離散データ生成モデル for ファッションテック

ファッションコーディネートの生成をGANで実現した話です。
GANで離散データを扱う上で大変だったことなど、奮闘の記録を話しました。

Transcript

1. ཭ࢄσʔλੜ੒Ϟσϧ for ϑΝογϣϯςοΫ ߥҪߒݾ
   'SPN
   גࣜձࣾχϡʔϩʔϓ
2. χϡʔϩʔϓͬͯʁ • ϑΝογϣϯº"*ͰϑΝογϣϯ ۀքͷൃలΛ͓ख఻͍
   • ਂ૚ֶशΛ࢝Ίͱͨ͠ػցֶशٕ ज़
   • ಠࣗͷσʔληοτΛ࣋ͭ
   • ਺ଟ͘ͷϞσϧ͞Μͱఏܞ
   • ϝσΟΞࣄۀ
   • 8FC։ൃʹ΋ڧΈ
   • ࠷৽ٕज़Λ࢖͍΍͍͢ܗͰఏڙ
3. ࣗݾ঺հ ໊લ
   ߥҪߒݾ
   (JU)VC
   IJSPBSB
   'BDFCPPL
   GBDFCPPLDPNIJSPBSB
   $50
   PG
   /&8301&
   ܦྺ
   4*FSͷ4&
   
