li>Encode the text into transformer-readable token ID integers using a tokenizer.</li><li>Use RunInference to get the vector embedding from a transformer-based language model.</li><li>Normalize the embedding.</li><li>Use RunInference to get anomaly prediction from a trained HDBSCAN clustering model.</li><li>Write the prediction to BigQuery so that the clustering model can be retrained when needed.</li><li>Send an email alert if an anomaly is detected.</li></ol><p>The following code snippet shows the first two steps of the pipeline:</p><div class=snippet><div class="notebook-skip code-snippet without_switcher"><a class=copy type=button data-bs-toggle=tooltip data-bs-placement=bottom title="Copy to clipboard"><img src=/images/copy-icon.svg></a><pre tabindex=0><code> docs = (
pipeline
| &#34;Read from PubSub&#34;
&gt;&gt; ReadFromPubSub(subscription=cfg.SUBSCRIPTION_ID, with_attributes=True)
| &#34;Decode PubSubMessage&#34; &gt;&gt; beam.ParDo(Decode())
)</code></pre></div></div><p>The next section describes the following pipeline steps:</p><ul><li>Tokenizing the text</li><li>Getting embedding using RunInference</li><li>Getting predictions from the HDBSCAN model</li></ul><h3 id=get-embedding-from-a-language-model>Get Embedding from a Language Model</h3><p>In order to do clustering with text data, first map the text into vectors of numerical values suitable for statistical analysis. This example uses a transformer-based language model called <a href=https://huggingface.co/sentence-transformers/stsb-distilbert-base>sentence-transformers/stsb-distilbert-base/stsb-distilbert-base</a>. This model maps sentences and paragraphs to a 768 dimensional dense vector space, and you can use it for tasks like clustering or semantic search.</p><p>Because the language model is expecting a tokenized input instead of raw text, start by tokenizing the text. Tokenization is a preprocessing task that transforms text so that it can be fed into the model for getting predictions.</p><div class=snippet><div class="notebook-skip code-snippet without_switcher"><a class=copy type=button data-bs-toggle=tooltip data-bs-placement=bottom title="Copy to clipboard"><img src=/images/copy-icon.svg></a><pre tabindex=0><code> normalized_embedding = (
docs
| &#34;Tokenize Text&#34; &gt;&gt; beam.Map(tokenize_sentence)</code></pre></div></div><p>Here, <code>tokenize_sentence</code> is a function that takes a dictionary with a text and an ID, tokenizes the text, and returns a tuple of the text and ID as well as the tokenized output.</p><div class=snippet><div class="notebook-skip code-snippet without_switcher"><a class=copy type=button data-bs-toggle=tooltip data-bs-placement=bottom title="Copy to clipboard"><img src=/images/copy-icon.svg></a><pre tabindex=0><code>Tokenizer = AutoTokenizer.from_pretrained(cfg.TOKENIZER_NAME)
def tokenize_sentence(input_dict):
&#34;&#34;&#34;
Takes a dictionary with a text and an id, tokenizes the text, and
returns a tuple of the text and id and the tokenized text
Args:
input_dict: a dictionary with the text and id of the sentence
Returns:
A tuple of the text and id, and a dictionary of the tokens.
