機械学習を使ったサービスのプロダクションコードを書く話 /ml-production

第8回 Machine Learning 15minutes! 2017/01/21

Transcript

1. ػցֶशΛ࢖ͬͨαʔϏεͷ ϓϩμΫγϣϯίʔυΛॻ͘࿩ PKSHA Technology ౔԰ ༞Ұ࿠
2. I am… ౔԰༞Ұ࿠ yuichiroTCY • ~2016/03 2015೥౓IPAະ౿ (εʔύʔΫϦΤʔλ) • 2016/07~ PKSHA Technology • BEDORE։ൃνʔϜ ର࿩Τϯδϯ୲౰ ػցֶश / Web / ૊ΈࠐΈ etc…
3. What’s BEDORE? νϟοτʹΑΔAIҰ࣍Ԡ౴Λ࣮ࢪ
 ΦϖϨʔλʔͱ૊Έ߹Θͤɺ
 ߴ඼࣭௿ίετରԠ Ի੠ղੳπʔϧͱ࿈ܞͨ͠
 ΦϖϨʔλʔͷφϨοδαϙʔτ Web্ͷFAQϖʔδʹ
 AIΤʔδΣϯτΛઃஔ͠ɺ
 ࣗݾղܾ཰޲্ ೔ຊޠʹಛԽͨ͠ର࿩ΤϯδϯͰ͢ɻ νϟοτ΍ΦϖϨʔλʔαϙʔτͷྖҬͰ͝׆༻͍͍͍ͨͩͯ·͢ɻ େखϝσΟΞاۀ༷ɺେखECاۀ༷ͳͲɺ෯޿͘ಋೖ௖͍͓ͯΓ·͢
4. ػցֶशΛ࢖ͬͨαʔϏεΛ ͭ͘Δͱ͖…
5. ΞϧΰϦζϜΤϯδχΞ / ιϑτ΢ΣΞΤϯδχΞ ιϑτ΢ΣΞΤϯδχΞ ΞϧΰϦζϜΤϯδχΞ ʮػցֶशΛ࢖ͬͯ̋̋Λͭ͘Δͧʂʯ ࿦จ ࣮ݧ ઃܭ ӡ༻ https://iconsmind.com/
6. ͲͪΒ΋େࣄ ॊೈੑͷ͋Δ
   ઃܭʹ͠Α͏ ͜ͷ࿦จͷख๏
   ྑͦ͞͏ʂ ࣮ݧͯ͠
   ͍͍݁Ռग़ͨʂ ਫ਼౓্͛ΔͨΊ
   ৽ख๏ࢼͦ͏ ιϑτ΢ΣΞͷ
   ඼࣭Λอͭʹ͸ʜ (PBM
   ͜ΜͳαʔϏεΛ
   Ϣʔβʔʹಧ͚͍ͨ
7. ΞϧΰϦζϜ / ιϑτ΢ΣΞΤϯδχΞͷ͍͋ͩ ιϑτ΢ΣΞΤϯδχΞ ΞϧΰϦζϜΤϯδχΞ ʮػցֶशΛ࢖ͬͯ̋̋Λͭ͘Δͧʂʯ ࿦จ ࣮ݧ ઃܭ ӡ༻ https://iconsmind.com/
8. ࠓ೔࿩͢͜ͱ ιϑτ΢ΣΞ ΤϯδχΞ ΞϧΰϦζϜ ΤϯδχΞ • ػցֶशαʔϏεͷϓϩμΫγϣϯߏ੒ͷҰྫ ✴ Πϯϑϥ ✴ γεςϜઃܭ • ΞϧΰϦζϜ / ιϑτ΢ΣΞΤϯδχΞͷք໘Λ
 ͍͔ʹͯ͠ઃܭ͢Δ͔ ✴ ιʔείʔυ ✴ ςετ ✴ ૊৫ ίίʂ
9. BEDOREγεςϜͷ֓ཁ
10. BEDOREͷର࿩γεςϜུ֓ #&%03&
    TZTUFN API Gateway Web Interface Reply server Database ΞΧ΢ϯτ Φʔφʔ \lSFRVFTUz
    l#&%03&ͬͯͳʹʁz^
    ˣ
    \lSFTQPOTFz
    l͋ͳͨͷ͓ख఻͍Λ
 ͍ͤͯͨͩ͘͞ձ࿩ΤϯδϯͰ͢ʂz^ ֎෦API
11. Reply αʔό 3FQMZ
    TFSWJDF લॲཧ ֎෦"1*ίʔϧ ػցֶश ϧʔϧϕʔε ग़ྗੜ੒
    
