Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
Deep Learningライブラリ 色々つかってみた感想まとめ
Search
Takanori Ogata
April 17, 2016
Technology
18
17k
Deep Learningライブラリ 色々つかってみた感想まとめ
Takanori Ogata
April 17, 2016
Tweet
Share
More Decks by Takanori Ogata
See All by Takanori Ogata
ACCELStarsピッチ資料
conta
0
110
Convolutional Pose Machines
conta
0
1.3k
Other Decks in Technology
See All in Technology
スイッチのBMC、つかってますか?
sonic
0
500
hacomonoらしさをデザインする
hacomono
PRO
2
110
チェックツールを導入したけど使ってもらえなかった話 #GAADjp
lycorptech_jp
PRO
1
150
Docker Compose で手軽に手元環境を実現する / Simplifying Local Environments with Docker Compose #CinemaDeLT
nabeo
0
290
ITベンダーから見る内製化支援の本質/in-house-dev
slsops
1
180
OpenTelemetry SpanProcessor を Let's カスタマイズ!
phaya72
1
100
WindowsでGenesisに挑戦した話
natsutan
0
130
TypeScriptで実践するクリーンアーキテクチャ ― WebからもCLIからも使えるアプリ設計 / CClean Architecture with Typescript Application
panda_program
10
2.5k
株式会社Awarefy(アウェアファイ)会社説明資料 / Awarefy-Company-Deck
awarefy
3
17k
SaaS公式MCPサーバーをリリースして得た学び
kawamataryo
5
1.4k
golang-migrate VS Atlas !? 技術選定のポイントと学び ~DBマイグレーションツール選定の実例を通して~ / golang-migrate vs Atlas ! What is the point of technology selection and what you can learn from the examples of DB migration tool selection?
nttcom
0
110
Lakehouse в Лемана Тех. От архитектуры до оптимизации
emeremyanina1234
0
480
Featured
See All Featured
Understanding Cognitive Biases in Performance Measurement
bluesmoon
29
1.7k
Thoughts on Productivity
jonyablonski
69
4.6k
I Don’t Have Time: Getting Over the Fear to Launch Your Podcast
jcasabona
32
2.3k
Mobile First: as difficult as doing things right
swwweet
223
9.6k
No one is an island. Learnings from fostering a developers community.
thoeni
21
3.3k
CoffeeScript is Beautiful & I Never Want to Write Plain JavaScript Again
sstephenson
160
15k
CSS Pre-Processors: Stylus, Less & Sass
bermonpainter
357
30k
Become a Pro
speakerdeck
PRO
28
5.3k
Being A Developer After 40
akosma
91
590k
jQuery: Nuts, Bolts and Bling
dougneiner
63
7.7k
Distributed Sagas: A Protocol for Coordinating Microservices
caitiem20
331
21k
VelocityConf: Rendering Performance Case Studies
addyosmani
329
24k
Transcript
Deep LearningϥΠϒϥϦ ৭ʑ͔ͭͬͯΈͨײ·ͱΊ @conta_
Self Introduction ॹํɹول (twitter: @conta_) CTO@ABEJA, Inc. Computer Visionͱ͔ɺMachine LearningΛͬͨ
ϓϩμΫτ։ൃΛ͍ͬͯ·͢ɻ
Deep Learning Library?
