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
PyTorchとCaffe2とONNXと深層学習モデルのデプロイについて
Search
Yuichiro Someya
January 27, 2018
Programming
1
2.9k
PyTorchとCaffe2とONNXと深層学習モデルのデプロイについて
https://machine-learning15minutes.connpass.com/event/73442/
Yuichiro Someya
January 27, 2018
Tweet
Share
More Decks by Yuichiro Someya
See All by Yuichiro Someya
にんげんがさき 基盤はあと / Developers over ML platform
ayemos
0
13k
機械学習をスモールスタートさせる方法 / small machine learning
ayemos
3
2k
アットホームな分析基盤の作り方 / Homemade Machine Learning Toolkits
ayemos
1
930
サービス開発、機械学習、クラウド / the trinity of machine learning
ayemos
0
3.3k
成長を止めない機械学習のやり方 / Don't stop 'til you get enough (data).
ayemos
15
5.1k
AWS で加速する機械学習 / Accelerate Machine Learning with AWS
ayemos
1
300
クックパッドの機械学習基盤 2018 / Machine Learning Platform at Cookpad ~ 2018 ~
ayemos
15
19k
クックパッドにおけるAWS GPUインスタンスの利用事例 / Powering by AWS GPU Instances in Cookpad Inc
ayemos
0
390
How we use GPUs in Cookpad
ayemos
0
120
Other Decks in Programming
See All in Programming
Importmapを使ったJavaScriptの 読み込みとブラウザアドオンの影響
swamp09
3
1.2k
CSC305 Lecture 13
javiergs
PRO
0
120
Kaigi on Rails 2024 - Rails APIモードのためのシンプルで効果的なCSRF対策 / kaigionrails-2024-csrf
corocn
4
2.8k
『ドメイン駆動設計をはじめよう』のモデリングアプローチ
masuda220
PRO
4
150
OpenTelemetryでRailsのパフォーマンス分析を始めてみよう(KoR2024)
ymtdzzz
3
1k
WEBエンジニア向けAI活用入門
sutetotanuki
0
290
From Subtype Polymorphism To Typeclass-based Ad hoc Polymorphism- An Example
philipschwarz
PRO
0
110
VR HMDとしてのVision Pro+ゲーム開発について
yasei_no_otoko
0
100
The future of development – Are our jobs getting harder or easier?
hollycummins
1
240
Kubernetes for Data Engineers: Building Scalable, Reliable Data Pipelines
sucitw
1
180
現場で役立つモデリング 超入門
masuda220
PRO
12
2.6k
Golang と Erlang
taiyow
8
1.8k
Featured
See All Featured
CoffeeScript is Beautiful & I Never Want to Write Plain JavaScript Again