    Web API Interface
12. ͜ΜͳཁٻΛຬͨ͢ίʔυΛͲ͏ॻ͚͹ྑ͍ΜͩΖ͏ʁ • ػցֶशॲཧ / ϧʔϧϕʔεॲཧ / ֎෦APIݺͼग़͠
 ͳͲΛڠௐͤ͞Δ
13. BEDOREͷର࿩γεςϜུ֓ʢ࠶ܝʣ #&%03&
    TZTUFN API Gateway Web Interface Database ΞΧ΢ϯτ Φʔφʔ Reply server ֎෦API
14. ΞΧ΢ϯτ Φʔφʔ ֶशαʔό͸෼͚Δ #&%03&
    TZTUFN API Gateway Web Interface Database Learning servers Invoke Save models Storage Load models Training Prediction ୹࣌ؒɾߴෛՙ ৗ࣌ɾ௿ʙதෛՙ ฒྻ࣮ߦ ଱ো֐ੑ Reply server ֎෦API
15. ͜ΜͳཁٻΛຬͨ͢ίʔυΛͲ͏ॻ͚͹ྑ͍ΜͩΖ͏ʁ • ػցֶशॲཧ / ϧʔϧϕʔεॲཧ / ֎෦APIݺͼग़͠
 ͳͲΛڠௐͤ͞Δ • ReplyαʔόͱֶशαʔόΛ෼͚Δ ੡඼ͱͯ͠… • ػೳ֦ுɾอक͸༰қʹ • όά͸ग़͞ͳ͍ • ਫ਼౓޲্ͷͨΊߴ଎ʹࢼߦࡨޡ
16. ػցֶशγεςϜͷઃܭ
17. ʢࢀߟʣ http://www.slideshare.net/takahirokubo7792/ss-65413290 icoxfog417ࢯʮػցֶशͰٽ͔ͳ͍ͨΊͷίʔυઃܭʯ
18. BEDOREࣜɾػցֶशσβΠϯύλʔϯʢҊʣ .PEFM 5SBJOFS %BUB1SPWJEFS
19. BEDOREࣜɾػցֶशσβΠϯύλʔϯʢҊʣ .PEFM
    
    .FUIPET
    w QSFQSPDFTT
    w USBJO
    w QSFEJDU
    w TBWF
    w SFTUPSF 
    1BSBNT
    ྫ
    w IJEEFO@MBZFST
    w ESPQPVU
    
    
    
    
    
    FUDʜ 5SBJOFS 
    1BSBNT
    ྫ
    w O@TUFQT
    w PQUJNJ[FS
    
    
    
    
    
    FUDʜ 
    .FUIPET
    w USBJO %BUB1SPWJEFS
    
    .FUIPET
    w HFU@USBJOJOH@EBUB ʢྫɿNeural NetworksΛ࢖ͬͨϞσϧͷ৔߹ʣ
20. ੹຿Λ෼཭͢Δ • Model • train, predict, preprocess, save, restore
 ͱ͍͏ΠϯλϑΣʔε͚ͩఆΊΔ • த਎͸ΞϧΰϦζϜΤϯδχΞ୲౰ • DataProvider • ֶशσʔλΛఏڙ͢Δget_training_dataͱ͍͏ΠϯλϑΣʔε͚ͩ ఆΊΔ • DB͔ΒಡΜͰ΋APIୟ͍ͯ΋CSVಡΜͰ΋OKɻ࢖͏ଆ͸ؔ஌͠ͳ͍ • Trainer • DataProvider͔ΒֶशσʔλΛऔಘͯ͠ɺ
 ֶशύϥϝʔλͱҰॹʹModelΫϥεͷtrain()ʹ౉͚ͩ͢
21. Training .PEFM
    
    .FUIPET
    w QSFQSPDFTT
    w USBJO
    w QSFEJDU
    w TBWF
    w SFTUPSF 
    1BSBNT
    ྫ
    w IJEEFO@MBZFST
    w ESPQPVU
    
    
    
    
    
    FUDʜ 5SBJOFS 
    1BSBNT
    ྫ
    w O@TUFQT
    w PQUJNJ[FS
    
    
    
    
    
    FUDʜ 
    .FUIPET
    w USBJO data_provider = DataProvider() trainer = Trainer(data_provider, train_params) model = Model(model_params) trainer.train(model) model.save(‘path/to/save’) %BUB1SPWJEFS
    
    .FUIPET
    w HFU@USBJOJOH@EBUB w %BUB1SPWJEFSHFU@USBJOJOH@EBUB
    ͔Β
 ֶशσʔλΛऔಘ͠ɺ
    w .PEFMUSBJO Λݺͼग़ͯ͠౉͢ɻ w ड͚औͬͨσʔλΛ
 QSFQSPDFTT Ͱલॲཧ͠ɺ
    w ֶशΛߦ͏
22. Prediction .PEFM
    