None
ʊਓਓਓਓਓਓਓਓਓʊ ʼɹଟ͗ͯͭ͢Β͍ɹʻ ʉY^Y^Y^Y^Y^Y^Y^Yʉ
Dive into Deep Learning
ˎײ͡ํʹݸਓ͕ࠩ͋Γ·͢
ࠓճհ͢ΔϥΠϒϥϦ
Caffe Caffe: UC Berkleyͷਓ͕࡞ͬͯΔɻDeepLearningք۾Ͱ͔ͳΓฮͳϥΠϒϥϦͰɺ ޭେ͖͍ ݴޠ: ɾCoreC++ɻPython, MatlabͷWrapper͕͋Δ ಛ: ɾجຊతʹProtocol
BufferͰωοτϫʔΫΛهड़
▪͍͍ͱ͜Ζ(ݸਓతײ) ɾModel ZooʹֶशࡁΈϞσϧ͕ͨ͘͞Μެ։͞ΕͯΔ ʢطʹCVPR2016ͷจͷϞσϧެ։͞ΕͯΔʣ ɾݚڀऀׂ͕ͱͬͯΔͷͰ࠷৽ͷݚڀՌ͕CaffeͰ࣮͞ΕͯͨΓ͢Δ ɾMulti-GPUʹରԠͨ͠ͷͰɺઃఆ̍ͭͰෳͷGPUΛར༻Մೳ ɾ࣮ߦׂ͕Γͱૣ͍ ɾωοτϫʔΫͷύϑΥʔϚϯεςετ͕Ͱ͖Δ(caffe testίϚϯυ)
▪ͭΒ͍ͱ͜Ζ(ݸਓతײ) ɾΧελϚΠζ͕c++ͱProtocol BufferɻɻɻϚξͭΒ͌Ηɻɻɻʢˎ̍ʣ =>ਓ͕ΧελϚΠζͨ͠ͷɺΘ͔ΒΜɻ ɾωοτϫʔΫΛProtocol BufferͰॻ͘ͷ͕ͭΒ͍ʢˎ̎ʣ =>GoogLeNet2000ߦɺResNet7000ߦɻɻɻ ʢProtocol Buffer৬ਓܳʣ ɾσʔληοτΛ࡞͢Δͷ͕େม
ɾΤϥʔ͕Θ͔Γʹ͍͘ ɾιʔείʔυΛಡΊͳ͍ͱશػೳ͑ͳ͍ɺಈ͖͕Θ͔Βͳ͍ ʢυΩϡϝϯτߋ৽͠Ζʂʣ ɾΠϯετʔϧ͕ͭΒ͍ ʢੲʹൺΔͱґଘؔͷOnOffͷΦϓγϣϯ͕͍ͨͨΊɺ ͍ͩͿϚγʣ ɾRNNΛѻ͏͜ͱͰ͖ͳ͍ʢຐվ͞ΕͨCaffeϕʔεͷͷ͋Δ͚Ͳɻɻɻʣ
▪༨ஊʢˎ̍ʣ ɾ࠷ۙPython Layer͕Ճ͞ΕͯPython͚ͩͰΧελϚΠζ Ͱ͖ΔΑ͏ʹຐվ͍ͯ͠Δ(No Documentation)
▪༨ஊʢˎ̎ʣ ɾPythonͰProtocol BufferΛੜ Ͱ͖ΔΑ͏ʹͳͬͨͨΊɺ ϧʔϓͨ͠هड़ׂ͕ͱ؆୯ʹͳͬͨ (No Documentation)
▪͜Μͳͻͱʹ͓͢͢Ί ɾ·ͣԿ͔ಈ͔͍ͨ͠ਓ ɾͱΓ͋͑ͣݚڀՌΛࢼ͍ͨ͠ਓ ɾ͕ඞཁͳਓ ɾC++ͱProtocol BufferΛษڧ͍ͨ͠ਓ ɾࠜؾڧ͘Կ͔ͱઓ͍͍ͨਓ
Tensorflow: G̋̋gleͷࢄߦྻܭࢉϥΠϒϥϦɻ ผʹDeep͚ͩ͡Όͳ͍Μ͔ͩΒͶʂ ݴޠ: ɾCoreC++ɻPythonͱC++ͲͪΒͰಈ͘ɻ ಛ: ɾࢄॲཧ͕؆୯ʹͰ͖Δ ɾGoogleͷϓϩμΫτͰԿར༻͞Ε͍ͯͯɺ҆ఆײ͕͋Δ
