科技英文:Managing Machine Learning in Production with Kubeflow and DevOps

--- title: 科技英文:Managing Machine Learning in Production with Kubeflow and DevOps --- #Managing Machine Learning in Production with Kubeflow and DevOps - David Aronchick, Microsoft 1.Managing Machine Learning in Production with Kubeflow and DevOps - David Aronchick, Microsoft Dev:development /Ops:operations=keep going 1 00:00:00,030 --> 00:00:01,990 hi I apologize I was not working on my 嗨，我很抱歉，我沒有為我工作 Apologize(V):sorry 2 00:00:02,189 --> 00:00:03,428 slides one minute I'm not one of those 滑了一分鐘，我不是其中之一 3 00:00:03,629 --> 00:00:05,049 presenters but I was very late so I 主持人，但我來晚了，所以我 4 00:00:05,250 --> 00:00:08,560 apologize thank you so much I am David 很抱歉，我是大衛 5 00:00:08,759 --> 00:00:10,390 Ron chick I lead open source machine 羅恩·小妞我領導開源機器 open source:free 6 00:00:10,589 --> 00:00:13,330 learning strategy and ml at Microsoft 微軟的學習策略和毫升 Ml:maching Learning 7 00:00:13,529 --> 00:00:16,780 and Azure and I was previously the lead 和Azure，而我以前是負責人 Pentagon:top secret 8 00:00:16,980 --> 00:00:18,940 p.m. for kubernetes and I helped start 下午為kubernetes，我幫助開始 p.m.:project manager 9 00:00:19,140 --> 00:00:20,890 the Q flow project and I'm here to talk Q flow項目，我在這裡談 10 00:00:21,089 --> 00:00:23,470 to you about how to bring your machine 向您介紹如何攜帶您的機器 11 00:00:23,670 --> 00:00:26,260 learning to production using Q flow and 使用Q流學習生產 12 00:00:26,460 --> 00:00:33,219 mo ops so at Microsoft because the 在Microsoft運作的原因是 13 00:00:33,420 --> 00:00:35,320 widget is not working all right I will 小部件無法正常工作，我會 14 00:00:35,520 --> 00:00:38,049 be operating from my laptop and 通過我的筆記型電腦操作 Operating:ops 15 00:00:38,250 --> 00:00:39,969 Microsoft's we do have a lot of 微軟的我們確實有很多 16 00:00:40,170 --> 00:00:43,178 experience bringing ml to production we 將ml投入生產的經驗 17 00:00:43,378 --> 00:00:44,739 bring together your data we bring 彙集您帶來的資料 18 00:00:44,939 --> 00:00:48,759 together cloud your models and our 一起遮蓋您的模型，以及 Cloud:雲端 19 00:00:48,960 --> 00:00:51,009 job is really to help customers large 真的是説明大客戶的工作 20 00:00:51,210 --> 00:00:52,869 customers small customers whatever it is 客戶小客戶，無論是什麼 21 00:00:53,070 --> 00:00:55,869 help you move through and this is 幫助您通過，這是 22 00:00:56,070 --> 00:00:57,579 something that we have a lot of 我們有很多東西 23 00:00:57,780 --> 00:01:00,038 experience doing we've have a lot of 經歷過，我們有很多 24 00:01:00,238 --> 00:01:02,979 internal experience around Microsoft mic Microsoft的內部經驗 25 00:01:03,179 --> 00:01:06,189 Research and an ml generally with many 研究和毫升一般與許多 26 00:01:06,390 --> 00:01:08,168 of the most recent benchmarks and 最新的基準 Benchmarks:test for your computer speed 27 00:01:08,368 --> 00:01:09,730 achievements from coming out of 的成就來自 28 00:01:09,930 --> 00:01:11,439 Microsoft Research we're really 微軟研究院，我們真的非常 29 00:01:11,640 --> 00:01:13,509 proud of that and of course we do give 為此感到自豪，我們當然會給予 30 00:01:13,709 --> 00:01:15,009 all those back to the research community 所有那些回到研究界的人 31 00:01:15,209 --> 00:01:17,528 in the form of open papers and notebooks 以打開的紙和筆記本的形式 Form:way 32 00:01:17,728 --> 00:01:21,909 and data and the reality is that ml does 和數據，而現實是ml確實 33 00:01:22,109 --> 00:01:24,849 touch every aspect of Microsoft today 觸及Microsoft的各個方面 Aspect:a part 34 00:01:25,049 --> 00:01:26,679 literally every one of these logos and 實際上，這些徽標中的每一個和 Literally:actually 35 00:01:26,879 --> 00:01:30,450 many more from your customers clients 客戶的更多客戶 36 00:01:30,650 --> 00:01:33,878 rich clients dead clients are you know 有錢人客戶死了你知道嗎 :everybody use ml 37 00:01:34,078 --> 00:01:35,230 thin clients whatever they may be 瘦客戶無論他們是什麼 38 00:01:35,430 --> 00:01:39,039 Xbox phone you name it we're using ml in 您命名的Xbox手機，我們正在使用ml in Xbox:like switch 39 00:01:39,239 --> 00:01:41,349 all of these various places and we're 所有這些不同的地方，我們 40 00:01:41,549 --> 00:01:45,219 using it at enormous scale you know a 大規模使用它，您知道 Enormous:so big 41 00:01:45,420 --> 00:01:47,259 hundred and eighty million office users 一億八千萬辦公室用戶 42 00:01:47,459 --> 00:01:52,209 today every day use often features 今天每天經常使用我的功能 43 00:01:52,409 --> 00:01:55,659 in office we have 18 billion queries 在辦公室，我們有180億個查詢 Queries:question and other question to…. 44 00:01:55,859 --> 00:01:58,058 asked of Cortana which is obviously rich 問到Cortana，這顯然很豐富 45 00:01:58,259 --> 00:02:00,730 NLP and other things and six point five NLP和其他東西和六點五 46 00:02:00,930 --> 00:02:03,549 trillion security events evaluated every 每萬億安全事件評估一次 trillion :1,000,000,000,000 evaluated:how good something is 47 00:02:03,750 --> 00:02:06,698 day there's simply no way that we could 一天根本沒有辦法 48 00:02:06,899 --> 00:02:09,130 operate and process this data 操作和處理此資料 49 00:02:09,330 --> 00:02:10,980 without something like machine learning 沒有像機器學習這樣的東西 50 00:02:11,180 --> 00:02:14,020 so this is the point in the slide 這就是幻燈片中的重點 51 00:02:14,219 --> 00:02:14,920 where everyone's like wow that does 每個人都喜歡哇 52 00:02:15,120 --> 00:02:17,590 sound really great except they also say 聽起來真的很棒，除了他們也說 53 00:02:17,789 --> 00:02:21,370 this right ml is hard and it's really on 正確的毫升很難，而且確實在 54 00:02:21,569 --> 00:02:23,560 us that those that build these platforms 我們那些建立這些平臺的人 55 00:02:23,759 --> 00:02:26,530 in the ml community to help folks get 在ml社區中幫助人們獲得 Folks:people 56 00:02:26,729 --> 00:02:29,590 there better because we know something 那裡更好，因為我們知道一些 57 00:02:29,789 --> 00:02:32,200 that a lot of new people to ml don't 有很多新朋友不願意 58 00:02:32,400 --> 00:02:35,650 know and that's the following today a 知道，那就是今天的下一個 59 00:02:35,849 --> 00:02:39,340 lot of people new to ml think that it's 很多剛接觸毫升的人都認為這是 60 00:02:39,539 --> 00:02:42,070 all about the model and I understand 關於模型的一切，我瞭解 61 00:02:42,270 --> 00:02:44,350 why every new article out there talks 為什麼每一篇新文章都在談論 62 00:02:44,550 --> 00:02:46,300 about well you know Google just released 關於您知道Google剛剛發佈的 63 00:02:46,500 --> 00:02:48,460 Bert or Microsoft to release this and 伯特或微軟發佈此 64 00:02:48,659 --> 00:02:50,650 you know alphago did this and it's 而且您知道alphago做到了，這是 Alpha:the leader and best 65 00:02:50,849 --> 00:02:52,180 all about this amazing model that they 他們關於這個驚人的模型的一切 66 00:02:52,379 --> 00:02:54,189 built but it's not it's about the data 內置的，但這不是關於資料的 67 00:02:54,389 --> 00:02:56,920 processing and cleaning and all the 處理和清潔以及所有 68 00:02:57,120 --> 00:02:58,509 various things that are involved and 涉及的各種事物 69 00:02:58,709 --> 