Upgrade to Pro — share decks privately, control downloads, hide ads and more …

ZOZO Researchの研究活動を支える Amazon SageMaker 利用 導入課...

ZOZO Researchの研究活動を支える Amazon SageMaker 利用 導入課題の解決と展開について / Using Amazon SageMaker to Support ZOZO Research Activities

Title
ZOZO Researchの研究活動を支える Amazon SageMaker 利用
導入課題の解決と展開について

Speaker
光野 達朗 Tatsuro Mitsuno (開発部SRE テックリード)
中村 拓磨 Takuma Nakamura (ZOZO Research リサーチサイエンティスト)

2019/06/14 AWS Summit Tokyo 2019 Day2 J2-03
https://aws.amazon.com/jp/summits/tokyo-2019/
#awssummit

Tatsuro Mitsuno

June 13, 2019
Tweet

More Decks by Tatsuro Mitsuno

Other Decks in Technology

Transcript

  1. ˜;0;05FDIOPMPHJFT *OD !10 ;0;03FTFBSDIͷϛογϣϯ ϑΝογϣϯΛ਺஋Խ͢Δɻ ʮ෰࡞Γʯͷݚڀ ΦʔτσβΠϯɺ৽ૉࡐɺ৽ੜ࢈ϥΠϯͳͲ ෰࡞Γશൠͷݚڀ ʮࣅ߹͏ʯͷݚڀ ࣅ߹͏෰ͱ͸ʁ

    ૉఢͳ૊Έ߹Θͤͱ͸ʁ ਪનΞϧΰϦζϜͷݚڀ ʮ࠾ੇʯͷݚڀ ଍ͳͲͷ຤୺෦ͷ࠾ੇɺಈ͖ɺے೑ͳͲͷܭଌɺ ମܕܭଌʹؔ͢Δஎ·͵ਐԽ
  2. ˜;0;05FDIOPMPHJFT *OD σʔλYػցֶश !12 Takuma Nakamura, Ryosuke Goto. Outfit Generation

    and Style Extraction via Bidirectional LSTM and Autoencoder. KDDW. https://arxiv.org/abs/1807.03133 தଜ୓ຏ, ࡈ౻ါً. (2019) “Set TransformerʹΑΔϑΝογϣϯίʔσΟωʔτબ୒”, ৘ใ࿦తֶश ཧ࿦ͱػցֶशݚڀձ (IBISML). https://speakerdeck.com/yukisaito/set-transformer-for-coordinating- outfits
  3. ˜;0;05FDIOPMPHJFT *OD ܭࢉج൫ͷδϨϯϚ Amazon Elastic Compute Cloud (Amazon EC2) ೥݄

    ʢ;0;03FTFBSDIൃ଍ʣ ೥/݄ ೥.݄    !15 (16ΠϯελϯεΛඞཁ୆਺ɺΠϯϑϥνʔϜʢ౰࣌ʣ͕༻ҙ͢Δ
  4. ˜;0;05FDIOPMPHJFT *OD ܭࢉج൫ͷδϨϯϚ Amazon Elastic Compute Cloud (Amazon EC2) ೥݄

    ʢ;0;03FTFBSDIൃ଍ʣ ೥/݄ ೥.݄    !16 (16ΠϯελϯεΛඞཁ୆਺ɺΠϯϑϥνʔϜʢ౰࣌ʣ͕༻ҙ͢Δ ࿦จࣥචͳͲͷΠϕϯτ࣌ʹෆ଍ ୆਺͕૿͑ɺશମͷՔಇ཰͕௿Լ े෼ͳ୆਺Λ४උ ୆਺࡟ݮ ʢ܁Γฦ͠ʣ
  5. ˜;0;05FDIOPMPHJFT *OD ܭࢉج൫ͷδϨϯϚ Amazon Elastic Compute Cloud (Amazon EC2) ೥݄

