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
Neural Network Consoleについて
Search
Atsushi Nakatsugawa
PRO
September 25, 2018
Technology
0
72
Neural Network Consoleについて
Sony社の提供するNeural Network Consoleについて
https://dl.sony.com
Atsushi Nakatsugawa
PRO
September 25, 2018
Tweet
Share
More Decks by Atsushi Nakatsugawa
See All by Atsushi Nakatsugawa
DevRelの基礎と戦略
moongift
PRO
0
18
DevRelの始め方
moongift
PRO
3
520
マンガで分かるDevRelオンライン
moongift
PRO
0
110
DevRel Online with Manga
moongift
PRO
0
65
DevRel 4コマ
moongift
PRO
0
100
DevRel 4 panel Manga
moongift
PRO
0
62
DevRelの基礎〜開発者マーケティング〜
moongift
PRO
0
25
DevRelの基礎〜DevX(開発者体験)〜
moongift
PRO
0
35
DevRelの基礎「DevRelの重要性」
moongift
PRO
0
37
Other Decks in Technology
See All in Technology
開発生産性を上げながらビジネスも30倍成長させてきたチームの姿
kamina_zzz
2
1.7k
適材適所の技術選定 〜GraphQL・REST API・tRPC〜 / Optimal Technology Selection
kakehashi
1
400
VideoMamba: State Space Model for Efficient Video Understanding
chou500
0
190
複雑なState管理からの脱却
sansantech
PRO
1
140
DMARC 対応の話 - MIXI CTO オフィスアワー #04
bbqallstars
1
160
初心者向けAWS Securityの勉強会mini Security-JAWSを9ヶ月ぐらい実施してきての近況
cmusudakeisuke
0
120
ドメインの本質を掴む / Get the essence of the domain
sinsoku
2
150
ISUCONに強くなるかもしれない日々の過ごしかた/Findy ISUCON 2024-11-14
fujiwara3
8
870
B2B SaaSから見た最近のC#/.NETの進化
sansantech
PRO
0
780
Engineer Career Talk
lycorp_recruit_jp
0
170