00:03:01,150 actually bringing something to 實際上帶來了一些東西 70 00:03:01,349 --> 00:03:04,689 production because that's the nature of 生產，因為這就是 71 00:03:04,889 --> 00:03:05,860 machine learning 機器學習 72 00:03:06,060 --> 00:03:08,860 it is these many many micro services 這是許多許多微服務 73 00:03:09,060 --> 00:03:10,360 each of which have a very specific 每個都有一個非常具體的 74 00:03:10,560 --> 00:03:12,310 functionality often very specific 功能通常非常具體 75 00:03:12,509 --> 00:03:14,710 tooling that do very specific things 做非常具體的事情的工具 76 00:03:14,909 --> 00:03:16,960 well but then need to be coupled 好，但是需要結合 77 00:03:17,159 --> 00:03:19,330 together in an intelligent way and if 以一種聰明的方式在一起，如果 Intelligent:smart 78 00:03:19,530 --> 00:03:22,180 you just focus on the model then you're 您只關注模型，然後 79 00:03:22,379 --> 00:03:25,450 gonna be in trouble and I know what 會遇到麻煩，我知道 80 00:03:25,650 --> 00:03:27,580 you're saying you're saying your data 你是說你是在說你的數據 81 00:03:27,780 --> 00:03:29,620 scientists and you don't care and I 科學家，你不在乎，我 82 00:03:29,819 --> 00:03:33,280 believe you a little bit but I'm here to 相信你一點，但我在這裡 83 00:03:33,479 --> 00:03:36,069 tell you you actually do and the reason 告訴你你實際做的事以及原因 84 00:03:36,269 --> 00:03:40,500 is tweets like this right models are 是這樣的推文，正確的模型是 IDGAF:I don’t give a fuck= I don’t care 85 00:03:40,699 --> 00:03:43,750 relatively easy to build but they are 相對容易構建，但它們 Relatively:more or less 86 00:03:43,949 --> 00:03:47,380 very hard to roll out because more often 很難推出，因為更多時候 hard to roll out:make people use 87 00:03:47,580 --> 00:03:49,689 than not data scientists operate in a 而不是資料科學家在 88 00:03:49,889 --> 00:03:51,460 way that they are familiar with they 他們熟悉的方式 89 00:03:51,659 --> 00:03:53,319 understand their tools and they build 瞭解他們的工具，他們建立 90 00:03:53,519 --> 00:03:55,660 using their tooling and local laptops 使用他們的工具和本地筆記型電腦 91 00:03:55,860 --> 00:03:58,210 and local clusters but they don't know 和本地集群，但他們不知道 92 00:03:58,409 --> 00:03:59,770 how to reach out and roll it to 如何伸出手並推向 93 00:03:59,969 --> 00:04:01,480 production and the reason is because 生產的原因是因為 94 00:04:01,680 --> 00:04:04,719 you have this separation right the data 你有這種分離權的資料 95 00:04:04,919 --> 00:04:06,219 scientists over here they're trying to 他們正在嘗試的科學家 96 00:04:06,419 --> 00:04:08,500 iterate as quickly as she can she wants 盡可能快地反覆運算 Iterate:explain 97 00:04:08,699 --> 00:04:10,120 to use frameworks and tooling she 使用框架和工具 98 00:04:10,319 --> 00:04:12,520 understands she wants to mix and match 知道她想混搭 99 00:04:12,719 --> 00:04:15,069 tools the absolute latest build of 工具的絕對最新版本 100 00:04:15,269 --> 00:04:17,170 tensorflow or PI Torture or tensorflow或PI Torture y瑪瑙或 101 00:04:17,370 --> 00:04:19,689 whatever it may be you know somebody 不管你認識誰 102 00:04:19,889 --> 00:04:21,430 from Carnegie Mellon just launched a 來自卡耐基梅隆大學的 Carnegie Mellon:有名的大學 103 00:04:21,629 --> 00:04:24,699 brand new tool around reproducibility 圍繞重現性的全新工具 Reproducibility:how easy to make copy 104 00:04:24,899 --> 00:04:26,540 and it's respectability and that's 這是受人尊敬的 105 00:04:26,740 --> 00:04:28,309 she should have the capabilities to do 她應該有能力去做 106 00:04:28,509 --> 00:04:30,829 that she also wants to not worry about 她也不想擔心 107 00:04:31,029 --> 00:04:32,210 management because it's just her laptop 管理因為這只是她的筆記型電腦 108 00:04:32,410 --> 00:04:33,800 you know something goes wrong she 刮鬍子你知道出事了她 109 00:04:34,000 --> 00:04:36,170 can flatten it and restart and on the 可以將其展平並重新啟動，然後在 110 00:04:36,370 --> 00:04:37,579 other hand she wants unlimited scale 另一方面，她想要無限的規模 111 00:04:37,779 --> 00:04:39,319 she's got a paper due you know on 她有一份論文，因為你知道 112 00:04:39,519 --> 00:04:41,718 Thursday at 5:00 p.m. she wants all the 星期四下午5:00，她想要所有 113 00:04:41,918 --> 00:04:43,100 GPUs in the world in order to achieve 為了實現世界上的GPU GPUs:graphics(圖形) processing unit 114 00:04:43,300 --> 00:04:46,369 the results she needs on the other hand 另一方面，她需要的結果 115 00:04:46,569 --> 00:04:49,100 you have the right and she needs 你有權利，她需要 116 00:04:49,300 --> 00:04:51,800 consistency she needs observability she 一致性，她需要可觀察性，她 117 00:04:52,000 --> 00:04:53,420 needs to reuse tooling that's already 需要重用已經存在的工具 118 00:04:53,620 --> 00:04:55,490 been approved by her organization 被她的組織批准 119 00:04:55,689 --> 00:04:57,350 because they have support and things 因為他們有支持和東西 120 00:04:57,550 --> 00:05:00,410 built in and she needs uptime if the if 內置，如果需要，她需要正常執行時間 Uptime:time in the computer 121 00:05:00,610 --> 00:05:02,240 things are constantly changing under the 在不斷變化的情況下 122 00:05:02,439 --> 00:05:05,778 hood with no records of what's going on 引擎蓋，沒有發生任何事情的記錄 under the hood:parts inside 123 00:05:05,978 --> 00:05:09,649 then that's not gonna work now I am here 那那不行了我現在在這裡 124 00:05:09,848 --> 00:05:12,139 to propose that we can bring them 提議我們可以帶他們去 125 00:05:12,339 --> 00:05:14,210 together and we're gonna do it 在一起，我們要做 126 00:05:14,410 --> 00:05:19,610 through ml ops now before I begin I want 在我開始之前，現在要通過ml ops 127 00:05:19,810 --> 00:05:22,100 to talk about what the e2e machine 說說什麼e2e機 e2e: end to end=from here to there 128 00:05:22,300 --> 00:05:23,389 learning lifecycle looks like now you 學習生命週期看起來像現在 129 00:05:23,589 --> 00:05:24,860 saw a bunch of like boxes connected 看到一堆像盒子一樣的盒子 130 00:05:25,060 --> 00:05:26,749 together with arrows there but the main 與那裡的箭頭，但主要 Arrows-> 131 00:05:26,949 --> 00:05:28,759 things that I think we need to identify 我認為我們需要確定的事情 132 00:05:28,959 --> 00:05:31,459 and solve for in an lifecycle is 並在生命週期中解決 133 00:05:31,658 --> 00:05:34,249 what you see here first you do have the 首先您在這裡看到的是 134 00:05:34,449 --> 00:05:36,050 development and deployment of the model 模型的開發和部署 135 00:05:36,250 --> 00:05:37,939 then you're gonna have to package it in 那麼你將不得不打包 136 00:05:38,139 --> 00:05:40,490 a way that can be used and migrated to migrated:move to new place 一種可以使用和遷移到的方式 137 00:05:40,689 --> 00:05:42,499 in production you want to validate that 在生產中，您要驗證 138 00:05:42,699 --> 00:05:44,499 model behavior before you roll it out 在推出之前模擬行為 139 00:05:44,699 --> 00:05:46,699 then you want to deploy the model and 然後您要部署模型並 Deploy:調派 140 00:05:46,899 --> 00:05:48,170 then you want to monitor it and in 然後您要監視它並在 141 00:05:48,370 --> 00:05:49,819 monitoring it is not just about 監控它不僅僅是 142 00:05:50,019 --> 00:05:51,410 having it out there and making sure it's 把它放在那裡並確保它 143 00:05:51,610 --> 00:05:53,930 up but it's in fact