    ʢ;0;03FTFBSDIൃ଍ʣ ೥/݄ ೥.݄    !17 ৗʹɺे෼ͳ୆਺Λ༻ҙ͢Δ ৗʹɺՔಇ཰Λߴ͘อͭ ཱ͍྆ͨ͠ Amazon SageMaker (SageMaker) ೥݄
  6. ˜;0;05FDIOPMPHJFT *OD "NB[PO4BHF.BLFS #VJME 5SBJO %FQMPZ w ϫϯΫϦοΫτϨʔχϯά
 ඞཁͳΠϯελϯελΠϓΛඞཁͳ୆਺ɺग़ྗ͸4΁ w

    ϫϯΫϦοΫσϓϩΠ w ࣗಈ"#ςετ w Ϛωʔδυ+VQZUFSϊʔτϒοΫ w ෯޿͍ϑϨʔϜϫʔΫΛαϙʔτ
 ʢ5FOTPS'MPXɺ"QBDIF.9/FUɺ$IBJOFSɺ1Z5PSDIɺ4DJLJUMFBSOɺ4QBSL.-ʣ "NB[PO4BHF.BLFS͸ɺػցֶशͷϫʔΫϑϩʔશମΛΧόʔ͢Δ׬શϚωʔδυܕαʔϏε
 DGIUUQTBXTBNB[PODPNKQTBHFNBLFS !18
  7. ˜;0;05FDIOPMPHJFT *OD "NB[PO4BHF.BLFS #VJME 5SBJO %FQMPZ w ϫϯΫϦοΫτϨʔχϯά
 ඞཁͳΠϯελϯελΠϓΛඞཁͳ୆਺ɺग़ྗ͸4΁ w

    ϫϯΫϦοΫσϓϩΠ w ࣗಈ"#ςετ w Ϛωʔδυ+VQZUFSϊʔτϒοΫ w ෯޿͍ϑϨʔϜϫʔΫΛαϙʔτ
 ʢ5FOTPS'MPXɺ"QBDIF.9/FUɺ$IBJOFSɺ1Z5PSDIɺ4DJLJUMFBSOɺ4QBSL.-ʣ "NB[PO4BHF.BLFS͸ɺػցֶशͷϫʔΫϑϩʔશମΛΧόʔ͢Δ׬શϚωʔδυܕαʔϏε
 DGIUUQTBXTBNB[PODPNKQTBHFNBLFS !19
  8. ˜;0;05FDIOPMPHJFT *OD ϫϯΫϦοΫτϨʔχϯά !20 5SBJOJOH" 5SBJOJOH# 5SBJOJOH$ w QYMBSHF w

    ୆ w εΫϦϓτ" w QYMBSHF w ୆ w εΫϦϓτ# w QYMBSHF w ୆ w εΫϦϓτ$ w ܭࢉࢿݯͷઃఆ w ֶशεΫϦϓτ ֶशͭͭʹର͠ɺ ඞཁͳܭࢉࢿݯΛׂΓ౰ͯΔ ֶशͷऴྃͱڞʹ5FSNJOBUF
  9. ˜;0;05FDIOPMPHJFT *OD !21 ݁Ռɿֶश͸εέʔϧͨ͠      

       5SBJOJOH+0# $PNQMFUFE 4UPQQFE ϲ݄Ͱطʹલ2௒͑
  10. ˜;0;05FDIOPMPHJFT *OD !27 ΫΤϦ ༧ଌ݁Ռ     

              ࠨ୺ͷΞΠςϜ͕HSPVOEUSVUI ݚڀ੒ՌϑΝογϣϯίʔσΟωʔτͷܽଛ༧ଌ
  11. ˜;0;05FDIOPMPHJFT *OD !29 Style A Style B Style C Style