【若手エンジニア応援LT会】ソフトウェアを学んできた私がインフラエンジニアを目指した理由
kazushi_ohata
0
150
CysharpのOSS群から見るModern C#の現在地
neuecc
2
3.3k
Featured
See All Featured
XXLCSS - How to scale CSS and keep your sanity
sugarenia
246
1.3M
Adopting Sorbet at Scale
ufuk
73
9.1k
Visualization
eitanlees
145
15k
Practical Orchestrator
shlominoach
186
10k
Cheating the UX When There Is Nothing More to Optimize - PixelPioneers
stephaniewalter
280
13k
Why Our Code Smells
bkeepers
PRO
334
57k
StorybookのUI Testing Handbookを読んだ
zakiyama
27
5.3k
Git: the NoSQL Database
bkeepers
PRO
427
64k
Site-Speed That Sticks
csswizardry
0
25
RailsConf & Balkan Ruby 2019: The Past, Present, and Future of Rails at GitHub
eileencodes
131
33k
What’s in a name? Adding method to the madness
productmarketing
PRO
22
3.1k
GitHub's CSS Performance
jonrohan
1030
460k
Transcript
Neural Network Console ϋϯζΦϯ 2018.09.19
NNCͱʁ جૅࣝ
ιχʔͷఏڙ͢ΔσΟʔϓϥʔχϯά։ൃڥ ▸ Windows ൛ & Ϋϥυ൛ ▸ ίʔσΟϯάϨε / ίϯϙʔωϯτͷυϥοά&υϩοϓ
▸ ੜͨ͠ϞσϧϥΠϒϥϦʢOSSʣΛ༻͍ͯϓϩάϥϜ͔ Βར༻Մೳ https://dl.sony.com Neural Network Consoleͱʁ
Neural Network Consoleͱʁ Neural Network LibrariesͷωοτϫʔΫΛGUIͰߏஙɺ࣮ߦͰ͖Δڥ ▸ ΫϥυWindowsΞϓϦͰGUIΛͬͯσΟʔϓϥʔχϯάΛߏங ▸ τϨʔχϯάσʔλɺςετσʔλΛ࣮ͬͯࡍʹ࣮ߦɺධՁ
▸ Ϟσϧ + PythonʹΑΔωοτϫʔΫΛهड़ͨ͠ίʔυΛμϯϩʔυʢWindowsʣ ▸ ϞσϧͱωοτϫʔΫΛҰͭʹͨ͠ϑΝΠϧΛμϯϩʔυʢΫϥυʣ ▸ ͲͪΒPython/C++/C͔Βར༻Մೳ
Neural Network Librariesͱʁ ιχʔͷ։ൃͨ͠σΟʔϓϥʔχϯάιϑτΣΞ ▸ Φʔϓϯιʔεͱͯ͠ެ։ʢGitHubʣ https://github.com/sony/nnabla ▸ ϥΠηϯεApache 2.0