taking all that data 上升，但實際上是在吸收所有資料 144 00:05:54,129 --> 00:05:56,809 and feeding it back into the original so 並把它送回原來的 145 00:05:57,009 --> 00:05:58,430 that you can now train it again and be 現在您可以再次訓練它並成為 146 00:05:58,629 --> 00:05:59,350 smarter about it 對此更聰明 147 00:05:59,550 --> 00:06:01,639 so you may say you might have heard this 所以你可能會說你可能聽說過 148 00:06:01,839 --> 00:06:04,819 before and the reality is you have you 以前，現實是你擁有你 149 00:06:05,019 --> 00:06:06,680 heard this several years ago when things 幾年前聽說過 150 00:06:06,879 --> 00:06:08,240 started getting kicked off around get 開始在周圍開始 kicked off: kickstart:there has many new project in WEBSITE 151 00:06:08,439 --> 00:06:10,968 ops and that was the idea that you could 操作，那是您可以 152 00:06:11,168 --> 00:06:14,749 start with get and record everything 從獲取開始並記錄一切 153 00:06:14,949 --> 00:06:16,129 that you were doing relative to your 你相對於你所做的 154 00:06:16,329 --> 00:06:19,218 overall pipeline iterate very quickly on 整個管道非常快速地反覆運算 155 00:06:19,418 --> 00:06:20,629 that pipeline and then once that 該管道，然後一旦 156 00:06:20,829 --> 00:06:22,730 pipeline had passed all its tests and 管道通過了所有測試，並且 157 00:06:22,930 --> 00:06:24,319 humans had looked at it and said yes 人們看著它說是 158 00:06:24,519 --> 00:06:26,809 this is ready to go you trigger a second 準備好了，您觸發第二個 Trigger:to start something 159 00:06:27,009 --> 00:06:29,149 cycle and that second cycle is rolling 週期，第二個週期正在滾動 160 00:06:29,348 --> 00:06:31,338 it out to production but again it's all 投入生產，但這又是全部 161 00:06:31,538 --> 00:06:33,259 driven off get you don't have someone 被趕走，你沒有人 162 00:06:33,459 --> 00:06:35,209 in the middle there introducing any new 在中間介紹任何新的 163 00:06:35,408 --> 00:06:37,160 changes because those wouldn't have been 變化，因為那些不會 164 00:06:37,360 --> 00:06:39,050 tested and relied or 經過測試並依賴 165 00:06:39,250 --> 00:06:41,090 tested and observed and made sure that 測試並觀察並確保 166 00:06:41,290 --> 00:06:43,340 they pass all your bars and what that 他們通過了你所有的酒吧，那 167 00:06:43,540 --> 00:06:45,170 really gets to you at is those lines 真正讓您著迷的是那些臺詞 168 00:06:45,370 --> 00:06:47,000 at the bottom you get velocity plus 在底部，您獲得速度加 Velocity:speed for something 169 00:06:47,199 --> 00:06:51,590 security so what we need is ml ops right 安全性，所以我們需要的是ml ops 170 00:06:51,790 --> 00:06:53,629 and ml ops is going to look like this 和ml ops看起來像這樣 171 00:06:53,829 --> 00:06:56,840 right you start with the data scientist 從資料科學家開始 172 00:06:57,040 --> 00:06:58,250 the data scientist is able to iterate 資料科學家能夠反覆運算 Iterate:explain 173 00:06:58,449 --> 00:07:00,319 exactly as quickly as she could and 盡其所能 174 00:07:00,519 --> 00:07:02,030 potentially even faster because you're 可能甚至更快，因為您 Potentially:something could but not sure 175 00:07:02,230 --> 00:07:03,470 giving her the Best of Breed tools and 給她最好的工具 176 00:07:03,670 --> 00:07:05,270 you're giving in to her in a system 你在系統中屈服於她 177 00:07:05,470 --> 00:07:06,560 where she doesn't have to think about 她不必思考的地方 178 00:07:06,759 --> 00:07:08,870 well this Python dependency didn't work 很好，這個Python依賴項不起作用 179 00:07:09,069 --> 00:07:10,370 with that other package that I had over 與我已經結束的其他包裹 180 00:07:10,569 --> 00:07:12,889 there you give her a path to install and 在那裡，您給她安裝的路徑， 181 00:07:13,089 --> 00:07:14,600 manage these things in a very regular 定期管理這些東西 182 00:07:14,800 --> 00:07:17,660 way and then from that you start the 方式，然後從那開始 183 00:07:17,860 --> 00:07:18,350 second half 下半場 184 00:07:18,550 --> 00:07:20,600 right you have her finish you have 對，你有她的成就，你有 185 00:07:20,800 --> 00:07:23,000 her check in and move forward through 她簽入並繼續前進 186 00:07:23,199 --> 00:07:25,040 dev that's integrating it into the 開發人員將其集成到 ML+DEV+OPS Integrating:整合 187 00:07:25,240 --> 00:07:27,259 application itself and then integrate 應用程式本身，然後集成 188 00:07:27,459 --> 00:07:29,569 again when it comes to rolling out to 再次涉及到 189 00:07:29,769 --> 00:07:32,629 production and when you do that you get 生產，當您這樣做時，您會得到 190 00:07:32,829 --> 00:07:34,430 these benefits right you get great 這些好處使您受益匪淺 191 00:07:34,629 --> 00:07:37,129 observability that means you exactly 可觀察性意味著您完全 192 00:07:37,329 --> 00:07:39,020 what's being rolled out so you're able 正在推出什麼，以便您能夠 193 00:07:39,220 --> 00:07:41,240 to observe it using standard toolings 使用標準工具進行觀察 194 00:07:41,439 --> 00:07:43,879 and reproducibility you get validation 和重現性得到驗證 195 00:07:44,079 --> 00:07:45,560 again because the stuff is checked in 再次因為東西被簽入 196 00:07:45,759 --> 00:07:48,560 and good you can do static validation 很好，您可以進行靜態驗證 Static:not moving validation:驗證 197 00:07:48,759 --> 00:07:50,870 you could do runtime validation by 您可以通過執行運行時驗證 198 00:07:51,069 --> 00:07:52,520 rolling it out to Canaries in a very 很快地將它推廣到金絲雀 199 00:07:52,720 --> 00:07:54,860 regimented way and then of course you 規整的方式，然後你當然 Regimented:array 200 00:07:55,060 --> 00:07:56,840 get reproducibility and auditability you 獲得可重複性和可審核性 Auditability:audit 201 00:07:57,040 --> 00:07:59,240 know that whatever that query was that 知道那個查詢是什麼 202 00:07:59,439 --> 00:08:01,069 prediction that happened at the end of 發生在末尾的預測 203 00:08:01,269 --> 00:08:03,079 that very long cycle you're able to 如此長的週期，您能夠 204 00:08:03,279 --> 00:08:05,300 trace back exactly what code rolled it 追溯確切的代碼 205 00:08:05,500 --> 00:08:07,790 out at every step of the way and in this 在這個過程中的每一步 206 00:08:07,990 --> 00:08:10,790 byte you get velocity plus security for 位元組，您可以獲得速度和安全性 207 00:08:10,990 --> 00:08:13,610 ml so some of the specific components 毫升，所以一些特定成分 208 00:08:13,810 --> 00:08:15,050 here and I'll just walk through this 在這裡，我將逐步流覽 209 00:08:15,250 --> 00:08:17,090 really quickly like I said you have the 真的很快，就像我說過的 210 00:08:17,290 --> 00:08:19,100 data scientist she checks in to get at 她簽到的資料科學家 211 00:08:19,300 --> 00:08:21,350 that point you do code set version or 那點你做代碼設置版本或 212 00:08:21,550 --> 00:08:23,210 you do code dataset and environment 您編寫資料集和環境代碼 213 00:08:23,410 --> 00:08:24,949 versioning you snapshot all of those 對所有這些版本進行快照 214 00:08:25,149 --> 00:08:27,230 things from there it automatically 從那裡自動進行的事情 215 00:08:27,430 --> 00:08:30,280 builds both the app and trains the model 同時構建應用程式和訓練模型 216 00:08:30,480 --> 00:08:32,750 and in that you're able to do hyper 而且你可以做超 Hyper:super 217 00:08:32,950 --> 00:08:34,218 parameter sweeps and all those kind of 參數掃描以及所有此類 Parameter: Sweeps:掃毒 218 00:08:34,418 --> 00:08:36,949 