    D Style A Style B Style C Sty Style A + Style B S +0.5⇥ <latexit sha1_base64="tcrm8is1sjUv+vEDWvmqA/D6eTA=">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</latexit> <latexit sha1_base64="tcrm8is1sjUv+vEDWvmqA/D6eTA=">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</latexit> <latexit sha1_base64="tcrm8is1sjUv+vEDWvmqA/D6eTA=">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</latexit> <latexit sha1_base64="tcrm8is1sjUv+vEDWvmqA/D6eTA=">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</latexit> 0.5⇥ <latexit sha1_base64="BLMXtHMoAf6LIq7CCRWO/fVnk4Q=">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</latexit> <latexit sha1_base64="BLMXtHMoAf6LIq7CCRWO/fVnk4Q=">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</latexit> <latexit sha1_base64="BLMXtHMoAf6LIq7CCRWO/fVnk4Q=">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</latexit> <latexit sha1_base64="BLMXtHMoAf6LIq7CCRWO/fVnk4Q=">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</latexit> = <latexit sha1_base64="mKNhi4l4+8NwtmcIDF9c3XG2cjo=">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</latexit> <latexit sha1_base64="mKNhi4l4+8NwtmcIDF9c3XG2cjo=">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</latexit> <latexit sha1_base64="mKNhi4l4+8NwtmcIDF9c3XG2cjo=">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</latexit> <latexit sha1_base64="mKNhi4l4+8NwtmcIDF9c3XG2cjo=">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</latexit> ݚڀ੒ՌϑΝογϣϯελΠϧͷΫϥελϦϯάͱԋࢉ
  12. ˜;0;05FDIOPMPHJFT *OD !30 Query item image Generated outfit Query style

    Mixture ratio Style A Style B Style C Style D Style A ݚڀ੒ՌϑΝογϣϯελΠϧຖͷίʔσΟωʔτੜ੒
  13. ˜;0;05FDIOPMPHJFT *OD !36 ։ൃ ֶश ෳ਺ͷαʔόͰฒߦͯ͠։ൃ ͯͦ͠ΕͧΕ࣮ߦ ಋೖલ ਪ࿦ ରࡦޙ

    ಋೖ ௚ޙ ʁ ʁ ʁ ʁ ʁ ʁ ར༻ଅਐͷҝͷऔΓ૊Έ &'4΍&$಺ͷετϨʔδʹ εφοϓγϣοτΛอଘ &'4ʹอଘ͞ΕֶͨशࡁΈ ϞσϧΛςετ༻&$Ͱ ϩʔυ࣮ͯ͠ߦ
  14. ˜;0;05FDIOPMPHJFT *OD !37 ։ൃ ֶश ෳ਺ͷαʔόͰฒߦͯ͠։ൃ ͯͦ͠ΕͧΕ࣮ߦ ಋೖલ ਪ࿦ ରࡦޙ

    ಋೖ ௚ޙ ʁ ʁ ʁ ʁ ʁ ʁ ར༻ଅਐͷҝͷऔΓ૊Έ &'4΍&$಺ͷετϨʔδʹ εφοϓγϣοτΛอଘ &'4ʹอଘ͞ΕֶͨशࡁΈ ϞσϧΛςετ༻&$Ͱ ϩʔυ࣮ͯ͠ߦ
  15. ˜;0;05FDIOPMPHJFT *OD !38 Πϯελϯεຖʹಠཱʹ࣮ݧͱίʔυमਖ਼Λ܁Γฦ͍ͯͨ͠ ͕࣌ؒܦͭʹ࿈Εͯίʔυͱ࣮ݧͷόʔδϣϯ؅ཧ͕ෳࡶʹͳΓɺ͔ͭϦιʔε΋࢖͍੾Ε͍ͯͳ͍ ։ൃ ಋೖલ ͜Ε·Ͱͷ։ൃํ๏ͱ4BHF.BLFS΁ͷظ଴ training A

    notebook training B training C 4BHF.BLFSΛಋೖͯ͠ ࣮ݧ؀ڥΛՕॴʹू໿ ͢Δ͜ͱΛظ଴ͨ͠ scripts developed on the instance A training A1 training A2 scripts' developed on the instance A scripts'' developed on the instance A scripts developed on the instance B training B1 training B2 scripts' developed on the instance B scripts'' developed on the instance B scripts developed on the instance C training C1 training C2 scripts' developed on the instance C scripts'' developed on the instance C
  16. ˜;0;05FDIOPMPHJFT *OD !39 ։ൃ ಋೖ௚ޙ ݦࡏԽͨ͠OPUFCPPLΠϯελϯεͷ໰୊ ☀δϣϒͱ͍͏୯ҐͰ࣮ݧΛ؅ཧͰ͖Δ ☀࣮ݧΛΫϩʔϯͰ͖Δ ☀՝ֹۚΛ࠷খԽͰ͖Δ +VQZUFS΍ΤσΟλͷΧελϚΠζ͕Ͱ͖ͳ͍