sstephenson
159
15k
Product Roadmaps are Hard
iamctodd
PRO
48
10k
Designing for humans not robots
tammielis
249
25k
Testing 201, or: Great Expectations
jmmastey
38
7k
Building Adaptive Systems
keathley
38
2.2k
The Pragmatic Product Professional
lauravandoore
31
6.3k
個人開発の失敗を避けるイケてる考え方 / tips for indie hackers
panda_program
92
16k
BBQ
matthewcrist
85
9.3k
Designing Experiences People Love
moore
138
23k
[Rails World 2023 - Day 1 Closing Keynote] - The Magic of Rails
eileencodes
32
1.8k
How to Think Like a Performance Engineer
csswizardry
19
1.1k
10 Git Anti Patterns You Should be Aware of
lemiorhan
653
59k
Transcript
1Z5PSDIͱ$B⒎Fͱ0//9ͱ ਂֶशϞσϧͷσϓϩΠʹ͍ͭͯ ΫοΫύουגࣜձࣾݚڀ։ൃ෦છ୩༔Ұ
ࣗݾհ છ୩༔Ұ<:VJDIJSP4PNFZB> ΫοΫύουגࣜձࣾݚڀ։ൃ෦ ϦαʔνΤϯδχΞ݄d "84ͱ͔ػցֶशͱ͔ .-NJOVUFTճ
UXJUUFSDPN!BZFNPT@Z HJUIVCDPNBZFNPT IUUQTXXXBZFNPTNF
None
ϨγϐɿສҎ্ ࠃͷ݄ؒར༻ऀɿສਓ
ରԠݴޠɿݴޠϲࠃ ւ֎ͷ݄ؒར༻ऀɿສਓ
"HFOEB ͢͜ͱ χϡʔϥϧωοτϫʔΫϞσϧͷσϓϩΠͷݱঢ়ͱ՝ʹ͍ͭͯ 1Z5PSDIͱ$B⒎Fͱ0//9 χϡʔϥϧωοτϫʔΫͷதؒදݱͱ
͞ͳ͍͜ͱ 1Z5PSDIɺ$B⒎Fɺ0//9ͷ͍ํ σϞ
ϞσϧσϓϩΠͷݱঢ়ͱ՝
എܠ ػցֶशΛར༻͢Δ જࡏ ਓ͕૿͍͑ͯΔ ྫΫοΫύουגࣜձࣾݚڀ։ൃ෦ ݄ͷൃ໊࣌ͷϝϯόʔ
݄ݱࡏࠃʹ໊ ւ֎ʹ໊ ϞσϧͷσϓϩΠ࡞ۀ͕ϘτϧωοΫʹͳ͖ͬͯͨ
5SBJO 5SBJO*OGFSͱ͍͏ॱ൪ ͋Δ͍ͦΕΛ܁Γฦ͢ ྫϓϩτλΠϓɺ1P$ ͳΔָ͘ʹ͘Γ͍ͨ *OGFS
σʔλऩू .PEFM ܇࿅ લॲཧ "1*࣮ ੳ %PDLFSԽ όον࣮ ςετʗࢹ
ϞσϧͷσϓϩΠख๏ྫ ΞϓϦέʔγϣϯίʔυʹΈࠐΉ 1ZUIPOQBDLBHFʹ͢Δ "1*Խ͢Δ
ΞϓϦέʔγϣϯίʔυʹΈࠐΉ ͍͍ͱ͜Ζ ௨৴Φʔόʔϔου͕ආ͚ΒΕΔ Α͘ͳ͍ͱ͜Ζ ΞϓϦέʔγϣϯίʔυΛ͍͡Δඞཁ͕͋Δ
1ZUIPOڥʹݶΒΕΔPSม͢Δඞཁ͕͋Δ Ұํ1Z$BMMͱ͍͏ͷ͋Δ άϧʔίʔυ͕ॏෳ͢Δ ͔݁͠ߏάϧʔ͕ް͔ͬͨΓ͢Δ IUUQTHJUIVCDPNNSLOQZDBMMSC
1ZUIPOQBDLBHFʹ͢Δ άϧʔ෦ΛؚΊͯ1ZUIPOQBDLBHFԽͯ͠ଞॴͰར༻Ͱ͖Δ Α͏ʹ͢Δ ެ։Ͱ͖ͳͷʹ͍ͭͯ1Z1*$MPVE ͳͲΛࣾʹ ཱͯΔͱ͍͏ख͕͋Δ IUUQTHJUIVCDPNTUFWFBSDQZQJDMPVE
1ZUIPOQBDLBHFʹ͢Δ ͍͍ͱ͜Ζ άϧʔίʔυͷॏෳ͕ආ͚ΒΕΔ Α͘ͳ͍ͱ͜Ζ ґવ1ZUIPOڥʹݶΒΕΔ
ख͕͔͔ؒΔ //ϑϨʔϜϫʔΫͷࣝɺ1ZUIPOͷQBDLBHJOHʹؔ͢Δ͕ࣝඞཁʹ ͳΔ ୭͕Δͷ͔ IUUQTHJUIVCDPNTUFWFBSDQZQJDMPVE
"1*Խ͢Δ Ϟσϧ୯ମΛ"1*Խ ϞσϧΛಡΈࠐΜͰɺ֎෦ϦΫΤετʹԠͯ͡ϞσϧΛ ݺͼग़͢ബ͘খ͞ͳ"1*Λ࣮͢Δ NJDSPTFSWJDF पลͷϩδοΫ
FHʰϞσϧ͕εύϜఆͨ͠Β%#ʹϑϥ άΛཱͯΔʱ ֎෦Ͱ࣮
"1*Խ͢Δ ͍͍ͱ͜Ζ Ϟσϧ࣮ͱΞϓϦέʔγϣϯ࣮Λ্ख͘ΓͤΔ ։ൃऀ্ख͚͘ΒΕΔ ΞϓϦέʔγϣϯ1ZUIPOͰͳͯ͘Α͍ NJDSPTFSWJDF