&#34;&#34;&#34;
text, uid = input_dict[&#34;text&#34;], input_dict[&#34;id&#34;]
tokens = Tokenizer([text], padding=True, truncation=True, return_tensors=&#34;pt&#34;)
tokens = {key: torch.squeeze(val) for key, val in tokens.items()}
return (text, uid), tokens</code></pre></div></div><p>Tokenized output is then passed to the language model to get the embeddings. To get embeddings from the language model, we use <code>RunInference()</code> from Apache Beam.</p><p><div class=snippet><div class="notebook-skip code-snippet without_switcher"><a class=copy type=button data-bs-toggle=tooltip data-bs-placement=bottom title="Copy to clipboard"><img src=/images/copy-icon.svg></a><pre tabindex=0><code> | &#34;Get Embedding&#34; &gt;&gt; RunInference(KeyedModelHandler(embedding_model_handler))</code></pre></div></div>where <code>embedding_model_handler</code> is:</p><div class=snippet><div class="notebook-skip code-snippet without_switcher"><a class=copy type=button data-bs-toggle=tooltip data-bs-placement=bottom title="Copy to clipboard"><img src=/images/copy-icon.svg></a><pre tabindex=0><code> embedding_model_handler = PytorchNoBatchModelHandler(
state_dict_path=cfg.MODEL_STATE_DICT_PATH,
model_class=ModelWrapper,
model_params={&#34;config&#34;: AutoConfig.from_pretrained(cfg.MODEL_CONFIG_PATH)},
device=&#34;cpu&#34;,
)</code></pre></div></div><p>We define <code>PytorchNoBatchModelHandler</code> as a wrapper to <code>PytorchModelHandler</code> to limit batch size to one.</p><div class=snippet><div class="notebook-skip code-snippet without_switcher"><a class=copy type=button data-bs-toggle=tooltip data-bs-placement=bottom title="Copy to clipboard"><img src=/images/copy-icon.svg></a><pre tabindex=0><code># Can be removed once: https://github.com/apache/beam/issues/21863 is fixed
class PytorchNoBatchModelHandler(PytorchModelHandlerKeyedTensor):
&#34;&#34;&#34;Wrapper to PytorchModelHandler to limit batch size to 1.
The tokenized strings generated from BertTokenizer may have different
lengths, which doesn&#39;t work with torch.stack() in current RunInference
implementation since stack() requires tensors to be the same size.
Restricting max_batch_size to 1 means there is only 1 example per `batch`
in the run_inference() call.
&#34;&#34;&#34;
def batch_elements_kwargs(self):
return {&#34;max_batch_size&#34;: 1}</code></pre></div></div><p>Because the <code>forward()</code> for <code>DistilBertModel</code> doesn&rsquo;t return the embeddings, we custom define the model_class <code>ModelWrapper</code> to get the vector embedding.</p><div class=snippet><div class="notebook-skip code-snippet without_switcher"><a class=copy type=button data-bs-toggle=tooltip data-bs-placement=bottom title="Copy to clipboard"><img src=/images/copy-icon.svg></a><pre tabindex=0><code>class ModelWrapper(DistilBertModel):
&#34;&#34;&#34;Wrapper to DistilBertModel to get embeddings when calling
forward function.&#34;&#34;&#34;
def forward(self, **kwargs):
output = super().forward(**kwargs)
sentence_embedding = (
self.mean_pooling(output,
kwargs[&#34;attention_mask&#34;]).detach().cpu().numpy())
return sentence_embedding
# Mean Pooling - Take attention mask into account for correct averaging
def mean_pooling(self, model_output, attention_mask):
&#34;&#34;&#34;
Calculates the mean of token embeddings
Args:
model_output: The output of the model.
attention_mask: This is a tensor that contains 1s for all input tokens and
0s for all padding tokens.
Returns:
The mean of the token embeddings.