things in an automated way you validate 驗證的自動化方式 219 00:08:37,149 --> 00:08:38,629 the model at that testing point using 使用該測試點的模型 220 00:08:38,830 --> 00:08:41,299 things like model validation and 諸如模型驗證和 221 00:08:41,500 --> 00:08:43,309 certification and things like that you 認證和類似的東西 222 00:08:43,509 --> 00:08:45,139 release the app and then finally you 釋放應用程式，然後最終您 223 00:08:45,339 --> 00:08:47,029 roll it out and so where that's what 推出它，那在哪裡 224 00:08:47,230 --> 00:08:48,169 we're gonna try and get to here today 我們今天要去這裡 225 00:08:48,370 --> 00:08:49,879 and we're gonna do it right here on 我們將在這裡繼續 226 00:08:50,080 --> 00:08:52,578 stage for you well technically 從技術上為您做好舞臺 Technically:具體 227 00:08:52,778 --> 00:08:56,659 so the question is this sounds pretty 所以問題是聽起來很漂亮 228 00:08:56,860 --> 00:08:58,969 good I'd like to take it on what should 好，我想拿什麼做 229 00:08:59,169 --> 00:09:01,519 I do in order to achieve that well I 我為了做到這一點而努力 230 00:09:01,720 --> 00:09:03,378 have a suggestion for you go join one of 有一個建議給你去參加其中之一 231 00:09:03,578 --> 00:09:05,029 these big companies because they already 這些大公司，因為他們已經 232 00:09:05,230 --> 00:09:08,029 have and they have it it's you know at 有，他們有，這是你知道的 233 00:09:08,230 --> 00:09:10,428 Google it's tf-x and at Facebook it's FB Google是TF-X，Facebook是FB 234 00:09:10,629 --> 00:09:13,488 Lerner flow at uber it's Michelangelo 尤納的勒納潮流是米開朗基羅 235 00:09:13,688 --> 00:09:15,318 Microsoft has our own which is called 微軟有我們自己的被稱為 236 00:09:15,519 --> 00:09:19,039 ether and what this is it's a very 乙太，這是一個非常 Ether:magic in the air 237 00:09:19,240 --> 00:09:21,738 standard data science platform where a 標準資料科學平臺 238 00:09:21,938 --> 00:09:25,698 data scientist will come in and you know 資料科學家會進來，你知道 239 00:09:25,899 --> 00:09:27,469 be able to interact at a very high level 能夠進行高水準的互動 240 00:09:27,669 --> 00:09:30,139 see other people's experiments see her 看別人的實驗見她 241 00:09:30,339 --> 00:09:32,868 own experiments iterate and version on 自己的實驗進行反覆運算並在其上進行版本控制 242 00:09:33,068 --> 00:09:35,448 her own experiments and then basically 她自己的實驗，然後基本上 243 00:09:35,649 --> 00:09:37,099 download and begin to use those 下載並開始使用那些 244 00:09:37,299 --> 00:09:39,019 experiments rather than having to start 實驗而不是必須開始 245 00:09:39,220 --> 00:09:40,578 from scratch every time she can start 每次她可以從頭開始 Start from scratch:star from nothing 246 00:09:40,778 --> 00:09:42,618 there and and move forward rather than 在那裡，前進而不是 247 00:09:42,818 --> 00:09:46,669 you know resetting everything so you 你知道重置一切，所以你 248 00:09:46,870 --> 00:09:48,799 have rejected my suggestions 拒絕了我的建議 249 00:09:49,000 --> 00:09:50,479 that's very harsh of you I don't want to 對你來說我很苛刻 Harsh:too difficult 250 00:09:50,679 --> 00:09:51,498 work at a company so I'm gonna help you 在公司工作，所以我會幫助你 251 00:09:51,698 --> 00:09:55,039 by building your own ml ops platform 通過建立自己的ml ops平臺 252 00:09:55,240 --> 00:09:57,128 using first we'll do it with C ICD 首先使用C ICD 253 00:09:57,328 --> 00:10:01,099 second we'll add in git and third we'll 第二，我們添加git，第三，我們添加 254 00:10:01,299 --> 00:10:04,459 add a soon you start with Q plug that 添加一個很快您開始使用Q外掛程式 255 00:10:04,659 --> 00:10:07,608 was get in the middle or get lab or 在中間或實驗室 256 00:10:07,808 --> 00:10:09,019 bitbucket or whatever it might be and 比特桶或其他可能的東西 257 00:10:09,220 --> 00:10:11,659 then finally we'll do it with a C ICD 最後，我們將使用C ICD 258 00:10:11,860 --> 00:10:14,868 platform by the way I hope I did my best 平臺，希望我能做到最好 259 00:10:15,068 --> 00:10:16,818 I'm trying to be very neutral here all 我試圖在這裡保持中立 Neutral: 中立 260 00:10:17,019 --> 00:10:20,058 logos are the same size and by the way 徽標的大小和方式相同 261 00:10:20,259 --> 00:10:23,628 if you are thinking of inventing a new C 如果您正在考慮發明新的C 262 00:10:23,828 --> 00:10:27,289 ICD platform there are so many there are ICD平臺有很多 263 00:10:27,490 --> 00:10:31,849 so many C ICD platforms it's enough just 這麼多的C ICD平臺就足夠了 264 00:10:32,049 --> 00:10:34,128 help another one help one that already 幫助另一個已經幫助的人 265 00:10:34,328 --> 00:10:39,318 exists I promise you okay so joking 存在，我保證你還可以，所以開玩笑 Exists:egg+sister 266 00:10:39,519 --> 00:10:42,738 aside let's go ml ops in the real I'm 拋開現實，我走了 267 00:10:42,938 --> 00:10:45,828 gonna take on a challenge that many 將會面臨許多挑戰 268 00:10:46,028 --> 00:10:48,318 people I don't know if Corey's here 我不知道寇里在這裡的人 269 00:10:48,519 --> 00:10:49,669 Corey are you here I know that you're 寇里你在這裡我知道你在 270 00:10:49,870 --> 00:10:51,108 attending cube con but I'm not sure you 參加立方體騙局，但我不確定你 Con:conference 活動 271 00:10:51,308 --> 00:10:52,789 hear Corey doesn't believe that at 聽到寇里不相信 272 00:10:52,990 --> 00:10:55,248 Multi-cloud exists I'm gonna say that it 多面存在，我要說 Multi:many 273 00:10:55,448 --> 00:10:56,959 does and I'm gonna say it does in the 會，我會說它在 274 00:10:57,159 --> 00:10:58,969 following way multi-cloud really does 按照多雲真正的方法 275 00:10:59,169 --> 00:11:00,558 exist in the real world and I'm gonna 存在于現實世界中，我要 276 00:11:00,759 --> 00:11:03,529 lay out a very standard scenario and 提出一個非常標準的方案 Scenario:common story 277 00:11:03,730 --> 00:11:05,379 then solve it for you on stage 然後在舞臺上為你解決 278 00:11:05,580 --> 00:11:08,409 so what when I say multi-cloud this is 所以當我說多雲時 279 00:11:08,610 --> 00:11:10,929 what it looks like you have at the top 頂部看起來像什麼 280 00:11:11,129 --> 00:11:12,939 you have cloud whatever cloud it might 無論有沒有雲，您都有雲 281 00:11:13,139 --> 00:11:14,889 be and at the bottom you have on Prem be and在Prem的最底端 on Prem:on the local hard drive 282 00:11:15,089 --> 00:11:17,109 and again that could be that's a special 再次可能那是一個特殊的 283 00:11:17,309 --> 00:11:18,639 case of a second cloud it could be a 第二個雲的情況可能是 284 00:11:18,839 --> 00:11:19,448 second cloud 第二朵雲 285 00:11:19,649 --> 00:11:20,769 it could be on Prem could be your local 它可能在Prem上可能是您本地的 286 00:11:20,970 --> 00:11:22,539 laptop doesn't matter and in the middle 筆記型電腦沒關係，在中間 287 00:11:22,740 --> 00:11:26,529 you have get the reason you've chosen 你有選擇的理由 288 00:11:26,730 --> 00:11:28,750 that top cloud is because it's 那頂雲是因為 289 00:11:28,950 --> 00:11:30,008 distributed is something that your 分發是你的事 Distributed:to give to many people 290 00:11:30,208 --> 00:11:31,899 datacenter can't offer you today maybe 資料中心今天可能無法為您提供服務 291 00:11:32,100 --> 00:11:32,828 it's up time 時間到了 up time:how much time it takes on the computer 292 00:11:33,028 --> 00:11:34,959 maybe it's locality because of 也許是因為 