▪͍͍ͱ͜Ζ(ݸਓతײ) ɾࢄॲཧ͕ΊͬͪΌ؆୯ʹͰ͖Δ(Distributed Tensorflow) ɾGoogle͕MLϓϥοτϑΥʔϜΛఏڙ։࢝ ɾ࠷ۙɺTensorflow͍·ͨ͠จ͕Α͘Ͱ͖͍ͯͯΔ ɾίΞ͕C++ͳͷͰAndroidͰಈ࡞͢Δ ɾDocker Container͕མͪͯΔͷͰɺDocker͑ΔͳΒ ΠϯετʔϧʹࠔΒͳ͍ ɾTensorboard͕ΦγϟϨ
▪ͭΒ͍ͱ͜Ζ(ݸਓతײ) ɾݰਓ͚ϥΠϒϥϦ =>Έ͕ͪΐͬͱෳࡶͳͷͰཧղ͠ͳ͍ͱ͍͜ͳͤͳ͍ =>ωοτϫʔΫΛॻ͘ͷʹҰ͔Βهड़͢Δඞཁ͕͋ΔɺTheanoతͳཱͪҐஔ ɾιʔείʔυ͕େنͳͨΊվ͕େมͦ͏ ʢҰԠυΩϡϝϯτ͋Δ͚Ͳʣ ɾDistributed TensorflowΛݸਓͷࢿݯͰ׆༻͢ΔͷࠔͳͷͰɺGoogleͷϓϥοτ ϑΥʔϜΛΘͳ͍ͱԸܙΛड͚ʹ͍͘ =>ࢄίϯϐϡʔςΟϯάͷIOϘτϧωοΫɺInfiniBandΛ͍ͬͺ͍ങ͑Δ͓ۚ࣋ͪ
ͳΒԸܙΛड͚ΒΕΔ͔
▪͜Μͳͻͱʹ͓͢͢Ί ɾΈͷ෦͔ΒDeep LearningΛษڧ͍ͨ͠ਓ ɾDeep Learningɹதʙ্ڃऀ͚ͷਓ ɾେنػցֶशΛͬͯΈ͍ͨਓ ɾେنػցֶशج൫Λ࡞Γ͍ͨਓ ɾMobileʹΈࠐΈ͍ͨਓ
Chainer: PFNͷDeep LearningϥΠϒϥϦɻ ݴޠ: ɾPython(+Cuda) ಛ: ɾDefine-by-Runͱ͍͏ख๏Λͱ͍ͬͯͯɺωοτϫʔΫΛޙ͔Βղ ऍ ɾ͢Β͍͠
▪͍͍ͱ͜Ζ(ݸਓతײ) ɾωοτϫʔΫͷهड़ͷॊೈੑ͕ߴ͍ ʢಛʹRNNܥඇৗʹॻ͖͍͢ʣ ɾ෦ͷಈ࡞͕Ͳ͏ͳͬͯΔ͔ඇৗʹΘ͔Γ͍͢ ɾσόοΫ͍͢͠ ɾφ͍ΞϧΰϦζϜ͕͍ͪૣ࣮͘͞ΕͯΔ ɾCupyͱ͍͏Cuda͕؆୯ʹ͑ΔߦྻԋࢉϥΠϒϥϦؚ͕·Ε͍ͯ ͯɺࣗલͷΞϧΰϦζϜΛൺֱత؆୯ʹߴԽͰ͖Δ (C++Ͱॻ͍ͯϥούʔͱ͔ͭ͘Βͳ͍͍ͯ͘) ɾதͷਓ͕͍͢͝
▪ͭΒ͍ͱ͜Ζ(ݸਓతײ) ɾωοτϫʔΫҎ֎ͷهड़ྔ͕ଟ͘ͳͬͯ͠·͏ʢֶशͷίʔυͱ͔ʣ ɾ࣮ߦʢ࠷ۙͦͦ͜͜ૣ͍ͬΆ͍ʣ ɾDeep Learning͔ͬͯͳ͍ͱଟ͍͜ͳͤͳ͍
▪͜Μͳͻͱʹ͓͢͢Ί ɾDeep LearingΛҰ͔ΒΨοπϦษڧ͍ͨ͠ਓ ɾDeep Learningɹதʙ্ڃऀͷਓ ɾݚڀͰTry and ErrorΛ܁Γฦ͠ͳ͕ΒΞϧΰϦζϜΛ։ൃ͍ͨ͠ਓ ɾෳࡶͳωοτϫʔΫΛهड़͍ͨ͠ਓ ʢωοτϫʔΫͰ݅จॻ͖͍ͨɺσʔλʹΑͬͯॲཧΛ͚͍ͨʣ
ɾRNNͱ͔NLPͱ͔Λॻ͖͍ͨ
▪MXNet: DMLC(Distributed (Deep) Machine Learning Community)͕࡞ͬͯ ΔɻXGBoostͷ࡞ݩͱͯ͠༗໊ɻ ▪ݴޠ: ɾCoreC++ɻWrapper͕ͨ͘͞Μ͋ΓɺPythonɺC++ɺScalaɺ RɺMatlabɺJuliaͱଟݴޠରԠɻ
▪ಛ: ɾଟݴޠʂ ɾ͕͔ͳΓૣ͍ʢॴײʣ ɾmshadow(ߦྻԋࢉ)ɺps-lite(ࢄॲཧ)ͷϥΠϒϥϦ͕ϕʔε