    TTIͰ͖ͳ͍ JQZOCϑΝΠϧͱHJUͷ૬ੑ͕ѱ͍ OPUFCPPLΠϯελϯεΛ։ൃ؀ڥͱͯ͠ݟΔͱɺॊೈੑʹ͚ܽΔ఺͕໨ཱͬͨ ͜Ε·Ͱͷ։ൃ؀ڥͷ֦ுੑ͕ࣦΘΕΔͱڞʹɺJQZOCϑΝΠϧͷ؅ཧίετ͕ՃΘͬͨ ֶशδϣϒ࣮ߦ ։ൃ training A training B training C foo.py bar.py baz.yml
  17. ˜;0;05FDIOPMPHJFT *OD !40 ։ൃ ରࡦޙ OPUFCPPLΠϯελϯεΛܦ༝͠ͳ͍ҝʹ͸ʜ training A training B

    training C foo.py bar.py baz.yml training A training B training C foo.py bar.py baz.yml OPUFCPPLΠϯελϯεͰͷ։ൃ࡞ۀΛ։ൃػʹҠߦ͠ɺ։ൃػ͔ΒδϣϒΛ࣮ߦ͢Δ ։ൃػʹ4BHF.BLFS4%,ΛΠϯετʔϧͯ͠ɺOPUFCPPLΠϯελϯεͱಉ౳ͷ໾ׂΛ༩͑Δ QJQJOTUBMMTBHFNBLFS
  18. ˜;0;05FDIOPMPHJFT *OD !41 ։ൃ ରࡦޙ 4%,Λར༻ͯ͠։ൃػ͔Β4BHF.BLFSʹ௚઀ֶशδϣϒΛ౤͛ΔΑ͏ʹͨ͠ OPUFCPPLΠϯελϯε͕ෆཁͱͳͬͨ OPUFCPPLʹهड़͍࣮ͯͨ͠ݧ৚݅ͱֶशδϣϒ࣮ߦʹؔ͢ΔॲཧΛ෼ղͨ͠ ࣮ݧ৚݅͸ZBNMʹ੾Γग़ͯ͠HJUͰ؅ཧɺֶशδϣϒ࣮ߦʹؔ͢Δॲཧ͸$-*ͱ࣮ͯ͠૷ͨ͠ ࣮ݧ৚݅

    ֶशδϣϒ࣮ߦ ࣮ݧ৚݅ZNM ֶश࣮ߦ$-* 4BHF.BLFS 4%,Λϥοϓ training A training B training C training D confA.yml confB.yml confC.yml confD.yml OPUFCPPLΛ࢖Θͣʹ։ൃػ͔ΒֶशδϣϒΛ࣮ߦ͢Δ FYBNQMFJQZOC ZBNMΛىಈҾ਺ͱ͢Δ ίϚϯυͰֶशδϣϒΛ࣮ߦ
  19. ˜;0;05FDIOPMPHJFT *OD !42 ։ൃ ֶश ෳ਺ͷαʔόͰฒߦͯ͠։ൃ ͯͦ͠ΕͧΕ࣮ߦ ಋೖલ ਪ࿦ ରࡦޙ

    ಋೖ ௚ޙ ։ൃޮ཰௿ԼͱJQZOC ϑΝΠϧͷ؅ཧίετ૿ ʁ ʁ ࣮ݧ৚݅ͱ࣮ߦΛ෼཭ ։ൃػ͔Β௚઀δϣϒ࣮ߦ ʁ ʁ &'4΍&$಺ͷετϨʔδʹ εφοϓγϣοτΛอଘ &'4ʹอଘ͞ΕֶͨशࡁΈ ϞσϧΛςετ༻&$Ͱ ϩʔυ࣮ͯ͠ߦ ར༻ଅਐͷҝͷऔΓ૊Έ
  20. ˜;0;05FDIOPMPHJFT *OD !43 ։ൃ ֶश ෳ਺ͷαʔόͰฒߦͯ͠։ൃ ͯͦ͠ΕͧΕ࣮ߦ ಋೖલ ਪ࿦ ରࡦޙ