License ▸ Python/C++/CͰσΟʔϓϥʔχϯάΛهड़Մೳ ▸ ONNXͷग़ྗʹରԠ
ྉۚମܥ WINDOWS / Ϋϥυ൛ͰҟͳΓ·͢ ▸ Windows ൛ແྉ ࡞ͨ͠Ϟσϧͷ༻ར༻Մ ▸ Ϋϥυ൛ैྔ՝ۚ
CPU : 85ԁ/࣌ؒ ɹGPU : 210ԁ/࣌ؒʙ ετϨʔδ : 200ԁ/10GBʙ
ར༻Ͱ͖Δٕज़ ը૾͚ͩͰͳ͘ɺςΩετʹ ▸ ͋Β͔͡ΊςΩετͷ୯ޠΛΠϯσοΫεԽʢ·ͨϕΫτϧ ܥྻʣʹมͯ͠ར༻ ▸ ෳཁૉͷೖྗΛֶश͠ɺΧςΰϦ͚ʢྫɿ͋Ίͷڭࢣ͋ ΓσΟʔϓϥʔχϯάʣ ▸ Ԡ༻࣍ୈͰखॻ͖ೝࣝɺମೝࣝɺϔουδΣενϟʔͳͲ
σΟʔϓϥʔχϯά։ൃͷਏ͍ͱ͜Ζ PythonʹΑΔίʔσΟϯάඞਢ ▸ σΟʔϓϥʔχϯά։ൃͱϓϩάϥϚʔҟͳΔεΩϧ͕ඞཁ ▸ σΟʔϓϥʔχϯάʹਖ਼ղͳ͘ɺ݅Λม͑ͯࢼ͢΄͔ͳ͍ ▸ ࢼ͍ͯ͘͠தͰɺਫ਼Λൺֱͨ͠ΓɺաڈͷωοτϫʔΫΛฤू͍ͨ͠ ▸ Ͱ͖͕͋ͬͨϞσϧɺωοτϫʔΫΛͳΔ͘ίʔσΟϯάϨεʹΈࠐΈ͍ͨ
χϡʔϥϧωοτϫʔΫͱʁ جૅࣝ
ͷΈΛίϯϐϡʔλ্Ͱ࠶ݱ ਓؒͷΑ͏ʹࢥߟ͢Δػցֶश ▸ ͷਆܦճ࿏ʢχϡʔϥϧωοτϫʔΫʣΛݩʹઃܭ ▸ େྔͷσʔλΛύλʔϯͱֶͯ͠शͤ͞ɺͦΕҎ֎ͷ݅ʹ͓͍ͯߴਫ਼ʹ ผ͢Δ ▸ ڭࢣ͋Γσʔλͱڭࢣͳ͠σʔλ
χϡʔϥϧωοτϫʔΫͷ֓ཁ ೖྗɺதؒʢӅΕʣɺग़ྗ ϊʔυ
χϡʔϥϧωοτϫʔΫͷ֓ཁ ೖྗɺதؒʢӅΕʣɺग़ྗ ೖྗ தؒ ग़ྗ
χϡʔϥϧωοτϫʔΫͷ֓ཁ ೖྗɺதؒʢӅΕʣɺग़ྗ ೖྗ தؒ ग़ྗ Τοδ ॏΈ
χϡʔϥϧωοτϫʔΫͷ֓ཁ ೖྗɺதؒʢӅΕʣɺग़ྗ ೖྗ தؒ ग़ྗ Τοδ ॏΈ N1 × ॏΈ1
+ N2 × ॏΈ2 + … → ׆ੑԽؔͰม
׆ੑԽؔ γάϞΠυؔɺۂઢؔɺRELUؔͳͲ ը૾Ҿ༻ɿʲReLU, PReLU, γάϞΠυ etc...ʳχϡʔϥϧωοτͰΑ͘͏׆ੑԽؔͷޮՌΛKerasͰௐͯΈͨ - ϓϩΫϥγετ
ॱ → ٯ ޡࠩؔʢަࠩΤϯτϦϐʔೋޡࠩʣͰޡࠩΛٻΊΔ ೖྗ தؒ ग़ྗ ॏΈΛߋ৽
ޯ߱Լ๏ ॏΈ ޡࠩ
ہॴ࠷దʹؕΔ߹ ॏΈ ޡࠩ ϛχόονʢֶशσʔλΛׂʣͰ͛Δ
σΟʔϓϥʔχϯά ଟ͘ͷதؒΛ࣋ͭ͜ͱͰɺΑΓෳࡶͳʹରԠ ೖྗ தؒ ग़ྗ
ΈࠐΈχϡʔϥϧωοτϫʔΫ ը૾ೝࣝͰΑ͘ΘΕΔख๏ 28x28 px άϨʔεέʔϧ ΈࠐΈ ϓʔϦϯά ΈࠐΈ ϓʔϦϯά શ݁߹
׆ੑԽؔ ग़ྗ
ΈࠐΈχϡʔϥϧωοτϫʔΫ ը૾ೝࣝͰΑ͘ΘΕΔख๏ 28x28 px άϨʔεέʔϧ ΈࠐΈ ϓʔϦϯά ΈࠐΈ ϓʔϦϯά શ݁߹
׆ੑԽؔ ग़ྗ ը૾ʹϑΟϧλΛ͔͚ͯ ಛΛநग़
ΈࠐΈχϡʔϥϧωοτϫʔΫ ը૾ೝࣝͰΑ͘ΘΕΔख๏ 28x28 px άϨʔεέʔϧ ΈࠐΈ ϓʔϦϯά ΈࠐΈ ϓʔϦϯά શ݁߹