Locality(adv) 293 00:11:35,159 --> 00:11:37,059 regulatory reasons maybe it's just 監管方面的原因可能僅僅是 Regulatory(V):規制 294 00:11:37,259 --> 00:11:39,189 closer to your IOT deployments right you 更接近您的物聯網部署 deploy =Distributed:to give to many people 295 00:11:39,389 --> 00:11:41,139 don't want to have that far big Latancy 不想擁有那麼大的萊頓 Latency:ping = ping test 296 00:11:41,339 --> 00:11:43,419 sees and so on but the reason used to 看到等等，但以前的原因 297 00:11:43,620 --> 00:11:44,679 have stuff on prem is because you 在prem上有東西是因為你 298 00:11:44,879 --> 00:11:45,909 actually have a lot of data that you've 實際上有很多資料 299 00:11:46,110 --> 00:11:48,128 been collecting for many years and it's 已經收集了很多年了 300 00:11:48,328 --> 00:11:49,899 just a lot a big pain in the buttom to 屁股上的痛苦很大 a big pain:a big problem again 301 00:11:50,100 --> 00:11:52,599 move it to the cloud we see this 將其移動到雲中，我們看到了 302 00:11:52,799 --> 00:11:54,399 scenario all the time and it's whether 情況一直存在，這是否 Scenario:common situation 303 00:11:54,600 --> 00:11:56,198 or not you know maybe that's a poor 還是你不知道那是一個窮人 304 00:11:56,399 --> 00:11:57,698 latency connection or maybe you just 延遲連接，也許你只是 Latency: 305 00:11:57,899 --> 00:12:00,159 don't want to spend you know many many 不想花你知道很多很多 306 00:12:00,360 --> 00:12:01,689 millions of dollars on petabytes of data PB級數據的數百萬美元小到大Megabyte->Gigabyte->terabyte(1000GB)->Petabytes(1000000GB) 307 00:12:01,889 --> 00:12:02,769 being held in the cloud because you 被困在雲端，因為你 308 00:12:02,970 --> 00:12:03,849 already have a bunch of disks and it's 已經有一堆光碟了 309 00:12:04,049 --> 00:12:06,149 doing its job you don't want to do that 做它的工作，你不想那樣做 310 00:12:06,350 --> 00:12:09,008 so and then we have our to cast of 所以然後我們要 311 00:12:09,208 --> 00:12:10,659 characters we have our data scientist 我們擁有資料科學家的角色 312 00:12:10,860 --> 00:12:13,448 she wants to iterate very quickly on the 她想非常快速地反覆運算 Iterate:explain 313 00:12:13,649 --> 00:12:15,819 the model and make sure that it works 模型並確保其有效 314 00:12:16,019 --> 00:12:17,799 and she's gonna want to train next to 她要在旁邊訓練 315 00:12:18,000 --> 00:12:20,198 the model but then the sres 模型，然後是sres 316 00:12:20,399 --> 00:12:21,698 and the ml engineers are going to want 和毫升工程師會想要 317 00:12:21,899 --> 00:12:25,899 to host in the cloud so let's do this so 託管在雲中，所以我們這樣做 318 00:12:26,100 --> 00:12:28,508 what she does is first she's gonna check 她要做的是首先要檢查 319 00:12:28,708 --> 00:12:32,500 in her code and to get will be 在她的代碼，並獲得將是 320 00:12:32,700 --> 00:12:34,809 watched by the CI CD pipeline for a new CI CD管道正在監視新的 321 00:12:35,009 --> 00:12:37,179 check-in and that will automatically 簽入，這將自動 322 00:12:37,379 --> 00:12:41,198 trigger a cube flow pipeline it that cue 觸發一個提示的多維資料集流管道 323 00:12:41,399 --> 00:12:43,029 flow pipeline will process the data 流管道將處理資料 324 00:12:43,230 --> 00:12:44,769 first in whatever way it needs 首先以任何需要的方式 325 00:12:44,970 --> 00:12:46,959 processing or if it's already been 正在處理或已經處理過 326 00:12:47,159 --> 00:12:48,819 processed and check pointed then it can 處理並檢查指出，然後可以 327 00:12:49,019 --> 00:12:50,319 skip to the next step and cue flows 跳到下一步並提示流程 328 00:12:50,519 --> 00:12:51,429 smart enough to do that 足夠聰明地做到這一點 329 00:12:51,629 --> 00:12:54,479 it then will run a training run 然後它將進行一次訓練 330 00:12:54,679 --> 00:12:58,120 automatically on that process data again 再次自動處理該過程資料 331 00:12:58,320 --> 00:12:59,859 using whatever parameters you've passed 使用您傳遞的任何參數 332 00:13:00,059 --> 00:13:02,409 in via get right again this isn't 通過再次獲得權利這不是 333 00:13:02,610 --> 00:13:03,609 something where the data scientist is 資料科學家所在的地方 334 00:13:03,809 --> 00:13:05,318 picking it on the fly you want to be 隨時隨地挑選它 335 00:13:05,519 --> 00:13:08,258 extremely prescriptive about this once 關於這一次非常規範 Prescriptive:don’t give wrong code 336 00:13:08,458 --> 00:13:10,059 that's done it's gonna register the 這樣就可以註冊 Register:報名 337 00:13:10,259 --> 00:13:12,459 model at a central point which can be 可以在一個中心點建模 338 00:13:12,659 --> 00:13:15,669 picked up and then rolled out to the 拿起然後展開到 339 00:13:15,870 --> 00:13:17,139 cloud and in this case you would roll it 雲，在這種情況下，您將其滾動 340 00:13:17,339 --> 00:13:19,120 out to something like a canary end point 到金絲雀終點 341 00:13:19,320 --> 00:13:21,009 or a staging endpoint to make sure that 或暫存端點以確保 342 00:13:21,210 --> 00:13:24,069 it runs a human being often will after 它使一個人經常會 343 00:13:24,269 --> 00:13:26,500 the tests of paws come by make sure 通過檢查爪子來確保 344 00:13:26,700 --> 00:13:29,469 everything is operating properly that 一切運行正常 345 00:13:29,669 --> 00:13:31,809 the human being will then say okay it's 然後，人類會說，好的 346 00:13:32,009 --> 00:13:33,669 time for me to trigger the next step and 是我觸發下一步的時間了， 347 00:13:33,870 --> 00:13:35,799 that's at that point it gets rolled out 到那時它被推出 348 00:13:36,000 --> 00:13:38,259 to a public service endpoint and use for 到公共服務端點並用於 349 00:13:38,460 --> 00:13:42,639 the model for production okay so that's 生產模型還可以，所以 350 00:13:42,840 --> 00:13:44,319 the high level let's see what this 高層次，讓我們看看這是什麼 351 00:13:44,519 --> 00:13:47,689 actually looks like when it's running 實際上看起來像是在運行時 352 00:13:47,779 --> 00:13:53,889 and yes you can okay well this is gonna 是的，你可以很好，這將會 353 00:13:54,090 --> 00:13:55,359 be very hard because I shoulder the 很難，因為我肩負著 354 00:13:55,559 --> 00:13:57,819 whole time alright so this is um you 整個時間都很好，所以這是你 355 00:13:58,019 --> 00:13:59,919 know the one downside of doing we're 知道做我們的一個缺點 Downside:壞處 356 00:14:00,120 --> 00:14:02,079 moving into ml is you have to record all 進入毫升是你必須記錄所有 357 00:14:02,279 --> 00:14:04,049 your videos because everything takes forever 您的視頻會永遠解決一切 358 00:14:04,250 --> 00:14:07,089 so I promise you can look at the 所以我保證你可以看看 359 00:14:07,289 --> 00:14:08,379 date I don't know where the data is on 日期，我不知道資料在哪裡 360 00:14:08,580 --> 00:14:10,000 anyone see a date on there this is 有人在那裡看到約會 361 00:14:10,200 --> 00:14:12,189 literally run last night so I promise 昨天晚上確實跑了，所以我保證 362 00:14:12,389 --> 00:14:14,109 you can go to this repo which is 你可以去這個倉庫 363 00:14:14,309 --> 00:14:15,879 public and download all the code and 公開並下載所有代碼， 364 00:14:16,080 --> 00:14:18,939 have a good time okay where are you 玩得開心吧，你在哪裡 365 00:14:19,139 --> 00:14:23,199 okay so first is this running it is okay 好的，所以首先運行它是可以的 366 00:14:23,399 --> 00:14:25,539 so first what you see here this is a 首先，您在這裡看到的是 367 00:14:25,740 --> 00:14:28,719 very