   8FC։ൃ
   
   ػցֶशͱ͔Λ΍ΔΤϯδ χΞ
   όοΫάϥ΢ϯυ
   ਓؒՊֶ౷ܭ
   ޷͖ͳਂ૚ֶशϑϨʔϜϫʔΫ
   5FOTPSqPX
   4POOFU
   ,FSBT͸ͪΐͬͱۤख
4. χϡʔϩʔϓͰѻ͍ͬͯΔϞσϧ • ਂ૚ֶशܥ • ը૾ೝࣝ • ෺ମݕग़ • Ϛϧνϥϕϧೝࣝ • ཭ࢄσʔλੜ੒Ϟσϧ • ࠷దԽ • όϯσΟοτΞϧΰϦζϜ • ࣗಈ෼ྨ • φΠʔϒϕΠζ • ౷ܭʹΑΔΞυϗοΫ෼ੳ • ॏճؼܥ/ܾఆ໦ܥ/SVMͳͲҰൠత͔ͭϥΠϒϥϦ͕࢖͑Δ΋ͷ
5. We are hiring! • ઌʹݴ͓ͬͯ͘ͱҰॹʹϑΝογϣϯςοΫʹऔΓ૊Ή ஥ؒΛઈࢍืूதͰ͢ • ओʹWebΤϯδχΞΛืूத • ػցֶश΍౷ܭ෼ੳʹڵຯ͕͋Δਓɺܦݧऀ΋େ׻ܴ • ·ͩ·ͩਓ਺͕গͳ͍ͷͰɺࣗ༝ͳ։ൃ؀ڥ͕଴͍ͬͯ ·͢ʂ • ग़ࣾ͸िࡾճͷΈͰ͋ͱ͸ϦϞʔτʂ
6. ࠓճऔΓ্͛Δ͜ͱ • ਂ૚ֶशܥ • ը૾ೝࣝ • ෺ମݕग़ • Ϛϧνϥϕϧೝࣝ • ཭ࢄσʔλੜ੒Ϟσϧ • ࠷దԽ • όϯσΟοτΞϧΰϦζϜ • ࣗಈ෼ྨ • φΠʔϒϕΠζ • ౷ܭʹΑΔΞυϗοΫ෼ੳ • ॏճؼܥ/ܾఆ໦ܥ/SVMͳͲҰൠత͔ͭϥΠϒϥϦ͕࢖͑Δ΋ͷ
7. എܠ • ౰࣌ը૾ೝࣝ͸·͋·͋ͷਫ਼౓Λୡ੒ɺҾ͖߹͍΋͋ͬ ͨ • ͔͠͠ɺը૾ೝࣝҰຊ͸ϏδωεతʹϋΠϦεΫ • ࠩผԽՄೳͳ΋ͷΛ࡞Εͳ͍͔ʁ • χϡʔϩʔϓ͸ίʔσΟωʔτσʔλΛେྔʹ͍࣋ͬͯΔ • GAN΋ྲྀߦ͍ͬͯΔ͠ɺੜ੒ͯ͠ϦίϝϯυͰ͖ΔΑ͏ ʹͯ͠ΈΑ͏ʂ
8. ࡞Γ͍ͨ΋ͷ • τοϓε΍ϘτϜεͳͲɺͭ ͳ͍͠ෳ਺ͷΞΠςϜΛೖྗ͢ Δ
   • ΫΤϦΞΠςϜΛؚΊͨΞΠς Ϝͷηοτ ίʔσΟωʔτ Λ ੜ੒
   • ྫ͑͹ɺΧʔτը໘Ͱͷ߹Θͤ ങ͍ૂ͍ͷϦίϝϯυͳͲʹ࢖ ͑Δ ΫΤϦ: ΞΠςϜू߹ X = {xi ; i ∈ ℕ} ݁Ռ: ΞΠςϜू߹ Y = {yi ; i ∈ ℕ}
9. GANͷ͓͞Β͍ • Generator vs DiscriminatorͷߏਤͰͦΕͧΕֶशΛߦ͍(͍͍ͩͨਂ૚ֶश)ɺΑΓ޼ ົͳfakeσʔλΛ࡞Δੜ੒ϞσϧͷϑϨʔϜϫʔΫ • Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, Yoshua Bengio. abs/1406.2661, 2014 V.S. Generator (ِ଄ऀ) Discriminator (ܯ࡯) min G max D V(D, G) = x∼pdata (x) [logD(x)] + z∼pz (z) [log(1 − D(G(z)))] Real or Fake
10. ౰ॳͷ໨࿦ݟ • GANͷݚڀ΍ࣄྫ͸ը૾ੜ੒͕ओ͕ͩɺϞσϧతʹͦ ͷଞͷλεΫʹ΋Ԡ༻Մೳͳ͸ͣ • ίʔσΟωʔτσʔλ͸ը૾ΑΓ΋σʔλαΠζ͕খ ͍͞ͷͰɺܰྔͳϞσϧ͕ར༻Մೳɺֶशࣗମ΋ૣ͍ ͸ͣʂ • σʔλ͸͋ΔͷͰద౰ʹϞσϧΛ૊ΜͰ૸ΒͤΕ͹͏ ·͍͘͘ʂ͸ͣʂ
11. ͦͯ͠΍͸ΓݱΕΔน • ऩଋ͠ͳ͍ • Generator͕Discriminatorʹॠࡴ͞Εɺֶश͠ͳ͍ • ࣄ࣮ͱͯ͠σʔλαΠζ͸খֶ͘͞शॲཧࣗମ͸ߴ଎ ͕ͩɺผͷ༷ʑͳ໰୊͕ݱΕΔ
12. ࿈ଓ vs ཭ࢄ • ը૾
    3(# CJU ͷϐΫηϧσʔλΛར ༻
    • ܭࢉ্͸࿈ଓͨ͠ൣғ<
    >ͷ஋Λ ͱΔqPBUϕΫτϧͱͯ͠ѻ͑Δ
    • ίʔσ
    ΧςΰϦ POFIPU ͱϥϕϧ NVMUJMBCFMT ͷ૊Έ߹ΘͤͰදݱ͞Ε Δ
    • ཭ࢄ஋Ͱɺ͔͠΋CPPMFBO
    εΧʔ τ͔ɺͦ͏Ͱͳ͍͔ͳͲ
