Suggestions and techniques for building neural networks with Go

Suggestions and techniques for building neural networks with Go Go
Conference Tokyo 2019 Autumn / October 28, 2019 @ po3rin

pon (@po3rin) Software Enginner @ Shiroyagi Corporation Golang / Python
/ Rust / Docker / AWS / Elasticsearch

QBHF നϠΪίʔϙϨʔγϣϯ͸ɺػցֶशɾࣗવ ݴޠॲཧ෼໺ͷϓϩμΫτʹྗΛೖΕ͍ͯΔձ ࣾͰ͢ɻ։ൃʹ͸ओʹ (P ΍ 1ZUIPO ͕࠾༻ ͞Ε͍ͯ·͢ɻ നϠΪίʔϙϨʔγϣϯ

Package for Machine Lerning in Go Ὂ (P (POVNͰχϡʔϥϧ ωοτϫʔΫΛ૊ΉͨΊͷί
ϯϙʔωϯτूʂ ./*45σʔληοτΛ(Pͷ ߦྻ࣮૷ͰಡΈࠐΉͨΊͷ ύοέʔδ

1: Why do data-science with Go?

Why Go ? Ὂ (PͷڧΈ͕ͦͷ··σʔλαΠΤϯεΛ(PͰ΍Δཧ༝ ίʔυ͕ಡΈ΍͘͢ϝϯς͠΍͍͢ ύϑΥʔϚϯε͕ྑ͍ ฒߦԽ͠΍͍͢ ҰํͰ΍͸Γ·ͩ·ͩࢿ࢈తʹ1ZUIPOʹ܉ഋ্͕͕Δͷ͸͔֬ɻɻ

Marchine Larning Packages Ὂ ͔͠͠ɺϞσϧΛಈ͔͢؀ڥ͸͔ͳΓॆ࣮͖͍ͯͯ͠ΔͷͰɺਪ࿦ΛϚΠΫϩαʔϏ εͱཱͯͯ͠Ε͹(P ϚΠΫϩαʔϏεͷخ͠Έ͕ಘΒΕΔɻ ֶशࡁΈϞσϧΛ5FOTPS'MPX͕ ఏڙ͢Δ$"1*Λ࢖༻ͯ͠ಈ͔͢ɻ ΋ͪΖΜ(P΋αϙʔτ͍ͯ͠Δɻ
1VSF(PͰॻ͔Ε͍ͯΔɻ0//9 ϑΥʔϚοτͰॻ͔Εͨχϡʔϥϧ ωοτΛΠϯϙʔτͯ͠ಈ͔ͤΔ

Why Go for Biginners ? Ὂ ॳ৺ऀͰɺجૅతͳͱ͜Ζ͔Βษڧ͔ͨͬͨ͠Β(P͕ ྑ͍ͱײͨ͡ɻ1ZUIPO͸ศརͳύοέʔδ͕ଟ͍ͷ Ͱɺ(PͰ࣮૷͍ͯ͘͠ͱࠜఈͷཧղ͕ͳ͍ͱ࣮૷Ͱ͖ ͳ͔ͬͨΓ͢Δɻ

Why Go ? Ὂ ྫ͑͹ɺ֬཰෼෍ΛݩʹཁૉΛϥϯμϜʹબ୒͢Δ࣮૷͸1ZUIPOͰ͸ߦͩ ͕ɺ(PͰ͸ࣗ෼ͰΞϧΰϦζϜΛษڧ࣮ͯ͠૷͢Δඞཁ͕͋ͬͨɻ͜͏͍ͬͨ ࡉ͔͍ॴͷ࢓૊Έ΋ษڧʹͳΔɻ https://qiita.com/po3rin/items/f28a8ce322cb1bd95175 ʮ༩͑ΒΕͨॏΈʹैͬͯϥϯμϜʹ஋Λฦ͢ʮ8FJHIUFE3BOEPN4FMFDUJPOʯΛ(PͰ࣮૷͢Δʂʯ

2: Basic of matrix calculation with Go

Matrix operation package choices Ὂ gonum.org/v1/gonum/mat (PͰߦྻΛѻ͏ͨΊͷ൚༻తͳ ύοέʔδɻ׆ൃʹϝϯς͞Εͯ Δɻૄߦྻ΍ࡾ֯ߦྻͳͲʹ΋ରԠ gorgonia.org/tensor
(Pͷػցֶश༻ϥΠϒϥϦ ʮHPSHPOJBʯͰ࢖͏͜ͱΛ໨తͱ ͨ͠ߦྻɻଟ࣍ݩʹ΋ରԠ