standard repo out there and it's 非常標準的倉庫在那裡 368 00:14:28,919 --> 00:14:30,309 got all the code in it and I let me show 裡面有所有代碼，我讓我展示 369 00:14:30,509 --> 00:14:31,870 you actually exactly what this code 您實際上正是這段代碼 370 00:14:32,070 --> 00:14:36,920 looks like here see this you cannot oh 看起來像這裡看到你不能哦 371 00:14:41,389 --> 00:14:46,059 okay so they always say never show code 好吧，所以他們總是說從不顯示代碼 372 00:14:46,259 --> 00:14:48,159 when you're doing a demo because 在進行演示時，因為 373 00:14:48,360 --> 00:14:49,779 everyone's eyes glaze over but these are 每個人的眼睛都凝視著，但這是 eyes glaze over:they don’t understand 374 00:14:49,980 --> 00:14:51,639 very smart room so I'm gonna trust you 非常聰明的房間，所以我會相信你 375 00:14:51,840 --> 00:14:53,949 your eyes are not going to leave over so 你的眼睛不會離開所以 376 00:14:54,149 --> 00:14:55,449 what you have here is this is a very 你在這裡是一個非常 377 00:14:55,649 --> 00:14:57,939 very standard CI CD pipeline this is 非常標準的CI CD管道，這是 378 00:14:58,139 --> 00:14:59,949 using Azure dev ops which is something 使用Azure開發人員操作 379 00:15:00,149 --> 00:15:02,409 that we use hosted dev ops CI CD on 我們使用託管的dev ops CI CD Host: the leader 380 00:15:02,610 --> 00:15:04,659 Azure but the Yambol is gonna look super 蔚藍但Yambol看起來超級好 381 00:15:04,860 --> 00:15:06,250 familiar because it follows very 熟悉，因為它遵循 382 00:15:06,450 --> 00:15:08,769 standard cid CD practices looks just 標準CID CD的做法看起來只是 383 00:15:08,970 --> 00:15:11,589 like jenkins and what you see here is at 就像詹金斯，你在這裡看到的是 384 00:15:11,789 --> 00:15:13,089 the top you see a trigger that means 在頂部，您會看到一個觸發器，這意味著 385 00:15:13,289 --> 00:15:14,620 it's triggering off the master once you 一旦你觸發了主人 386 00:15:14,820 --> 00:15:18,279 check in and here you have individual 簽入，這裡有個人 387 00:15:18,480 --> 00:15:20,199 steps this in fact does a build of 步驟這實際上是建立 388 00:15:20,399 --> 00:15:23,829 container and we do three builds here 容器，我們在這裡進行三個構建 389 00:15:24,029 --> 00:15:26,039 because that's what cube flow requires 因為那是多維資料集流所需要的 390 00:15:26,240 --> 00:15:28,809 cube flow needs every one of your steps 多維資料集流需要您執行的每一步 391 00:15:29,009 --> 00:15:31,089 to be built into a container and then 放在一個容器裡然後 392 00:15:31,289 --> 00:15:32,939 you roll that for 你為 393 00:15:33,139 --> 00:15:34,829 and into your cute flow pipeline and 並進入您可愛的流程管道， 394 00:15:35,029 --> 00:15:37,469 what you're rolling in what the pipeline 您在管道中滾動什麼 395 00:15:37,669 --> 00:15:42,449 looks like is this and actually why 看起來是這樣，實際上是為什麼 396 00:15:42,649 --> 00:15:43,879 don't I get to this in just a second so 我不是一秒鐘就能做到這一點嗎 397 00:15:44,080 --> 00:15:48,329 we'll go back to our video here and like 我們將在這裡回到我們的視頻 398 00:15:48,529 --> 00:15:52,859 I said what you see is a standard repo 我說你看到的是標準回購 399 00:15:53,059 --> 00:15:56,969 or is my thing here and that was just 還是我在這裡的東西，那僅僅是 400 00:15:57,169 --> 00:15:58,709 the code that you saw me pointing at 您看到我指向的代碼 401 00:15:58,909 --> 00:16:01,289 earlier with standard pipelines and this 早期使用標準管道，這 402 00:16:01,490 --> 00:16:03,779 is a very important one it does burritos 是一個非常重要的墨西哥卷餅 403 00:16:03,980 --> 00:16:05,909 versus tacos you upload and it'll tell 與您上傳的炸玉米餅，它將告訴您 404 00:16:06,110 --> 00:16:06,870 you whether or not something Sabrina 你是否有薩布麗娜的東西 405 00:16:07,070 --> 00:16:08,519 virgin taco we're just a hardware heart 原始的炸玉米餅，我們只是硬體的心臟 406 00:16:08,720 --> 00:16:10,289 problem that you would think you know a 你以為自己知道的問題 407 00:16:10,490 --> 00:16:13,229 half-open burrito what you get the idea 半開卷餅你有什麼主意 408 00:16:13,429 --> 00:16:15,659 all right so we moving on here and you 好吧，我們繼續在這裡和你 409 00:16:15,860 --> 00:16:17,309 can see what it's doing is this code 可以看到這段代碼在做什麼 410 00:16:17,509 --> 00:16:19,289 that's in the repo right now is doing 現在在回購中 411 00:16:19,490 --> 00:16:21,569 transfer learning using very standard 使用非常標準的轉移學習 412 00:16:21,769 --> 00:16:23,459 transfer learning tool called mobile net 轉移學習工具稱為移動網 413 00:16:23,659 --> 00:16:28,579 and so we layer on top of that so 所以我們在此之上 414 00:16:28,779 --> 00:16:31,169 like I said this is just incredibly 就像我說的那樣 Incredibly:so much can’t believe 415 00:16:31,370 --> 00:16:34,289 pipeline code and these are the pipeline 管道代碼，這些是管道 416 00:16:34,490 --> 00:16:35,849 steps that you'll see here and I've into 您將在這裡看到的步驟，我已經進入 417 00:16:36,049 --> 00:16:38,099 those in just a second but the idea is 那些只是一秒鐘，但想法是 418 00:16:38,299 --> 00:16:41,849 that we want to show how when you check 我們想告訴您如何檢查 419 00:16:42,049 --> 00:16:43,139 in it's going to kick all these things 在其中踢所有這些東西 Kick:throw 420 00:16:43,340 --> 00:16:45,750 off so we come back here and we're gonna 離開，所以我們回到這裡，我們要 421 00:16:45,950 --> 00:16:52,259 hit merge on our pull request and it's 在我們的拉取請求上點擊合併， Merge:put somethings together Request:要求 pull request:ask something from server 422 00:16:52,460 --> 00:16:53,279 gonna kick off there we go 要去那裡開始 kick off:to start 423 00:16:53,480 --> 00:16:57,120 so we're hitting merge and off we go 所以我們要合併然後離開 424 00:16:57,320 --> 00:16:59,759 and there you go so this is the UI for a 然後去，所以這是 425 00:16:59,960 --> 00:17:02,490 DevOps again all we did was hit merge DevOps我們所做的一切再次被合併 426 00:17:02,690 --> 00:17:05,309 and it all this kicked off so at this 這一切開始了 427 00:17:05,509 --> 00:17:07,200 point we're gonna go forward you can see 點，我們要前進，你可以看到 428 00:17:07,400 --> 00:17:09,059 there right there build number nine has 那邊有九號樓 429 00:17:09,259 --> 00:17:11,549 just kicked off and that is you know May 剛剛開始，那就是你知道的梅 430 00:17:11,750 --> 00:17:13,649 21st so you know I'm being honest here 21號所以你知道我在這裡很誠實 21st:20+first Honest: onest 431 00:17:13,849 --> 00:17:15,059 and you can see how the things are 你會看到事情如何 432 00:17:15,259 --> 00:17:17,339 aspiring here so in this case it's going 嚮往這裡，所以在這種情況下 Aspiring:to dream for something 433 00:17:17,539 --> 00:17:18,960 to go through those three build steps 完成這三個構建步驟 434 00:17:19,160 --> 00:17:20,220 where it builds each one of the 它在其中構建每個 435 00:17:20,420 --> 00:17:22,740 containers using standard as your build 使用標準容器的容器 436 00:17:22,940 --> 00:17:25,499 tools and again this can be docker this 工具，這又可以是docker this 437 00:17:25,699 --> 00:17:26,759 could be builder this could be anything 可能是建設者，這可能是任何事情 438 00:17:26,959 --> 00:17:29,399 that makes sense for you and you've all 這對您有意義，而您所有人 makes sense:有道理 439 00:17:29,599 --> 00:17:31,139 seen docker build many many times so I'm 看到碼頭工人建造了很多次，所以我 440 00:17:31,339 --> 00:17:35,220 