▪͍͍ͱ͜Ζ(ݸਓతײ) ɾࢄॲཧ(1Node, Multi-GPUɺMulti-NodeɺMulti-GPUͲͪΒ ʣ͕ΊͬͪΌ؆୯ʹͰ͖Δ(Example͋Γ) ɾS3ϞσϧσʔλΛอଘ͢Δػೳ͕͋Δ ɾૣ͍ʢImageNet full datasetΛGeForce GTX 980*4Ͱ8.5)
ɾͳͥૣ͍͔͕υΩϡϝϯτͰྗઆ͞Ε͍ͯΔ ɾଟ࠷ଟݴޠ͕ਐΜͰ͍Δ ɾC++Ͱॻ͔ΕͯΔͷͰ Mobile(iOS, Android)Ͱಈ͘
▪ͭΒ͍ͱ͜Ζ(ݸਓతײ) ɾΤϥʔ͕Θ͔Γʹ͍͘ɺຊʹΘ͔Γʹ͍͘ ɾυΩϡϝϯτ͕গͳ͍ =>ಛघͳֶशσʔλΛ࡞ͬͨΓ͢Δͷେม =>͍͜͠ͱΛ͠Α͏ͱ͢ΔͱιʔεΛಡ·ͳ͚ΕͳΒͳ͍
▪͜Μͳͻͱʹ͓͢͢Ί ɾDeep Learningɹதʙ্ڃऀ͚ͷਓ ɾΛٻΊ͍ͯΔਓ ɾPythonɺC++Ҏ֎Ͱར༻͍ͨ͠ਓ
▪Keras: PythonͷDeep LearningϥΠϒϥϦɻ ࠷ۙv1.0͕ϦϦʔε͞Εͨɻ ▪ݴޠ: ɾPython ▪ಛ: ɾTorchʹࣅͨهड़ํ๏ɻ ɾߦྻԋࢉͷόοΫΤϯυTheanoͱTensorFlowΛར༻͍ͯͯ͠ɺ Γସ͑Δ͜ͱ͕Ͱ͖Δ
▪͍͍ͱ͜Ζ(ݸਓతײ) ɾωοτϫʔΫهड़͕؆୯ɺॊೈ ϕʔεͷAPI͕ͨ͘͞Μ४උ͞Ε͍ͯΔͨΊɺهड़ྔগͳ͘ࡁ Ήɻ؆୯ͳωοτϫʔΫͰ͋ΕAPIΈ߹ΘͤͰͳΜͱ͔ͳΔɻ v1.0.0͔Β functional APIͳΔͷ͕ग़དྷͯɺ ඇৗʹײతʹωοτϫʔΫΛهड़Ͱ͖ΔΑ͏ʹͳͬͨ ɾֶश͕؆୯ ScikitͷΑ͏ʹfit()ؔݺͼग़ͤΑΖͬͯ͘͘͠ΕΔ
ɾιʔε͕ಡΈ͍͢
▪ͭΒ͍ͱ͜Ζ(ݸਓతײ) ɾMulti-GPUඇରԠ TheanoΛBackendͱͯͬͯ͠ΔͱMulti-GPUͭΒ͍ɻ Tensorflowͷ͓͔͛ͰMulti-GPU͕؆୯ʹͰ͖ΔΑ͏ʹͳͬͨʁ ɾPython͔͠ରԠ͍ͯ͠ͳ͍
▪͜Μͳͻͱʹ͓͢͢Ί ɾDeep LearningΛΓ͍ͨਓશൠ ɾ͋·Γࡉ͔͍͜ͱؾʹͤͣʹαΫοͱωοτϫʔΫΛ࡞ Γ͍ͨਓ ˎݸਓతʹҰ൪͓͢͢Ί
·ͱΊ ▪Caffe ɾͱΓ͋͑ͣDeep LearingʢCNNʣΓ͍ͨਓ ɾݚڀՌΛࢼ͍ͨ͠ਓ ▪Tensorflow ɾࢄίϯϐϡʔςΟϯάΓ͍ͨਓ ▪Chainer ɾΞϧΰϦζϜ։ൃ͍ͨ͠ਓ ɾຊؾͰDeep
LearningΛษڧ͍ͨ͠ਓ ▪MXNet ɾ͕ඞཁͳਓ ɾMobileͰಈ͔͍ͨ͠ਓ ▪Keras ɾͱΓ͋͑ͣDeep Learingษڧ͍ͨ͠ਓ ɾΊΜͲ͍͘͞ͷͰ͋Δఔڥ͕४උ͞Ε͍ͯͯཉ͍͠ͱࢥ͏ਓ
We are hiring! → https://www.wantedly.com/companies/abeja