Matrix operation (1) Ὂ (Pʹ͸1ZUIPOʹ͋ΔΑ͏ͳԋࢉࢠΦʔόʔϩʔυ͕ͳ͍ͷͰؔ਺ͰԋࢉΛߦ͏ɻ

Matrix operation (2) Ὂ ߦྻΛදݱ͢Δߏ଄ମͰ͋Δ NBU%FOTF͸಺෦Ͱશཁૉͷ εϥΠε<>qPBUͱߦ਺ྻ਺ ͳͲͷ৘ใΛอ͍࣋ͯ͠Δ͚ͩɻ

Matrix operation (3) Ὂ NBU%FOTF͸.BUSJYΠϯλʔϑΣʔεΛ࣮૷͍ͯ͠Δɻ͜ΕʹΑΓσʔλߏ଄ͷҧ͍ ΛҙࣝͤͣʹߦྻΛར༻Ͱ͖Δɻଞʹ΋7FD%FOTFʢϕΫτϧΛදݱʣ΍5SJ%FOTF ʢࡾ֯ߦྻΛදݱʣͳͲ.BUSJYΛ࣮૷͢Δ༷ʑͳߏ଄ମ͕͋Δɻ

Matrix operation (4) Ὂ ࣍ݩͷߦྻ͸HPOVNNBUͰ͸ߦྻͷεϥΠεͰ͔͠දݱͰ͖ͳ͍ɻଟ࣍ݩߦྻ ʹରԠ͢ΔύοέʔδΛ࡞Δ৔߹͸UFOTPSUFOTPSʹ܉ഋ্͕͕Δɻ

࠷ॳʹݟͨͱ͓Γɺ.BUSJYΠϯλʔϑΣʔεͷ࣮ଶͱͯ͠ %FOTF͕͋ΔɻNBUϑΟʔϧυͷܕͰ͋ΔCMBT(FOFSBMͱ ͸ͳΜͰ͠ΐ͏͔ɻ Use BLAS (1) Ὂ

Use BLAS (2) Ὂ #-"4ͱ͸ɺ#BTJD-JOFBS"MHFCSB4VCQSPHSBNTͷུͰɺجຊతͳઢ ܗԋࢉͷϥΠϒϥϦɻ#-"4ʹ͸͞·͟·ͳछྨ͕͋Γ0QFO#-"4 ΍"5-"4ͳͲ͕ଘࡏ͢Δɻ

Use BLAS (3) Ὂ (POVNͰ͸6TFͱ͍͏ؔ਺Ͱ࢖͏#-"4Λ͙͢ʹ੾Γସ͑ΕΔɻ

Use BLAS (4) Ὂ ͨͱ͑͹%PU ཁૉ͝ͱͷੵͷ߹ܭΛܭࢉ͢Δ Ͱ͸಺෦Ͱ͸CMBT%EPU Λݺͼग़͍ͯ͠Δɻ౰વCMBT%EPU͸ΠϯλʔϑΣʔεͰఆٛ͞Εͨϝιο υͳͷͰCMBTͷ࣮ମ͝ͱʹҧ͏ॲཧΛݺ΂Δɻ

Use BLAS (5) Ὂ ͨͱ͑͹ཁૉͷϕΫτϧಉ࢜Λ%PUؔ਺ʹ͔͚Δ࣌ͷϕϯνϚʔΫͰ ͸Լهͷ؀ڥͰ͸ഒ΄Ͳ0QFO#-"4ͷํ͕ૣ͘ͳ͍ͬͯ·͢ɻ

2JJUBʹ(P (POVNʹΑΔߦྻܭࢉΛ ମܥతʹ·ͱ·͍ͬͯΔɻߦྻܭࢉͷΤ ϥʔϋϯυϦϯά΍ςετͳͲ΋Χόʔ͞ Ε͍ͯ·͢ɻ https://qiita.com/po3rin/items/82c9c0195f9e3e38e2f1 Study Gonum Ὂ

IUUQTXXXDIFBUPHSBQIZDPNDIFXYZDIFBUTIFFUTEBUBTDJFODFJOHPB Study matrix operation using Go Ὂ (PSHPOJB࡞ऀʹΑΔ ʮOVNQZʯʮHPOVNNBUʯ ʮHPSHPOJBUFOTPSʯʹ͓͚Δ
ߦྻԋࢉͷνʔτγʔτ͕ૉ੖ Β͍͠ɻOVNQZͷ͜Εͬͯ(P ͩͱͲ͏΍ΔΜͩʁ͕΄΅ࡌͬ ͯΔɻ

3: Architecture for building neural networks with Go

Example of Machine Lerning in Go Ὂ (P (POVNͰχϡʔϥϧωοτϫʔΫΛ૊ΉͨΊͷίϯϙʔωϯτूͰ͋ ΔHPOOQͷ࣮૷ܦݧ͔Β࣮૷ΞʔΩςΫνϟΛߟ͑Δɻ