    ಋೖ ௚ޙ JQZOCϑΝΠϧͷ؅ཧίετ ૿ͱ։ൃޮ཰௿Լ ʁ ʁ ࣮ݧ৚݅ͱ࣮ߦΛ෼཭ ։ൃػ͔Β௚઀δϣϒ࣮ߦ ʁ ʁ &'4΍&$಺ͷετϨʔδʹ εφοϓγϣοτΛอଘ &'4ʹอଘ͞ΕֶͨशࡁΈ ϞσϧΛςετ༻&$Ͱ ϩʔυ࣮ͯ͠ߦ ར༻ଅਐͷҝͷऔΓ૊Έ
  21. ˜;0;05FDIOPMPHJFT *OD !47 TOBQTIPU TOBQTIPU@PCKFDU -PH3FQPSU 1MPU3FQPSU ֶश ରࡦޙ ϑΝΠϧసૹͱϩάసૹͷ࢓૊Έ

    ࠷৽ͷεφοϓγϣοτ͸#PUPΛར༻ͯ͠4ʹసૹ͢Δ ֶशͷϩά͸*ODPNJOH8FC)PPLΛར༻ͯ͠4MBDLʹϙετ͢Δ $IBJOFSͷग़ྗΛ֎෦πʔϧʹ࿈ܞ
  22. ˜;0;05FDIOPMPHJFT *OD !50 ։ൃ ֶश ෳ਺ͷαʔόͰฒߦͯ͠։ൃ ͯͦ͠ΕͧΕ࣮ߦ ಋೖલ &'4΍&$಺ͷετϨʔδʹ εφοϓγϣοτΛอଘ

    ਪ࿦ &'4ʹอଘ͞ΕֶͨशࡁΈ ϞσϧΛςετ༻&$Ͱ ϩʔυ࣮ͯ͠ߦ ରࡦޙ ಋೖ ௚ޙ ֶश్தʹεφοϓγϣοτ ΛಘΔ͜ͱ͕Ͱ͖ͳ͍ ʁ ࣮ݧ৚݅ͱ࣮ߦΛ෼཭ ։ൃػ͔Β௚઀δϣϒ࣮ߦ %-ϑϨʔϜϫʔΫͷ֦ுػ ೳΛ࢖͍ఆظతʹεφοϓ γϣοτΛ4ʹసૹ ʁ JQZOCϑΝΠϧͷ؅ཧίετ ૿ͱ։ൃޮ཰௿Լ ར༻ଅਐͷҝͷऔΓ૊Έ
  23. ˜;0;05FDIOPMPHJFT *OD !51 ։ൃ ֶश ෳ਺ͷαʔόͰฒߦͯ͠։ൃ ͯͦ͠ΕͧΕ࣮ߦ ಋೖલ &'4΍&$಺ͷετϨʔδʹ εφοϓγϣοτΛอଘ

    ਪ࿦ ରࡦޙ ಋೖ ௚ޙ ֶश్தʹεφοϓγϣοτ ΛಘΔ͜ͱ͕Ͱ͖ͳ͍ ʁ ࣮ݧ৚݅ͱ࣮ߦΛ෼཭ ։ൃػ͔Β௚઀δϣϒ࣮ߦ %-ϑϨʔϜϫʔΫͷ֦ுػ ೳΛ࢖͍ఆظతʹεφοϓ γϣοτΛ4ʹసૹ ʁ &'4ʹอଘ͞ΕֶͨशࡁΈ ϞσϧΛςετ༻&$Ͱ ϩʔυ࣮ͯ͠ߦ JQZOCϑΝΠϧͷ؅ཧίετ ૿ͱ։ൃޮ཰௿Լ ར༻ଅਐͷҝͷऔΓ૊Έ
  24. ˜;0;05FDIOPMPHJFT *OD !52 ಋೖલ ਪ࿦ &'4ʹग़ྗ͞ΕֶͨशࡁΈϞσϧ Λϩʔυͯ͠ςετΛ࣮ߦ ϑΝΠϧͷҠಈͳ͠ͰςετΛ࣮ߦͰ͖ͯ ͍ͨҝɺͦ͜·Ͱ໘౗ʹײ͍ͯ͡ͳ͔ͬͨ model

