Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Modern Machine Learning Delivery - ABEJA Platformで実現する、継続的機械学習とデリバリー

ABEJA
March 04, 2019

Modern Machine Learning Delivery - ABEJA Platformで実現する、継続的機械学習とデリバリー

SIX 2019 dev-b-4
Takanori Ishikawa @ABEJA, Inc.

「Modern Machine Learning Delivery - ABEJA Platformで実現する、継続的機械学習とデリバリー」

ABEJA Platformでは出来上がった機械学習モデルのAPIホスティングと運用が簡単にできるだけでなく、スケーラブルなビジネスに発展させるための機能を提供しています。本セッションでは、モデルの個別最適化、Fine Tuning等のABEJA Platformを活用する上での技術ノウハウをお伝えしています。

ABEJA

March 04, 2019
Tweet

More Decks by ABEJA

Other Decks in Technology

Transcript

  1. DAY 1 “技” Developer Day Modern Machine Learning Delivery ABEJA

    Platformで実現する継続的機械学習とデリバリー Takanori Ishikawa (ABEJA, Inc.)
  2. #abejasix “It is not the strongest of the species that

    survives, nor the most intelligent that survives. It is the one that is most adaptable to change.”
  3. “It is not the strongest of the species that survives,

    nor the most intelligent that survives. It is the one that is most adaptable to change.”
  4. “It is not the strongest of the species that survives,

    nor the most intelligent that survives. It is the one that is most adaptable to change.”
  5. "ਓྨͷਐԽ vector γϧΤοτ and outline" © Natasha Sinegina (Licensed under

    CC BY-SA 4.0) 道具を使う ⽕を使う 認知⾰命 農業⾰命 科学⾰命 産業⾰命 情報⾰命
  6. "ਓྨͷਐԽ vector γϧΤοτ and outline" © Natasha Sinegina (Licensed under

    CC BY-SA 4.0) 道具を使う ⽕を使う 認知⾰命 農業⾰命 科学⾰命 産業⾰命 情報⾰命 250万年前 30万年前 7万年前 1万2千年前 500年前 200万年前 70年前
  7. "ਓྨͷਐԽ vector γϧΤοτ and outline" © Natasha Sinegina (Licensed under

    CC BY-SA 4.0) Society 1.0 Society 2.0 Society 3.0 Society 4.0 250万年前 1万2千年前 200万年前 70年前
  8. 世界のモバイルデータトラフィックの推移予測 0 20 40 60 80 2017 2018 2019 2020

    2021 2022 (エクサバイト/⽉間) ग़యɿCisco Visual Networking Index: Global Mobile Data Traffic Forecast Update, 2017–2022 White Paper
  9. ग़యɿThe Internet of Things: a movement, not a market (IHS

    Markit) 世界の IoT デバイス数の 推移及び予測 1250億 2030年 270億 2017年
  10. Πϯλʔωοτར༻ঢ়گͷਪҠ 0% 20% 40% 60% 80% 100% 1997 1999 2001

    2003 2005 2007 2009 2011 2013 2015 2017 ग़యɿʮ௨৴ར༻ಈ޲ௐࠪʯʢ૯຿লʣ
  11. Enabling AI implementation into businesse With AI, we provide various

    solutions such as commodities classification, maintenance support efficiency, behavior analysis of adept workers, and more. We have supported AI implementation for over 150 companies, ranging industries such as manufacturing, infrastructure, logistics, and retail. With our experience and expertise, we provide overall support from AI application to utilization.
  12. AI service provided by ABEJA Enabling AI implementation into businesse

    With AI, we provide various solutions such as commodities classification, maintenance support efficiency, behavior analysis of adept workers, and more. We have supported AI implementation for over 150 companies, ranging industries such as manufacturing, infrastructure, logistics, and retail. With our experience and expertise, we provide overall support from AI application to utilization. ABEJA Platform supports efficient application of Deep Learning in the operation process The ABEJA Insights for retail analyses your custome'sr behavior, such as customer demographic attributes, visitor flow analysis and prediction of repeating customer visits to provide you with a deeper understanding of your customer base. ABEJA has supported leading companies to implement over 100 End-to-End Machine Learning & Deep Learning projects.
  13. Data gathered Integration & Analysis Operate retail shops scientifically, transcending

    ‘experience’ and ‘gut feeling’ ABEJA Insight for Retail Visualization Action Cloud Dashboard What to know ✓ Underperforming vs. benchmark stores ✓ Reason for underperformance ✓ Identify measures and policies for improvement What to do ✓ Identify underperforming stores and improvement plans quickly ✓ Specify know-how of best practices in your shops and expand horizontally in your organization Easy to mange data visualization Camera Image Analysis Utilizing Deep Learning technology POS Data Weekly e-mail report
  14. Product Features Machine Status Customer Demographic Simple UI Unlimited Categories

    Easily Outsourced Image, Video, Text, Voice ABEJA Platform Annotation Utilize to Tag Benefits Supported Data Types Our Annotation tool and outsourcing service provides a GUI tool that makes creating usable data sets that are suitable for Machine and Deep learning very easy .
  15. 1. Data saving 2. Annotation 3. Training 4. Deploy 5.