    .FUIPET
    w QSFQSPDFTT
    w USBJO
    w QSFEJDU
    w TBWF
    w SFTUPSF 
    1BSBNT
    ྫ
    w IJEEFO@MBZFST
    w ESPQPVU
    
    
    
    
    
    FUDʜ 5SBJOFS 
    1BSBNT
    ྫ
    w O@TUFQT
    w PQUJNJ[FS
    
    
    
    
    
    FUDʜ 
    .FUIPET
    w USBJO model = Model.restore(‘path/to/load’) result = model.predict(‘Hello, BEDORE!’) %BUB1SPWJEFS
    
    .FUIPET
    w HFU@USBJOJOH@EBUB w ड͚औͬͨσʔλΛ
 QSFQSPDFTT Ͱલॲཧ͠ɺ
    w ࣝผΛߦ͏
23. ModelͷՄൖੑ .PEFM
    
    1BSBNT
    ྫ
    w IJEEFO@MBZFST
    
    <b
    `>
    w ESPQPVU
    
    
    
    
    
    
    
    FUDʜ 
    5SBJOFE
    NPEFM
    w /FUXPSL
    QBSBNFUFST
    FUDʜ save() .PEFM
    
    1BSBNT
    ྫ
    w IJEEFO@MBZFST
    
    <b
    `>
    w ESPQPVU
    
    
    
    
    
    
    
    FUDʜ 
    5SBJOFE
    NPEFM
    w /FUXPSL
    QBSBNFUFST
    FUDʜ restore()
24. ModelͷՄൖੑ .PEFM
    
    1BSBNT
    ྫ
    w IJEEFO@MBZFST
    
    <b
    `>
    w ESPQPVU
    
    
    
    
    
    
    
    FUDʜ 
    5SBJOFE
    NPEFM
    w /FUXPSL
    QBSBNFUFST
    FUDʜ save() .PEFM
    
    1BSBNT
    ྫ
    w IJEEFO@MBZFST
    
    <b
    `>
    w ESPQPVU
    
    
    
    
    
    
    
    FUDʜ 
    5SBJOFE
    NPEFM
    w /FUXPSL
    QBSBNFUFST
    FUDʜ restore() TBWF
    
    SFTUPSF
    w .PEFMΫϥε͕࣋ͭ͋ΒΏΔύϥϝʔλ
 ʢϋΠύʔύϥϝʔλɾֶशࡁΈϞσϧɾલॲཧ༻ࣙॻͳͲʣΛ
 ̍σΟϨΫτϦʹύοΫˍϦετΞ
    w 1JDLMFͳͲͷγϦΞϥΠβ΍
 ϥΠϒϥϦ
    5FOTPS'MPXͳͲ ͷTBWF
    
    SFTUPSFΛ׆༻
25. BEDOREγεςϜʹ͓͚ΔModelϑΝΠϧͷऔΓճ͠ #&%03&
    TZTUFN API Gateway Web Interface Database Learning servers Invoke Save models Storage Load models ֶश͸ઐ༻αʔόͰߦ͍ɺ
 ֶशࡁϞσϧ ϑΥϧμݸʂ Λ
 4UPSBHFʹอଘɺ
    3FQMZαʔόʹϩʔυ͢Δɻ
    w ෛՙͷେֶ͖͍शॲཧ
 ʹઐ༻ͷߴՐྗΠϯελϯε
 ɹΛεϙοτతʹ࢖༻
    w աڈͷϞσϧ͸શͯอଘ͞ΕΔ
 ʹϩʔϧόοΫՄೳ Reply server
26. ࣮ݧ΋ָʹ .PEFM
    
    .FUIPET
    w QSFQSPDFTT
    w USBJO
    w QSFEJDU
    w TBWF
    w SFTUPSF 
    1BSBNT
    ྫ
    w IJEEFO@MBZFST
    w ESPQPVU
    
    
    
    
    
    FUDʜ 5SBJOFS 
    1BSBNT
    ྫ
    w O@TUFQT
    w PQUJNJ[FS
    
    
    
    
    