&#34;&#34;&#34;
token_embeddings = model_output[
0] # First element of model_output contains all token embeddings
input_mask_expanded = (
attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float())
return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(
input_mask_expanded.sum(1), min=1e-9)</code></pre></div></div><p>After getting the embedding for each piece of Twitter text, the embeddings are normalized, because the trained model is expecting normalized embeddings.</p><div class=snippet><div class="notebook-skip code-snippet without_switcher"><a class=copy type=button data-bs-toggle=tooltip data-bs-placement=bottom title="Copy to clipboard"><img src=/images/copy-icon.svg></a><pre tabindex=0><code> | &#34;Normalize Embedding&#34; &gt;&gt; beam.ParDo(NormalizeEmbedding())</code></pre></div></div><h3 id=get-predictions>Get Predictions</h3><p>The normalized embeddings are then forwarded to the trained HDBSCAN model to get the predictions.</p><div class=snippet><div class="notebook-skip code-snippet without_switcher"><a class=copy type=button data-bs-toggle=tooltip data-bs-placement=bottom title="Copy to clipboard"><img src=/images/copy-icon.svg></a><pre tabindex=0><code> predictions = (
normalized_embedding
| &#34;Get Prediction from Clustering Model&#34;
&gt;&gt; RunInference(model_handler=clustering_model_handler)
)</code></pre></div></div><p>where <code>clustering_model_handler</code> is:</p><div class=snippet><div class="notebook-skip code-snippet without_switcher"><a class=copy type=button data-bs-toggle=tooltip data-bs-placement=bottom title="Copy to clipboard"><img src=/images/copy-icon.svg></a><pre tabindex=0><code> clustering_model_handler = KeyedModelHandler(
CustomSklearnModelHandlerNumpy(
model_uri=cfg.CLUSTERING_MODEL_PATH, model_file_type=ModelFileType.JOBLIB
)
)</code></pre></div></div><p>We define <code>CustomSklearnModelHandlerNumpy</code> as a wrapper to <code>SklearnModelHandlerNumpy</code> to limit batch size to one and to override <code>run_inference</code> so that <code>hdbscan.approximate_predict()</code> is used to get anomaly predictions.</p><div class=snippet><div class="notebook-skip code-snippet without_switcher"><a class=copy type=button data-bs-toggle=tooltip data-bs-placement=bottom title="Copy to clipboard"><img src=/images/copy-icon.svg></a><pre tabindex=0><code>class CustomSklearnModelHandlerNumpy(SklearnModelHandlerNumpy):
# limit batch size to 1 can be removed once: https://github.com/apache/beam/issues/21863 is fixed
def batch_elements_kwargs(self):
&#34;&#34;&#34;Limit batch size to 1 for inference&#34;&#34;&#34;
return {&#34;max_batch_size&#34;: 1}
# run_inference can be removed once: https://github.com/apache/beam/issues/22572 is fixed
def run_inference(self, batch, model, inference_args=None):
&#34;&#34;&#34;Runs inferences on a batch of numpy arrays.
Args:
batch: A sequence of examples as numpy arrays. They should
be single examples.
model: A numpy model or pipeline. Must implement predict(X).
Where the parameter X is a numpy array.
inference_args: Any additional arguments for an inference.
Returns:
An Iterable of type PredictionResult.
&#34;&#34;&#34;
_validate_inference_args(inference_args)
vectorized_batch = np.vstack(batch)
predictions = hdbscan.approximate_predict(model, vectorized_batch)
return [PredictionResult(x, y) for x, y in zip(batch, predictions)]</code></pre></div></div><p>After getting the model predictions, decode the output from <code>RunInference</code> into a dictionary. Next, store the prediction in a BigQuery table for analysis, update the HDBSCAN model, and send an email alert if the prediction is an anomaly.</p><div class=snippet><div class="notebook-skip code-snippet without_switcher"><a class=copy type=button data-bs-toggle=tooltip data-bs-placement=bottom title="Copy to clipboard"><img src=/images/copy-icon.svg></a><pre tabindex=0><code> _ = (
predictions
| &#34;Decode Prediction&#34; &gt;&gt; beam.ParDo(DecodePrediction())
| &#34;Write to BQ&#34; &gt;&gt; beam.io.WriteToBigQuery(
table=cfg.TABLE_URI,
schema=cfg.TABLE_SCHEMA,
write_disposition=beam.io.BigQueryDisposition.WRITE_APPEND,
create_disposition=beam.io.BigQueryDisposition.CREATE_IF_NEEDED,