    Serving 6. Inference 7. Insight ABEJA Platform 1):4*$ "-803-% $ :#&3803-% *P5"DUVBUPS *P54FOTPS &YJTUJOH%BUB 0UIFS4ZTUFN #JH%BUB %FQMPZ .POJUPSJOH &EHF"* *OGFSSJOH $MPVE"* -FBSOJOH*OGFSSJOH *OCPVOE*P5 0VUCPVOE*P5 ABEJA Platform is a platform that provides functions to improve the efficiency of machine learning and deep learning implementation and operation. We provide all necessary functions for pipeline with CLI, SDK, API, GUI and accelerate implementation.
  16. 1. Data saving 2. Annotation 3. Training 4. Deploy 5.

    Serving 6. Inference 7. Insight ABEJA Platform 1):4*$ "-803-% $ :#&3803-% *P5"DUVBUPS *P54FOTPS &YJTUJOH%BUB 0UIFS4ZTUFN #JH%BUB %FQMPZ .POJUPSJOH &EHF"* *OGFSSJOH $MPVE"* -FBSOJOH*OGFSSJOH *OCPVOE*P5 0VUCPVOE*P5 ABEJA Platform is a platform that provides functions to improve the efficiency of machine learning and deep learning implementation and operation. We provide all necessary functions for pipeline with CLI, SDK, API, GUI and accelerate implementation.
  17. • Python 3 • DNN フレームワーク • TensorFlow, MXNet, PyTorch,

    Keras • Docker による環境構築 • カスタム・イメージのサポート ABEJA Platform / Python ⾔語とフレームワーク
  18. ։ൃ؀ڥ͔Βຊ൪؀ڥ΁ͷҾ͖౉͠ データ、モデル、結果のバージョン管理 冗⻑性やGPUリソースの担保、 エッジ側との連携プロセス構築 ⼤量データの取得に必要なAPIや負担分 散の仕組みや準備、セキュリティ担保 教師データの作成に必要なツールと⼈材の準備 データウェアハウス の準備と管理 データのバリエーション(正確性)の確認

    0からのモデル設計 GPU環境の準備 と⾼度な分散化 デプロイ後のモデルの挙動を監視し、 必要に応じてモデルをアップデート データ 取得 データ 蓄積 データ 確認 教師 データ 作成 モデル 設計 学習 評価 デプロイ 推論 再学習 教師 データ 作成 デプロイ 推論 再学習
  19. • TFX (Google) • FBLearner (Facebook) • Michelangelo (Uber) •

    Bighead (Airbnb) 他社事例 • Kubeflow • MLflow • DVC オープンソース
  20. DataLake Dataset Database Notebook Shared File System Training Jobs Code

    Local Training Evaluation Trained Models Deployed Models Metrics Filter and Queue Pipeline in ABEJA Platform
  21. DataLake Dataset Database Notebook Shared File System Training Jobs Code

    Local Training Evaluation Trained Models Deployed Models Metrics Filter and Queue Pipeline in ABEJA Platform
  22. DataLake Dataset Database Notebook Training Jobs Code Local Training Evaluation

    Trained Models Deployed Models Metrics Filter and Queue Shared File System Pipeline in ABEJA Platform
  23. DataLake Dataset Database Notebook Training Jobs Code Local Training Evaluation

    Trained Models Deployed Models Metrics Filter and Queue Shared File System Pipeline in ABEJA Platform
  24. ೔ຊͰ΋൒෼Ҏ্ͷاۀ͕Ϋϥ΢υαʔϏεΛར༻ 0% 50% 100% 2011 2012 2013 2014 2015 2016

    2017 ग़యɿʮ௨৴ར༻ಈ޲ௐࠪʯʢ̐ʣΫϥ΢υίϯϐϡʔςΟϯάαʔϏεͷར༻ঢ়گʢ૯຿লʣ 56%
  25. Tomorrow will be announced in many sessions how the technology

    introduced today is actually used by clients. Please come tomorrow by all means GO Day2 !! - for ABEJA Platform
  26. The contents introduced today and the products and services that

    support the backside of these, We have prepared a booth at the 3F exhibition hall and tell it. Please drop by during the session. GO EXPO 2F 3F Room A Room B Room C Room D Hall ٖؒك٦ة٦ WC ㉀锑 ٕ٦ي ㉀锑 ٕ٦ي ㉀锑 ٕ٦ي ㉀锑 ٕ٦ي Room E ٖؒك٦ة٦ WC ♧菙勻㜥罏「➰ ٝ؟٦ أؙ 闌怴罏 「➰ 1F 2F 3F Floor Maps Room A Room B Room C Room D Hall ٖؒك٦ة٦ WC ㉀锑 ٕ٦ي ㉀锑 ٕ٦ي ㉀锑 ٕ٦ي ㉀锑 ٕ٦ي Room E ٖؒك٦ة٦ WC Room W ♧菙勻㜥罏「➰ أهٝ؟٦ رأؙ 闌怴罏 「➰ WC ٖؒك٦ة٦ Here
  27. After the lecture is over, we are waiting at the

    Ask the Speaker section of the exhibition area. If you have any questions, please come to this corner after the session ends. See you Ask the Speaker !! ABEJA 17 6 5 4 3 1 2 9 10 11 12 7 8 16 15 ABEJA Ask the Speaker 14 3F Hall ABEJAծ ABEJA Deep Learning ABEJA
  28. Thank you! Please give us feedback on this session if

    you like ID of this session dev-b-4 Feedback will be used to develop products and deliver more information https://goo.gl/forms/erEBAsrQK4XKEv352