Architecture (2) Ὂ ܇࿅༻ͷΦϒδΣΫτΛ.PEFMͱ0QUJNJ[FSΛ౉ͯ͠ॳظԽ ͠·͢ɻॳظԽؔ਺͸ύϥϝʔλ͕ඇৗʹଟ͘ͳΔͷͰ 'VODUJPOBM0QUJPO1BUUFSOTͰ࣮૷͓ͯ͘͠ͱྑ͍ɻ 5SBJOFS .PEFM 0QUJNJ[FS

Architecture (3) Ὂ 5SBJOFS 0QUJNJ[FS "⒏OF-BZFS 3FMV-BZFS "⒏OF-BZFS "⒏OF-BZFS ϨΠϠʔͷ࣮૷͸*OUFSGBDFͱͯ͠ఆ͓ٛͯ͠
͖ɺύοέʔδΛར༻͢ΔϢʔβʔ͕ಠࣗͷϨ ΠϠʔΛ࣮૷ͯ͠૊ΈࠐΊΔΑ͏ʹ͓ͯ͘͠

Architecture (4) Ὂ 8FJHIU΍#JBTͳͲͷύϥϝʔλΛߋ৽͢ Δ0QUJNJ[FSɻ΋ͪΖΜ͜͜΋Ϣʔβʔ͕೚ ҙͷΞϧΰϦζϜΛར༻Ͱ͖ΔΑ͏ʹ *OUFSGBDFͰఆٛ͢Δɻ 5SBJOFS .PEFM (SBEJFOU%FTDFOU

Architecture (7) Ὂ .BUSJYΠϯλʔϑΣʔεʹ͸ "EE΍4VCͳͲͷجຊతͳԋࢉ ͕࣮૷͞Ε͍ͯͳ͍ͨΊ (POVNͰχϡʔϥϧωοτ ϫʔΫΛߏங͍ͯ͘͠ͱҾ਺΍ ฦΓ஋Λɺߏ଄ମͰѻ͏͔Πϯ λʔϑΣʔεͰѻ͏͔೰ΈͲ͜ɻ

Architecture (9) Ὂ (POVN͚ͩͰ͸े෼ͳߦྻԋࢉ ͕ఏڙ͞Ε͍ͯͳ͍ͷͰɺ HPOOQͰ͸NBUVUJMύοέʔδ Λ࣮૷͍ͯ͠Δɻͦͷ৔߹͸ HPOVNNBUͷΑ͏ʹΠϯλʔ ϑΣʔεͰड͚ɺߏ଄ମͰฦ͢ ಺෦ͰܕεΠον͓ͯ͘͠ͱ࢖
͍΍͍͢ɻ

Architecture (8) Ὂ ϨΠϠʔʹؔͯ͠͸༷ʑͳϨΠ Ϡʔͷ࣮૷Λίϯϙʔωϯτͱ͠ ͯ࢖͍ճͤΔΑ͏ʹ*OUFSGBDFͰ ड͚ͯ*OUFSGBDFͰฦ࣮͢૷ʹ͠ ͍ͯΔɻ ϨΠϠʔ͸ߦྻΛड͚ Δ͔ϕΫτϧΛड͚Δ͔෼͔Βͳ
͍

Architecture (9) Ὂ 5FOTPS'MPXͷ࣮૷ͷΑ͏ʹσʔλϑϩʔάϥϑͬΆ͘ઃܭ͢Δͱ͍͏ߟ͑΋͋Δɻ

Architecture (9) Ὂ ࣮ࡍʹHPSHPOJBHPSHPOJB Ͱ͸νϟωϧʹΑΔσʔλϑ ϩʔͱͯ͠άϥϑΛߏங͠ɺ࣮ ߦ͢Δ࣮૷Λ͍࣋ͬͯΔɻ

Architecture (10) Ὂ ͜͏͢Δͱฒߦॲཧͳάϥϑॲ ཧΛ࣮૷Ͱ͖Δɻ HPOOQͰ͸ ·ͩશͯγʔέϯεͳͷͰฒߦ ॲཧΛαϙʔτ͍͖͍ͯͨ͠ ࢀߟʮ5IJOLMJLFBWFSUFYVTJOH(PTDPODVSSFODZGPSHSBQIDPNQVUBUJPOʯ https://blog.owulveryck.info/2019/10/14/think-like-a-vertex-using-gos-concurrency-for-graph-
computation.html

4: Store trained models

Store trained models (1) Ὂ (PͰֶशͨ͠Ϟσϧͷอଘ΍෼ࢄදݱͷอଘํ๏Λߟ͑Δɻ(PͰσʔλߏ଄ΛγϦΞϥ Πζ͢ΔͨΊʹ͍͔ͭ͘ͷύοέʔδ͕༻ҙ͞Ε͍ͯΔɻ ɾFODPSEJOHDTW ɾFODPSEJOHKTPO ɾFODPSEJOHYNM