׆ੑԽؔ ग़ྗ ൣғͷ࠷େฏۉ Λऔͬͯѹॖ
ΈࠐΈχϡʔϥϧωοτϫʔΫ ը૾ೝࣝͰΑ͘ΘΕΔख๏ 28x28 px άϨʔεέʔϧ ΈࠐΈ ϓʔϦϯά ΈࠐΈ ϓʔϦϯά શ݁߹
׆ੑԽؔ ग़ྗ
ΈࠐΈχϡʔϥϧωοτϫʔΫ ը૾ೝࣝͰΑ͘ΘΕΔख๏ 28x28 px άϨʔεέʔϧ ΈࠐΈ ϓʔϦϯά ΈࠐΈ ϓʔϦϯά શ݁߹
׆ੑԽؔ ग़ྗ
ΈࠐΈχϡʔϥϧωοτϫʔΫ ը૾ೝࣝͰΑ͘ΘΕΔख๏ 28x28 px άϨʔεέʔϧ ΈࠐΈ ϓʔϦϯά ΈࠐΈ ϓʔϦϯά શ݁߹
׆ੑԽؔ ग़ྗ y:9 → 0 y:9 → 1 0.000197 0.999722
NNC͓קΊϙΠϯτ × 3 ϙΠϯτ
NNCͷศརͳͱ͜Ζ1 ίʔσΟϯάϨεʂ ▸ ϓϩάϥϚͰ͋ͬͯϓϩάϥϛϯάͤͣʹࢼͤΔͷັྗʂ ▸ Ϋϥυ൛ͳΒΠϯετʔϧ͢Δඞཁͳ͠ ▸ ϊʔυΛD&D͍͚ͯͩ͘͠ͷ؆୯ૢ࡞ ▸ Ͱֶ͖͕͋ͬͨश݁ՌPython/C++/C͔Βར༻Մ
NNCͷศརͳͱ͜Ζ2 ֶश݁Ռͷཤྺཧ ▸ ػցֶशʹਖ਼ղͳ͍ ▸ ࢼߦࡨޡͯ͠࠷ྑͷ݁ՌΛੜΈग़͢ ▸ ࢼߦࡨޡͷ݁ՌΛهɺޙ͔Βݺͼग़ͤΔ ▸ ෳͷ݁ՌΛάϥϑͰൺֱͰ͖Δ
NNCͷศརͳͱ͜Ζ3 ࣗಈ୳ࡧ࣮ߦ ▸ ωοτϫʔΫߏΛม͑ΔͱࣗಈͰධՁΛ։࢝ͯ͘͠ΕΔ ▸ ਫ਼ͱҰॹʹϦιʔεফඅʢϑοτϓϦϯτʣ࠷దԽ ▸ ϑοτϓϦϯτ࠷దԽΈࠐΈ༻్ʹ͓͍ͯॏཁ
ϋϯζΦϯʹ͍ͭͯ ຊ
શൠ ࢿྉΛݟͳ͕Βɺ͋ͳͨͷϖʔεͰߦͬͯԼ͍͞ ▸ ࣭ෆ໌͕͋Εɺ͙͢ʹखΛڍ͍͛ͯͩ͘͞ ࣌ؒΛϜμʹͤͣɺ࣭͍ͯͩ͘͠͞ʂ ▸ શ෦ऴΘΒͳͯ͘ߏ͍·ͤΜ Ή͠Ζগ͠ଟΊʹͯ͋͠Γ·͢ ▸ ࢿྉ࣋ͪؼΕ·͢
ࣗΦϑΟεͰଓ͖Λߦ͍ͬͯͩ͘͞ʂࢿྉΦϯϥΠϯʹ͋Γ·͢ ▸ ࣭GoogleάϧʔϓͰͥͻʂ
ϋϯζΦϯ1 4ͱ9Λผ͢Δ ▸ खॻ͖Ͱॻ͔Εͨ4͔9ͷը૾Λఆ͠·͢ ▸ 28x28pxɺάϨʔεέʔϧͷը૾Λར༻͠·͢
ϋϯζΦϯ2 ͢ΈΕͷछྨΛผ͠·͢ ▸ 3छྨ͋Δ͢ΈΕΛɺՖͼΒͱΨΫͷ͞ɺ෯ʹԠͯ͡ྨ͚͠·͢ ▸ ೖྗσʔλ4छྨɺ݁Ռ3ͭͷྨʹͳΓ·͢
ϋϯζΦϯ3 ίϯιʔϧ/ίϚϯυϓϩϯϓτͰNNCΛ࣮ߦ͠·͢ ▸ ࡞ͨ͠ϞσϧɺωοτϫʔΫΛμϯϩʔυ͠·͢ ▸ ςετ༻ͷCSVσʔλɺը૾Λ༻͍ͯϩʔΧϧͰίϚϯυΛ࣮ߦ͠·͢ ▸ PythonͱNeural Network LibrariesΛ͍·͢
ϋϯζΦϯ4 WEBΞϓϦέʔγϣϯʹNNCͷϞσϧ/ωοτϫʔΫΛΈࠐΈ·͢ ▸ खॻ͖ೝࣝΞϓϦέʔγϣϯΛ࡞͠·͢ ▸ PythonΛ༻͍·͢ʢίʔσΟϯάຆͲߦ͍·ͤΜʣ ▸ Neural Network LibrariesΛ͍·͢
ϋϯζΦϯ5 0ʙ9ͷը૾Λผ͠·͢ ▸ ReLUMaxPoolingͳͲΛ༻͍ͯɺΑΓෳࡶͳωοτϫʔΫΛߏங͠·͢ ▸ Ϟσϧ͕͏·͘Ͱ͖ͨΒɺϋϯζΦϯ4ͷϞσϧͱೖΕସ࣮͑ͯߦͯ͠Έ·͠ΐ͏ ▸ ίʔυม͑ͣʹϞσϧ͚ͩΛࠩ͠ସ͑ͯ4/9͚ͩͰͳ͘ɺ0ʙ9ͷೝ͕ࣝՄೳʹ ͳΓ·͢