    C model A model B test A test B test C model B training B model A training A model C training C test scripts test data &$؀ڥͰͷϞσϧͷςετ͸&'4͕த৺ ֶश༻Πϯελϯε ςετ༻Πϯελϯε
  25. ˜;0;05FDIOPMPHJFT *OD !53 4ʹग़ྗ͞ΕֶͨशࡁΈϞσϧΛ ςετ༻&$ʹίϐʔͯ͠ςετΛ ࣮ߦ .-Πϯελϯεʹ&'4͕Ϛ΢ϯτͰ͖ͳ͍ ͨΊɺֶशࡁΈϞσϧͷ഑෍͕4ܦ༝ͱͳ Γɺίϐʔͷख͕ؒ૿͑ͨ model

    C model A model B teat A test B test C model A training A model B training B model C training C test scripts test data खಈͰίϐʔ 4த৺ͷϞσϧ഑෍͸खಈͷ࡞ۀ͕ൃੜͨ͠ ಋೖ௚ޙ ਪ࿦ ςετ༻Πϯελϯε
  26. ˜;0;05FDIOPMPHJFT *OD όονม׵δϣϒΛར༻ͯ͠ςετΛ࣮ߦ͢ΔΑ͏ʹͨ͠ !54 ରࡦޙ ਪ࿦ ֶश݁ՌΛͦͷ··Ϟσϧʹొ࿥ ֶश݁Ռͷ4ϩέʔγϣϯΛࢦఆͯ͠Ϟσ ϧΛ࡞੒͢Δ ొ࿥ͨ͠ϞσϧΛࢀরͯ͠όον

    ม׵δϣϒΛఆٛ Ϟσϧͱ4ʹ഑ஔ͞ΕͨσʔλΛར༻ͯ͠ ςετΛ࣮ߦՄೳ model C model A model B test A test B test C training A training A training A registered model C registered model A registered model B test data test data test data
  27. ˜;0;05FDIOPMPHJFT *OD όονม׵ػೳΛར༻ͯ͠ςετΛ࣮ߦ͢ΔΑ͏ʹͨ͠ !55 model C model A model B

    test A test B test C training A training A training A registered model C registered model A registered model B test data test data test data ରࡦޙ ਪ࿦ ֶश݁ՌΛͦͷ··Ϟσϧʹొ࿥ ֶश݁Ռͷ4ϩέʔγϣϯΛࢦఆͯ͠Ϟσ ϧΛ࡞੒͢Δ ొ࿥ͨ͠ϞσϧΛࢀরͯ͠όον ม׵δϣϒΛఆٛ Ϟσϧͱ4ʹ഑ஔ͞ΕͨσʔλΛར༻ͯ͠ ςετΛ࣮ߦՄೳ Ϟσϧͷ࡞੒
  28. ˜;0;05FDIOPMPHJFT *OD όονม׵ػೳΛར༻ͯ͠ςετΛ࣮ߦ͢ΔΑ͏ʹͨ͠ !56 model C model A model B

    test A test B test C training A training A training A registered model C registered model A registered model B test data test data test data ରࡦޙ ਪ࿦ ֶश݁ՌΛͦͷ··Ϟσϧʹొ࿥ ֶश݁Ռͷ4ϩέʔγϣϯΛࢦఆͯ͠Ϟσ ϧΛ࡞੒͢Δ ొ࿥ͨ͠ϞσϧΛࢀরͯ͠όον ม׵δϣϒΛఆٛ Ϟσϧͱ4ʹ഑ஔ͞ΕͨσʔλΛར༻ͯ͠ ςετΛ࣮ߦՄೳ όονม׵δϣϒ
  29. ˜;0;05FDIOPMPHJFT *OD Ϟσϧ࡞੒͔Βόονม׵δϣϒ·ͰΛ$-*Ͱ࣮૷ !57 model C model A model B

    test A test B test C training A training A training A registered model C registered model A registered model B test data test data test data ରࡦޙ ਪ࿦ 4%,Λ༻͍ͯ։ൃػ͔Β ςετΛ࣮ߦ ςετ࣮ߦ$-*
  30. ˜;0;05FDIOPMPHJFT *OD !58 ։ൃ ֶश ෳ਺ͷαʔόͰฒߦͯ͠։ൃ ͯͦ͠ΕͧΕ࣮ߦ ಋೖલ ਪ࿦ &'4ʹอଘ͞ΕֶͨशࡁΈ