Store trained models (2) Ὂ (Pͷ؀ڥ಺Ͱ׬݁͢ΔͷͰ͋Ε͹ɺFODPSEJOHHPC͕͓͢͢Ίɻσʔλߏ଄͑͞ ෼͔Ε͹ͦͷ··γϦΞϥΠζͰ͖Δɻ·ͨόΠφϦΤϯίʔσΟϯάͳͷͰޮ཰ੑ ΋ߴ͍ɻ ࢀߟʮ(PCTPGEBUBʯ5IF(P#MPH https://blog.golang.org/gobs-of-data

Store trained models (3) Ὂ ྫ͑͹8PSE7FDͰ֫ಘ͠ ͨ୯ޠͷ෼ࢄදݱΛ(PͷΞϓ ϦέʔγϣϯͰར༻͍ͨ͠৔ ߹͸ӈͷΑ͏ͳγϯϓϧͳ ίʔυ͚ͩͰ࠶ར༻Մೳͳܗ
ͰอଘͰ͖Δɻ

Store trained models (4) Ὂ ಠࣗͷΠϯλʔϑΣʔεܕΛHPCͰอଘ͍ͨ͠৔߹͸HPC3FHJTUFSΛ࢖࣮ͬͯଶͱ ͳΔσʔλߏ଄Λڭ͑Δඞཁ͕͋Δɻ࣮ࡍʹHPOOQʹ͓͍ͯ͸NBU.BUSJYΛJOJUؔ਺ ಺Ͱొ࿥͍ͯ͠Δ

Store trained models (5) Ὂ ҰํͰόΠφϦͰσʔλΛอଘ͢ΔϑΥʔϚο τͰ͋Δ.FTTBHF1BDLͷ(P࣮૷ͳͲ΋બ୒ ࢶʹೖΓ·͢ɻฐࣾͷ(PͰֶशͨ͠ϞσϧͰ͸ ͪ͜ΒΛ࢖ͬͯอଘ͍ͯ͠Δɻ

4: Use existing machine learning packages

Marchine Larning Packages Ὂ ಠ࣮ࣗ૷ͤͣͱ΋(PͰػցֶशΛ͢Δ؀ڥ͕੔͍ͭͭ͋Δɻ ίϯηϓτͱͯ͠͸5FOTPS'MPXʹ͍ۙɺάϥϑΛߏங࣮ͯ͠ߦ͍ͯ͘͠ɺ಺෦ͷߦ ྻԋࢉͰ͸HPSHPOJBUFOTPS͕࢖ΘΕ͍ͯΔɻHPOVNNBUͱͷޓ׵ੑ΋͋Δɻ

Marchine Larning Packages Ὂ (PͷطଘύοέʔδͰطଘϞσϧΛ࢖͏؀ڥ΋੔͍ͭͭ͋Δɻ ֶशࡁΈϞσϧΛ5FOTPS'MPX͕ఏڙ ͢Δ$"1*Λ࢖༻ͯ͠ಈ͔͢ɻ΋ͪΖΜ (P΋αϙʔτ͍ͯ͠Δɻ

Marchine Larning Packages Ὂ ֶशࡁΈϞσϧ 1SPUPDPM#V⒎FST 5FOTPS'MPX 1ZUIPOͰֶश 5FOTPS'MPX (PͰਪ࿦
1SPUPDPM#V⒎FSTܗࣜͷֶशࡁΈϞσϧͰ ͋Ε͹5FOTPS'MPX (PͰಈ͔ͤΔɻ

data-science Channel in Gophers Slack Ὂ (PͰσʔλαΠΤϯεΛ΍Δ ਓ͕ؒू·Δ4MBDLνϟωϧ͕ ͋Δɻ(PपΓͷσʔλαΠΤ ϯε৘ใ͸͜͜ʹूத͢Δͷ
ͰΦεεϝ

6: Conclusion

Conclusion Ὂ ͜͜·Ͱ(PͰͷػցֶशͷ࣮૷Λߦྻͷ࢓૊Έ͔Β঺հ͍͖ͯ͠·ͨ͠ɻ ࣮͸·ͩ·ͩ(PͰͷ਺஋ܭࢉ͸ൃల్্Ͱ͢ɻ΋͜͠ͷൃදͰগ͠Ͱ΋ڵຯΛ ࣋ͬͨํ͸ੋඇҰॹʹ(PʹΑΔσʔλαΠΤϯεΛ΍͍͖ͬͯ·͠ΐ͏ɻ

Suggestions and techniques for building neural networks with Go Go
Conference Tokyo 2019 Autumn / October 28, 2019 @ po3rin

Suggestions and techniques for building neural networks with Go

Suggestions and techniques for building neural networks with Go

More Decks by po3rin

Other Decks in Programming

Featured

Transcript