going to fast forward through those okay 快點走過去 441 00:17:35,420 --> 00:17:37,589 and then this last step is where you 然後這最後一步是您 442 00:17:37,789 --> 00:17:39,809 take all those containers and now you've 拿走所有這些容器，現在您已經 443 00:17:40,009 --> 00:17:42,899 uploaded them to in this case a CR but 在這種情況下將它們上傳到CR， 444 00:17:43,099 --> 00:17:45,930 it could be any docker registry okay 可能是任何Docker註冊表都可以 445 00:17:46,130 --> 00:17:48,598 so now that is done and now we have the 所以現在完成了，現在我們有了 446 00:17:48,798 --> 00:17:50,129 interesting bit because that is the 有趣的是，因為那是 Bit:small thing 447 00:17:50,329 --> 00:17:52,559 build pipeline our steps here are first 建立管道我們的步驟是首先 448 00:17:52,759 --> 00:17:54,419 off a build pipeline where it takes 離開需要的構建管道 449 00:17:54,619 --> 00:17:57,509 those containers them and then we do 這些容器，然後我們做 450 00:17:57,710 --> 00:17:59,129 what's called a release pipeline and you 所謂的發佈管道，你 Release:to let something go 451 00:17:59,329 --> 00:18:01,049 want those to be separate right you want 想要那些分開的權利 452 00:18:01,250 --> 00:18:02,250 to make sure the build is complete that 確保構建完成 453 00:18:02,450 --> 00:18:04,169 all the artifacts pass and already go 所有的文物都過去了，已經過去了 454 00:18:04,369 --> 00:18:05,940 and now we're going to do this release 現在我們要發佈此版本 455 00:18:06,140 --> 00:18:07,829 pipeline and the release pipeline is 管道和發佈管道是 456 00:18:08,029 --> 00:18:11,279 where it reaches out to cube flow and 它伸向立方體流動的地方 reaches out:try to catch something from far away 457 00:18:11,480 --> 00:18:13,680 executes and you can see this here it's 執行，您可以在這裡看到它 Executes:.exe 458 00:18:13,880 --> 00:18:16,769 reaching out using a swagger client that 使用招搖的客戶 459 00:18:16,970 --> 00:18:19,139 we wrote to connect directly to the cue 我們寫了直接連接到提示 460 00:18:19,339 --> 00:18:20,789 flow API you can see there it's 流API，您可以看到它的 461 00:18:20,990 --> 00:18:24,479 forwarding the connection and here you 轉發連接，在這裡您 462 00:18:24,679 --> 00:18:26,098 can see the run these are the rough 可以看到運行這些是粗糙的 463 00:18:26,298 --> 00:18:28,769 he had previously we run 84 we had a lot 他以前有我們跑84我們有很多 464 00:18:28,970 --> 00:18:31,019 of debugging not cue flows fault our 調試不提示流故障我們 Fault:mistake 465 00:18:31,220 --> 00:18:33,429 fault 故障 466 00:18:34,390 --> 00:18:37,858 and presto we're at run eighty-five 前提是我們現在快八十五了 Presto(口語):immediately馬上 467 00:18:38,058 --> 00:18:41,430 right so again I want to call back 對，所以我想再次致電 468 00:18:41,630 --> 00:18:43,108 because this is a small but special 因為這很小但是很特別 469 00:18:43,308 --> 00:18:45,180 thing all I did was kick hit get all I 我所做的就是踢得到我所有 470 00:18:45,380 --> 00:18:46,919 did was check in and merge my final PR 做的是檢查併合並我的最終PR 471 00:18:47,119 --> 00:18:48,799 and it kicked off all of this goodness 它開始了所有的美好 472 00:18:49,000 --> 00:18:51,389 so now we were sending through here and 所以現在我們通過這裡發送 473 00:18:51,589 --> 00:18:54,119 you can see a rerunning know okay so 你可以看到一個重新運行的知道，所以 Rerunning:run again 474 00:18:54,319 --> 00:18:56,579 this is a standard cue put pipeline I'd 這是我想要的標準提示放置管道 475 00:18:56,779 --> 00:18:58,019 already described it does pre 已經描述過了 476 00:18:58,220 --> 00:18:59,789 processing here it's mounting in Azure 在這裡處理它正在Azure中安裝 477 00:18:59,990 --> 00:19:02,190 blob and you can see it watching the 一滴，你可以看到它看著 478 00:19:02,390 --> 00:19:03,180 logs right there 記錄在那裡 479 00:19:03,380 --> 00:19:05,309 and you can see the artifacts once 你可以一次看到文物 480 00:19:05,509 --> 00:19:06,899 that's done it automatically moves 完成後，它會自動移動 481 00:19:07,099 --> 00:19:09,750 forward the containers been created and 轉發已創建的容器，並 482 00:19:09,950 --> 00:19:11,399 here it's actually doing a training run 這實際上是在進行訓練 483 00:19:11,599 --> 00:19:13,259 right now it's using the latest 現在正在使用最新的 484 00:19:13,460 --> 00:19:16,049 tensorflow 2.0 you can see and not just 你可以看到的不僅僅是tensorflow 2.0 2.0:two point o 485 00:19:16,250 --> 00:19:17,700 tensorflow 2.0 but let me just highlight tensorflow 2.0，但讓我強調一下 486 00:19:17,900 --> 00:19:19,680 that right there it's using GPUs 就在那裡，它正在使用GPU 487 00:19:19,880 --> 00:19:22,559 natively on kubernetes directly through 直接通過kubernetes在本地 488 00:19:22,759 --> 00:19:27,180 this check it ok and it's going to 這項檢查確定，它將 489 00:19:27,380 --> 00:19:29,190 operate and run quite fast much faster 快速運行和運行 490 00:19:29,390 --> 00:19:30,899 than it ran when it was running on my 比它在我的電腦上運行時所運行的 491 00:19:31,099 --> 00:19:36,180 local machine forward there and now it's 本地機器向前走，現在 492 00:19:36,380 --> 00:19:38,190 doing registration and this is where you 做註冊，這是你的地方 493 00:19:38,390 --> 00:19:40,588 start to mix and match between things 開始在事物之間融合和匹配 494 00:19:40,788 --> 00:19:42,000 running in queue flow and things that 在佇列流中運行以及 495 00:19:42,200 --> 00:19:44,608 you may want to run in the cloud so in 您可能想在雲中運行，所以在 496 00:19:44,808 --> 00:19:46,379 this case it's going to do this last 在這種情況下，它將持續到最後 497 00:19:46,579 --> 00:19:49,889 register step in Azure we have a service 在Azure中註冊步驟我們有一項服務 498 00:19:50,089 --> 00:19:51,809 called the model management service this 稱為模型管理服務 499 00:19:52,009 --> 00:19:54,419 is basically a sophisticated storage 基本上是一個複雜的存儲 Sophisticated:very developed 500 00:19:54,619 --> 00:19:56,399 location which understands machine 瞭解機器的位置 501 00:19:56,599 --> 00:19:59,159 learning concepts first-class way 學習概念的一流方法 Concepts:idea 502 00:19:59,359 --> 00:20:01,229 and by that what I mean is you're able 我的意思是說你有能力 503 00:20:01,429 --> 00:20:02,789 to understand whether or not something 瞭解是否有東西 504 00:20:02,990 --> 00:20:05,369 needs versioning how to compare versions 需要版本控制如何比較版本 505 00:20:05,569 --> 00:20:06,960 within each other what the performance 彼此之間的表現 506 00:20:07,160 --> 00:20:08,460 of something is and you're able to 的東西是，你能夠 507 00:20:08,660 --> 00:20:11,819 visualize it in a first-class way and so 以一流的方式視覺化 Visualize:to see something with your eye close 508 00:20:12,019 --> 00:20:13,049 what it's going to do here at this last 這最後要在這裡做什麼 509 00:20:13,250 --> 00:20:15,659 step here are three versions that we ran 這是我們運行的三個版本 510 00:20:15,859 --> 00:20:19,379 already and once it's done you can see 已經，一旦完成，您可以看到 511 00:20:19,579 --> 00:20:22,129 here this is the model registry service 這是模型註冊服務 512 00:20:22,329 --> 00:20:27,428 it's going to kick it off any minute 它會在任何時候開始 any minute:very soon 513 00:20:28,599 --> 00:20:32,069 yeah okay so you can see here it's using 是的，所以你可以看到這裡正在使用 514 00:20:32,269 --> 