Α͘ͳ͍ͱ͜Ζ ख͕͔͔ؒΔ
"1*αʔόʔ ʰ1ZUIPOQBDLBHFԽɺ"1*Խ͍ͣΕ͔ͷํ๏ʹΑΓɺϞσϧ ͷ࣋ͪग़͕͠ग़དྷΔʱ ͍ͣΕख͕͔͔ؒΓɺ1P$ͷ߹ಛʹࠔΔ Γग़ͯ͠ڞ௨ج൫Խ͢Δʁ ͪΐͬͱͨ͠ϝλσʔλKTPO NPEFMQC .PEFMͷμϯϓϑΝΠϧ
NPEFMQC ͳʹ͔ 1ZUIPOQBDLBHF NPEFMQC
"1*αʔόʔ ͜Εݫ͍͠ NPEFMQC ͳʹ͔ 1ZUIPOQBDLBHF NPEFMQC
ͪͳΈʹ (PPHMF$MPVE.-&OHJOFͱ͍͏͕͋Δ IUUQTDMPVEHPPHMFDPNNMFOHJOF 5FOTPSqPXɺ($1റΓ͋Δ͕Α͘ಈ͘ ެࣜʮόον༻ɻϦΞϧλΠϜਪ#FUBͰ4-"ͳ͍Αʯ "1*αʔόʔ
ͪΐͬͱͨ͠ϝλσʔλ NPEFMQC .PEFMͷμϯϓϑΝΠϧ NPEFMQC
͜͜·Ͱ·ͱΊ Ϟσϧ։ൃऀ͕૿͖͑ͯͨ݁ՌɺϞσϧͷσϓϩΠͱ͍͏ ϘτϧωοΫ͕ݟ͖͑ͯͨ σϓϩΠ࡞ۀΛڞ௨ج൫ͱ͍͏ܗͰΈԽ͍ͨ͠ͱ͜Ζ ޓ͍ʹޓੑͷ ΄ͱΜͲ ͳ͍//ϑϨʔϜϫʔΫ͕
ࢁ͋Δ
1Z5PSDIͱ$B⒎Fͱ0//9
//ϑϨʔϜϫʔΫͷಛੑͱબ ͳͥ//ϑϨʔϜϫʔΫෳ͋Δʁ
//ϑϨʔϜϫʔΫͷಛੑͱબ ʰ࣮ݧͷඇϓϩμΫγϣϯڥʹ͓͍ͯਪͷ࣮ߦ ΑΓॊೈੑ͕༏ઌ͞ΕΔʱ ʰσϓϩΠઌͰɺ࣮ߦڥ αʔόʔʗΤοδ ʹద͠ɺ ਪʹ࠷దԽ͞ΕͨϑϨʔϜϫʔΫ͕·͍͠ʱ
ॊೈੑͱʁ Α͘ʹͳΔͷɺ੩తμΠφϛοΫ
1Z5PSDIͱ$B⒎F 1Z5PSDI ໋ྩతͰ μΠφϛοΫɺ࣮ݧ͖ $B⒎F એݴతͰ
੩తͳϑϨʔϜϫʔΫ ΫϩεϓϥοτϑΥʔϜͰ࣮ߦՄೳ FHJ04 "OESPJE ࣮ݧͱਪ͍͚ΔϞσϧͷՄൖੑ͕ॏཁࢹ͞Εͯ͘Δ
0//9 χϡʔϥϧωοτϫʔΫͷΦʔϓϯͳ*3 தؒදݱ ఆٛ ࣮ମ1SPDPM#V⒎FS
ྫ͑ 1Z5PSDI ࣮ݧ 0//9$B⒎F ϓϩμΫγϣϯ IUUQTSFTFBSDIGCDPNGBDFCPPLBOENJDSPTPGUJOUSPEVDFOFXPQFOFDPTZTUFNGPSJOUFSDIBOHFBCMFBJGSBNFXPSLT IUUQTEFWFMPQFSTHPPHMFDPNQSPUPDPMCV⒎FST 0//9'SPOUFOE 0//9#BDLFOE
ࣅͨΑ͏ͳऔΓΈ IUUQUWNMBOHPSH IUUQTHJUIVCDPNENMDUWN IUUQTHJUIVCDPNENMDOOWN IUUQTXXXLISPOPTPSHOOFG 57.4UBDL 57. //7.
ͦͷଞͷରԠϑϨʔϜϫʔΫ μΠφϛοΫ $IBJOFS 1Z5PSDI FYQPSUରԠॏࢹ ੩త 5' $B⒎F
ɺΤοδ༻ $PSF.- JNQPSUରԠॏࢹ IUUQTHJUIVCDPNPOOYUVUPSJBMT
ͦͷଞͷରԠϑϨʔϜϫʔΫ ྫ͑$B⒎Fͷ࣮ߦڥΛ༻ҙ͓͚ͯ͠ɺ$B⒎F 1ZUPSDI $IBJOFSͷϞσϧ͕࣮ߦͰ͖Δɻ ڞ௨࣮ߦڥເͰͳ͍ 5FOTPSqPXؤுΕ IUUQTHJUIVCDPNPOOYUVUPSJBMT
0//9CBDLFOE "1*αʔόʔ ເ NPEFMQC ͳʹ͔ 1ZUIPOQBDLBHF 0//9GSPOUFOE NPEFMPOOY ͪΐͬͱͨ͠ϝλσʔλKTPO
·ͱΊ ϞσϧͷσϓϩΠʹ·ͭΘΔ՝ͷͭͱͯ͠ɺͦΕͧΕʹ ಛੑ͕ҟͳΔʰෳͷ//ϑϨʔϜϫʔΫ͕ଘࡏ͢ΔࣄʱΛڍ ͛ͨ 0//9Λར༻ͯ͠ϞσϧΛσϓϩΠ͢Δҝͷڞ௨ج൫Λߏங ͢ΔΞΠσΞΛհͨ͠
એ ͦΜͳج൫ΛҰॹʹ࡞Γ·ͤΜ͔ IUUQTDPPLQBEKPCT
ࢀߟจݙ IUUQTSFTFBSDIGCDPNQVCMJDBUJPOTBQQMJFENBDIJOF MFBSOJOHBUGBDFCPPLBEBUBDFOUFSJOGSBTUSVDUVSF QFSTQFDUJWF IUUQTSFTFBSDIGCDPNGBDFCPPLBOENJDSPTPGU JOUSPEVDFOFXPQFOFDPTZTUFNGPSJOUFSDIBOHFBCMFBJ GSBNFXPSLT
IUUQTHJUIVCDPNPOOYUVUPSJBMT