    ϞσϧΛςετ༻&$Ͱ ϩʔυ࣮ͯ͠ߦ ରࡦޙ ಋೖ ௚ޙ ֶश్தʹεφοϓγϣοτ ΛಘΔ͜ͱ͕Ͱ͖ͳ͍ ֶशࡁΈϞσϧΛ4͔Β ςετ༻&$ʹμ΢ϯϩʔ υ͢Δ޻ఔ͕૿͑ͨ ࣮ݧ৚݅ͱ࣮ߦΛ෼཭ ։ൃػ͔Β௚઀δϣϒ࣮ߦ %-ϑϨʔϜϫʔΫͷ֦ுػ ೳΛ࢖͍ఆظతʹεφοϓ γϣοτΛ4ʹసૹ όονม׵δϣϒػೳΛར༻ ͯ͠&$Λհͣ͞ςετ࣮ߦ &'4΍&$಺ͷετϨʔδʹ εφοϓγϣοτΛอଘ JQZOCϑΝΠϧͷ؅ཧίετ ૿ͱ։ൃޮ཰௿Լ ར༻ଅਐͷҝͷऔΓ૊Έ
  31. ˜;0;05FDIOPMPHJFT *OD !59 ։ൃ ֶश ෳ਺ͷαʔόͰฒߦͯ͠։ൃ ͯͦ͠ΕͧΕ࣮ߦ ಋೖલ &'4΍&$಺ͷετϨʔδʹ εφοϓγϣοτΛอଘ

    ਪ࿦ &'4ʹอଘ͞ΕֶͨशࡁΈ ϞσϧΛςετ༻&$Ͱ ϩʔυ࣮ͯ͠ߦ ରࡦޙ ಋೖ ௚ޙ JQZOCϑΝΠϧͷ؅ཧίετ ૿ͱ։ൃޮ཰௿Լ ֶश్தʹεφοϓγϣοτ ΛಘΔ͜ͱ͕Ͱ͖ͳ͍ ֶशࡁΈϞσϧΛ4͔Β ςετ༻&$ʹμ΢ϯϩʔυ ͢Δ޻ఔ͕૿͑ͨ ࣮ݧ৚݅ͱ࣮ߦΛ෼཭ ։ൃػ͔Β௚઀δϣϒ࣮ߦ %-ϑϨʔϜϫʔΫͷ֦ுػ ೳΛ࢖͍ఆظతʹεφοϓ γϣοτΛ4ʹసૹ όονม׵δϣϒػೳΛར༻ ͯ͠&$Λհͣ͞ςετ࣮ߦ ར༻ଅਐͷҝͷऔΓ૊Έ
  32. ˜;0;05FDIOPMPHJFT *OD ։ൃػΛத৺ͱͯ͠࡞ۀΛγϯϓϧʹ͢Δ͜ͱʹ੒ޭͨ͠ ։ൃ؀ڥ͸͜Ε·Ͱͷ؀ڥΛྲྀ༻͠ɺֶशਪ࿦͸$-*Ͱ࣮ߦ͠ɺεφοϓγϣοτͷอଘ΋ࣗಈԽͨ͠ ར༻ଅਐͷҝͷऔΓ૊Έޙ !60 ./train ./test train.py test.py

    conf.yml test.py conf.yml test data train data train.py conf.yml job job job w ։ൃ؀ڥ͸ ୆ʹू໿ ֶशδϣϒ όονม׵δϣϒ w ֶशͱςετ͸։ൃػ ͔Β4%,Ͱ࣮ߦ w ֶशաఔ͸֎෦ πʔϧʹग़ྗ w ιʔείʔυͱϞσϧ ͸δϣϒຖʹอଘ