))
_ = predictions | &#34;Alert by Email&#34; &gt;&gt; beam.ParDo(TriggerEmailAlert())</code></pre></div></div><div class=feedback><p class=update>Last updated on 2024/05/03</p><h3>Have you found everything you were looking for?</h3><p class=description>Was it all useful and clear? Is there anything that you would like to change? Let us know!</p><button class=load-button><a href="https://docs.google.com/forms/d/e/1FAIpQLSfID7abne3GE6k6RdJIyZhPz2Gef7UkpggUEhTIDjjplHuxSA/viewform?usp=header_link" target=_blank>SEND FEEDBACK</a></button></div></div></div><footer class=footer><div class=footer__contained><div class=footer__cols><div class="footer__cols__col footer__cols__col__logos"><div class=footer__cols__col__logo><img src=/images/beam_logo_circle.svg class=footer__logo alt="Beam logo"></div><div class=footer__cols__col__logo><img src=/images/apache_logo_circle.svg class=footer__logo alt="Apache logo"></div></div><div class=footer-wrapper><div class=wrapper-grid><div class=footer__cols__col><div class=footer__cols__col__title>Start</div><div class=footer__cols__col__link><a href=/get-started/beam-overview/>Overview</a></div><div class=footer__cols__col__link><a href=/get-started/quickstart-java/>Quickstart (Java)</a></div><div class=footer__cols__col__link><a href=/get-started/quickstart-py/>Quickstart (Python)</a></div><div class=footer__cols__col__link><a href=/get-started/quickstart-go/>Quickstart (Go)</a></div><div class=footer__cols__col__link><a href=/get-started/downloads/>Downloads</a></div></div><div class=footer__cols__col><div class=footer__cols__col__title>Docs</div><div class=footer__cols__col__link><a href=/documentation/programming-guide/>Concepts</a></div><div class=footer__cols__col__link><a href=/documentation/pipelines/design-your-pipeline/>Pipelines</a></div><div class=footer__cols__col__link><a href=/documentation/runners/capability-matrix/>Runners</a></div></div><div class=footer__cols__col><div class=footer__cols__col__title>Community</div><div class=footer__cols__col__link><a href=/contribute/>Contribute</a></div><div class=footer__cols__col__link><a href=https://projects.apache.org/committee.html?beam target=_blank>Team<img src=/images/external-link-icon.png width=14 height=14 alt="External link."></a></div><div class=footer__cols__col__link><a href=/community/presentation-materials/>Media</a></div><div class=footer__cols__col__link><a href=/community/in-person/>Events/Meetups</a></div><div class=footer__cols__col__link><a href=/community/contact-us/>Contact Us</a></div></div><div class=footer__cols__col><div class=footer__cols__col__title>Resources</div><div class=footer__cols__col__link><a href=/blog/>Blog</a></div><div class=footer__cols__col__link><a href=https://github.com/apache/beam>GitHub</a></div></div></div><div class=footer__bottom>&copy;
<a href=https://www.apache.org>The Apache Software Foundation</a>
| <a href=/privacy_policy>Privacy Policy</a>
| <a href=/feed.xml>RSS Feed</a><br><br>Apache Beam, Apache, Beam, the Beam logo, and the Apache feather logo are either registered trademarks or trademarks of The Apache Software Foundation. All other products or name brands are trademarks of their respective holders, including The Apache Software Foundation.</div></div><div class="footer__cols__col footer__cols__col__logos"><div class=footer__cols__col--group><div class=footer__cols__col__logo><a href=https://github.com/apache/beam><img src=/images/logos/social-icons/github-logo-150.png class=footer__logo alt="Github logo"></a></div><div class=footer__cols__col__logo><a href=https://www.linkedin.com/company/apache-beam/><img src=/images/logos/social-icons/linkedin-logo-150.png class=footer__logo alt="Linkedin logo"></a></div></div><div class=footer__cols__col--group><div class=footer__cols__col__logo><a href=https://twitter.com/apachebeam><img src=/images/logos/social-icons/twitter-logo-150.png class=footer__logo alt="Twitter logo"></a></div><div class=footer__cols__col__logo><a href=https://www.youtube.com/channel/UChNnb_YO_7B0HlW6FhAXZZQ><img src=/images/logos/social-icons/youtube-logo-150.png class=footer__logo alt="Youtube logo"></a></div></div></div></div></div></footer></body></html>