    FUDʜ 
    .FUIPET
    w USBJO
    w FWBM %BUB1SPWJEFS
    
    .FUIPET
    w HFU@USBJOJOH@EBUB 1. DataProvider.get_training_data()͔ΒಘΒΕͨσʔλΛtrain / testʹ෼͚Δ 2. Model.train()Ͱֶश͢Δ 3. Model.predict()Ͱςετ͢Δ
27. ࣮ݧ΋ϓϩμΫγϣϯͷֶशͱಉ༷ͷॲཧɾΠϯϑϥͰ #&%03&
    TZTUFN Web Interface Database Experiment servers Invoke Save models Storage ࣮ݧ΋ֶशαʔόͱಉ༷ʹ
 ΦϯσϚϯυΠϯελϯεͰ
    w σʔλಡΈࠐΈॲཧ౳͸
 ֶशͱಉ͡΋ͷΛར༻
 ʹϩδοΫ΍ύϥϝʔλͷ
 ɹมߋʹઐ೦
 ʹσʔλ͸ຊ൪ͱಉ͡΋ͷ Reply server
28. .PEFM
    Model mixins ػցֶशͷຊ࣭Ͱ͸ͳ͍͕ศརͳػೳ͸
 Mixinͱͯࠩ͠͠ࠐΊΔΑ͏ʹ͓ͯ͘͠ .JYJO 
    .FUIPET
    w QSFQSPDFTT
    w USBJO
    w QSFEJDU
    w TBWF
    w SFTUPSF 
    1BSBNT
    w IJEEFO@MBZFST
    w ESPQPVU
    
    
    
    
    
    FUDʜ .JYJO .JYJO
29. Model mixins • Label Transformer • ೚ҙͷ؅ཧIDΛ࿈൪ͷIntegerʹม׵ • Cache Loader • restore() ʹΩϟογϡػೳΛ෇͚Δ Mixinͱͯ͠༻ҙ͓ͯ͘͠ͱศརͩͬͨػೳͷྫ
30. ModelͷΧϓηϧԽ • தͷ೉͍͜͠ͱ͸ΞϧΰϦζϜΤϯδχΞʹ
 ͓Ͷ͕͍͢Δ • தͰAPIΛݺ΅͏͕Կ͠Α͏͕OK • ֎෦ʹAPIΛ࣮૷͠ɺModel͸ͦΕΛݺͿϥούʔʹ ప͢Δͱ͔
31. ModelͷΧϓηϧԽ • train, predict, save, restore (, preprocess) ͕
 ࢖͑Ε͹ྑ͍ • ΠϯλϑΣʔε෦෼͸
 ιϑτ΢ΣΞ / ΞϧΰϦζϜΤϯδχΞͷ
 ڠۀ͕ඞཁ͔΋
32. ςετՄೳੑ ςετͷͳ͍ίʔυ͸੡඼ʹͰ͖ͳ͍ ʮςετͷແ͍ίʔυ͸ϨΨγʔίʔυʯ http://www.shoeisha.co.jp/book/detail/9784798116839 https://www.amazon.co.jp/exec/obidos/ASIN/4894717115/hatena-blog-22/
33. ςετՄೳੑ ςετͷͳ͍ίʔυ͸੡඼ʹͰ͖ͳ͍ ʮςετͷແ͍ίʔυ͸ϨΨγʔίʔυʯ ModelΫϥεͷ੾Γ෼͚ʹΑͬͯ ૄ݁߹㱺ςετՄೳੑUP • ModelΫϥεΛMockͯ͠ଞίϯϙʔωϯτΛςετ • ModelΫϥε͕୲͏ػցֶशͷ”ςετ”ʹਫ਼౓ධՁ΋
 ୯ମͰ࣮ߦՄೳ
34. 3FQMZ
    TFSWJDF ΞϧΰϦζϜΤϯδχΞ / ιϑτ΢ΣΞΤϯδχΞ ModelΫϥεͷ੹຿ͷ෼཭ʹ྆ऀͷ੹຿ͷ෼཭ .PEFM
    QSFQSPDFTT USBJO QSFEJDU TBWF SFTUPSF ιϑτ΢ΣΞ ΤϯδχΞ https://iconsmind.com/ ΞϧΰϦζϜ ΤϯδχΞ
35. ػցֶशϓϩμΫτͷνʔϜ ιϑτ΢ΣΞΤϯδχΞ ΞϧΰϦζϜΤϯδχΞ อकੑ ҆ఆੑ ֦ுੑ ࣮ߦޮ཰ ࣮ݧ ݚڀ։ൃ ࠷৽ख๏΁ͷཧղ ઃܭ
36. ·ͱΊ ద੾ͳ • ίʔυͷઃܭ • ੹຿ͷ੾Γ෼͚ ʹΑͬͯɺ ΞϧΰϦζϜ / ιϑτ΢ΣΞΤϯδχΞ͕ڠྗͯ͠ ߴ඼࣭ɾߴੑೳͳϓϩμΫτΛ
 σϦόϦʔͰ͖ΔΑ͏ʹ͠Α͏
37. Ξ ϧ ΰ Ϧ ζ Ϝ Τ ϯ δ χ Ξ ื ू த ͝ڵຯ͕͋Δํ͸ࢲ΁ɺ·ͨ͸Wantedly͔Β͝࿈བྷ͍ͩ͘͞
38. ͝ڵຯ͕͋Δํ͸ࢲ΁ɺ·ͨ͸Wantedly͔Β͝࿈བྷ͍ͩ͘͞ https://www.wantedly.com/projects/75747