00:20:35,039 in this case the azure SDK natively 在這種情況下，本機azure SDK 515 00:20:35,240 --> 00:20:37,079 inside of cube flow it's able to reach 在多維資料集流內部，它可以達到 516 00:20:37,279 --> 00:20:38,879 out and you can see they're using our 出去，你可以看到他們正在使用我們的 517 00:20:39,079 --> 00:20:40,829 service principles service passwords the 服務原則服務密碼 Principles:rules 518 00:20:41,029 --> 00:20:42,960 subscription information it's able to 能夠訂閱的資訊 Subscription:訂閱 519 00:20:43,160 --> 00:20:45,539 take that hc5 file and simply push it up 提取該hc5文件，然後將其向上推 520 00:20:45,740 --> 00:20:48,119 to the model registry service which then 到模型註冊服務，然後 521 00:20:48,319 --> 00:20:50,039 will kick off the second half of the 將在下半場開始 522 00:20:50,240 --> 00:20:55,559 solution and you can see there so that's 解決方案，您可以在那裡看到，所以 523 00:20:55,759 --> 00:20:57,690 roughly what it is you know rolling 大概是什麼，你知道滾動 524 00:20:57,890 --> 00:21:00,029 things out here's where you can see some 在這裡可以看到一些東西 525 00:21:00,230 --> 00:21:01,079 of the richness that we're able to 我們能夠做到的豐富 526 00:21:01,279 --> 00:21:02,598 upload and the metadata that we have 上傳以及我們擁有的中繼資料 Metadata:data about data 527 00:21:02,798 --> 00:21:06,990 relative to these things and no I don't 相對於這些東西，不，我不 528 00:21:07,190 --> 00:21:13,108 need you they were real so that's what 需要你，他們是真實的，那就是 529 00:21:13,308 --> 00:21:15,389 model that's what Ops looks like in 這就是ops的樣子 530 00:21:15,589 --> 00:21:16,950 reality right and from that point 現實的權利，從那時起 531 00:21:17,150 --> 00:21:18,089 forward you know I just rolled out a 轉發，你知道我剛剛推出了 532 00:21:18,289 --> 00:21:19,289 service endpoint everyone seen that 每個人都看到的服務端點 533 00:21:19,490 --> 00:21:23,069 right nothing special there but that's 沒什麼特別的，但是那是 534 00:21:23,269 --> 00:21:24,358 what we're talking about this is 我們正在談論的是 535 00:21:24,558 --> 00:21:26,159 something where you had a cube flow 立方體流動的地方 536 00:21:26,359 --> 00:21:29,190 cluster that was in no way connected to 完全沒有連接的集群 537 00:21:29,390 --> 00:21:31,289 you know any of the more sophisticated 你知道任何更複雜的 538 00:21:31,490 --> 00:21:33,598 azor or other components and with just azor或其他組件 539 00:21:33,798 --> 00:21:36,539 running standard kubernetes and standard 運行標準的kubernetes和標準 540 00:21:36,740 --> 00:21:38,819 cube flow and the reality is that 立方體流動，而現實是 541 00:21:39,019 --> 00:21:41,309 gives you the opportunity to do the 給你機會去做 542 00:21:41,509 --> 00:21:42,869 training in a way that makes sense for 進行有意義的培訓 543 00:21:43,069 --> 00:21:45,299 your organization you don't have to take 您的組織，您不必 544 00:21:45,500 --> 00:21:46,769 your data you don't have to change the 您的資料，您不必更改 545 00:21:46,970 --> 00:21:48,450 way your processes work you don't have 您沒有的流程工作方式 546 00:21:48,650 --> 00:21:51,598 to adopt some crazy new standard just 採取一些瘋狂的新標準 Adopt:to take something not yours and raise it 547 00:21:51,798 --> 00:21:53,848 because that looks slightly faster this 因為這看起來快一點 Slightly:a little 548 00:21:54,048 --> 00:21:55,979 is an open platform where you can pick 是一個開放的平臺，您可以在其中選擇 549 00:21:56,179 --> 00:21:57,839 and choose and plug in the components 然後選擇並插入元件 550 00:21:58,039 --> 00:21:59,789 that make sense to you and that's the 對你有意義，那就是 551 00:21:59,990 --> 00:22:01,319 power that things like open source 像開源這樣的東西 552 00:22:01,519 --> 00:22:03,750 and things like Q flow provide now I 現在像Q flow這樣的東西 553 00:22:03,950 --> 00:22:05,219 know what you're saying it sounds like a 知道你在說什麼聽起來像 554 00:22:05,419 --> 00:22:07,529 lot of work I would like to recall this 我想回顧很多工作 Recall:remember 555 00:22:07,730 --> 00:22:09,659 statement here you know what's a lot of 聲明在這裡，你知道很多 556 00:22:09,859 --> 00:22:10,829 work eleven months sitting around 工作十一個月 557 00:22:11,029 --> 00:22:11,940 without your model shipping out to 無需將模型運送到 558 00:22:12,140 --> 00:22:12,399 production 生產 559 00:22:12,599 --> 00:22:15,039 that's a lot of work this you know okay 這是很多工作，你知道嗎 560 00:22:15,240 --> 00:22:16,509 you have to spend a few minutes making 你必須花幾分鐘的時間 561 00:22:16,710 --> 00:22:18,129 sure that service principles are able to 確保服務主體能夠 Principles: 主體 562 00:22:18,329 --> 00:22:20,558 cross clouds but other than that you'll 跨雲，但除此之外， 563 00:22:20,759 --> 00:22:24,698 be fine and you're not just fine 很好，你不只是很好 564 00:22:24,898 --> 00:22:27,549 right you're actually moving faster and 對，實際上您的移動速度更快， 565 00:22:27,750 --> 00:22:29,438 and that's what we're really trying to 這就是我們真正想要的 566 00:22:29,638 --> 00:22:32,318 provide here data scientists understand 提供資料科學家瞭解 data scientists:algorithm 567 00:22:32,519 --> 00:22:36,519 their problem areas but too often they 他們的問題領域，但他們經常 568 00:22:36,720 --> 00:22:38,318 have thrown things over the wall and 把東西扔在牆上， 569 00:22:38,519 --> 00:22:41,169 asked production to ultimately you know 要求生產，最終你知道 Ultimately:in the end 570 00:22:41,369 --> 00:22:43,568 deal with a lot of the hardness what 處理很多硬度什麼 Hardness:difficult thing 571 00:22:43,769 --> 00:22:45,188 we're saying is we want to bring 我們是說我們要帶 572 00:22:45,388 --> 00:22:47,799 through things like ml ops we want to 通過像ml ops這樣的事情，我們想要 573 00:22:48,000 --> 00:22:50,108 bring the same advances that get ops 帶來同樣的進步 574 00:22:50,308 --> 00:22:52,808 brought to app developers we want to 帶給我們想要的應用程式開發人員 575 00:22:53,009 --> 00:22:55,568 help data scientists be part of how to 説明資料科學家成為如何 576 00:22:55,769 --> 00:22:57,789 bring things to production and get into 把東西帶入生產並進入 577 00:22:57,990 --> 00:23:00,039 end ownership and even better than that 最終擁有權甚至比這更好 578 00:23:00,240 --> 00:23:02,469 help them learn real software 説明他們學習真正的軟體 579 00:23:02,669 --> 00:23:04,659 engineering best practices like for too 工程最佳實踐，例如 580 00:23:04,859 --> 00:23:06,308 long we've said hey data 很久以來，我們已經說過嘿資料 581 00:23:06,509 --> 00:23:08,108 scientists you know don't worry about 你認識的科學家不用擔心 582 00:23:08,308 --> 00:23:11,259 you know sweet best practices you know 你知道甜蜜的最佳實踐 583 00:23:11,460 --> 00:23:13,959 come back when you're older no that is 當你長大回來時，那不是 584 00:23:14,159 --> 00:23:16,000 not okay these are these are software 不好，這些是軟體 585 00:23:16,200 --> 00:23:17,740 engineers they want to help get things 他們想要幫助的工程師 586 00:23:17,940 --> 00:23:19,240 into reduction they want their models to 減少他們想要他們的模型 587 00:23:19,440 --> 00:23:21,399 be used just as much as you do we need 盡可能多地使用，我們需要 588 00:23:21,599 --> 00:23:23,709 to help them bridge the gap but not by 幫助他們彌合差距，但不能 bridge the gap: 彌合差距