    • ۃ୺ͳྫͱ͔ͯͭ͠͠ϥϕϧ͕ଘ ࡏ͠ͳ͍ঢ়گΛߟ͑Δͱɺཱํମͷ ֤௖఺ʹ͔͠܇࿅σʔλ͕ଘࡏ͠ͳ ͍ঢ়ଶ εΧʔτ ϗϫΠτ ϛϞϨ
13. Dense vs Sparse • ը૾
    શͯͷϐΫηϧʹԿΒ͔ͷ3(# ஋͕ଘࡏ͢Δ
    • ଟ༷ੑ΋ඇৗʹߴ͍
    • ίʔσ
    ΞΠςϜ਺͕ฏۉݸɺ֤Ξ ΠςϜʹ͚ͭΒΕΔϥϕϧ਺͸ฏۉ ݸఔ౓
    • σʔλۭؒͷ͏ͪɺτϨʔχϯά σʔλ͕ଘࡏ͢ΔྖҬ͕ۃ୺ʹภͬ ͍ͯΔ
    • ྫ͑͹ֶशॳظʹੜ੒͞ΕಘΔʮϥ ϕϧݸΛ࣋ͭΞΠςϜʯͳͲ͸ %JTDSJNJOBUPSʹΑΓଈ٫Լ͞ΕΔ τοϓε ϘτϜε ๧ࢠ Կ΋ͳ͍ΤϦΞ
14. ໰୊఺Λ੔ཧ • ཁ͸σʔλͷ࣍ݩ਺ʹରͯ͠σʔλ ͷଘࡏ͢ΔྖҬ͕ۃ୺ʹখ͘͞ɺ͔ ͭඈͼ஍ʹͳ͍ͬͯΔ
    • %JTDSJNJOBUPS
    • ༰қʹࣝผ͕ͭ͘ঢ়گ ඈͼ஍Ҏ ֎ʹ͍Δ΍ͭ͸શͯGBLF
    • (FOFSBUPS
    • ·ͣ࠷దղͷํ֯͢ΒΘ͔Βͳ͍
    • ݟ͚ͭͯ΋ہॴ࠷దղʹ༰қʹؕ Γɺඈͼ஍Λӽ͑Δͷ͕ͱͯ΋೉ ͍͠ Global Minimum पลͷloss஋͕ݢฒΈߴ͍ͷͰ ࠷దղͷ͋Δํ֯͢ΒΘ͔Βͳ͍ Loss Local Minimum
15. ֶशՄೳͳޯ഑ΛಘΔͨΊͷଛࣦؔ਺͕ཉ͍͠ • ෼෍ؒͷڑ཭Λߟ͑Δଛࣦؔ਺Λ࣋ͭGAN • Wasserstein GAN: Earth Mover's distance • ਅͷσʔλͱϞσϧͷ෼෍͕શ͘ॏͳΒͳ͍৔߹ɺKL divergence͕ఆٛͰ͖ͳ͍໰୊ʹରԠ͢ΔͨΊɺEarth Mover's distanceΛར༻ • Martin Arjovsky, Soumith Chintala, Léon Bottou. abs/1701.07875, 2017 • Cramer GAN: Cramer distance (Energy distance) • Wasserstein GANͷ༻͍Δڑ཭ؔ਺ʹόΠΞε͕ଘࡏ͢ΔͱࢦఠɺCramer distanceΛར༻͢Δ͜ͱͰ͜ΕΛղফ (࣮ࡍ ͷܭࢉ͸Energy distanceΛ࢖͍ͬͯΔ) • Marc G. Bellemare, Ivo Danihelka, Will Dabney, Shakir Mohamed, Balaji Lakshminarayanan, Stephan Hoyer, Rémi Munos. abs/1705.10743, 2017 • MMD GAN: Maximum mean discrepancy • ্هͱ͸গ͠ҟͳΔܥྻͷϞσϧ(ࢲࢹ఺)ͰɺDiscriminator͕ग़ྗͨ͠தؒදݱΛΨ΢γΞϯΧʔωϧΛ࢖ͬͯແݶ࣍ݩ ۭؒʹඈ͹͠ɺͦͷ্ۭؒͰͷαϯϓϦϯά఺ͷฏۉಉ࢜Λൺֱ͢Δ • Chun-Liang Li, Wei-Cheng Chang, Yu Cheng, Yiming Yang, Barnabás Póczos. abs/1705.08584, 2017 ࡉ͔ͯ͘͢Έ·ͤΜ
16. ݁Ռ • ݁Ռɺਫ਼౓͸Wasserstein GAN < Cramer GAN < MMD GAN • ͨͩ͠Ϟσϧͦͷ΋ͷͷධՁͱͯ͠͸ඍົͰ͢ɻ • ্ه͸୯ʹ࣮૷ͨ࣌͠ܥྻॱͰɺଞͷ෇ਵ͢Δ࣮ ૷΋৭ʑͱมΘ͍ͬͯΔ • Ұ؏ͨ͠ධՁؔ਺ͷఆ͕ٛ೉͘͠ɺͦ΋ͦ΋ൺֱ ࣗମ͕ײ֮త
17. ߟ࡯: ͳͥMMD GAN͔ • Wasserstein GANͱCramer GAN͸FakeͱRealͷαϯϓϧؒͷൺֱ͢Δ • MMD GAN͸RealͱFakeͦΕͧΕͷϛχόονͷαϯϓϧ఺Λߴ࣍ݩ ۭؒʹࣹӨ͠ɺͦͷॏ৺ͷڑ཭Λ࢖͏ • ࠓճͷϞσϧ(཭ࢄ͔ͭεύʔεͳσʔλ)Ͱ͸ɺݸʑͷαϯϓϧ఺Λ ࢖͏ΑΓ΋ɺ౷ܭ஋ͱͯ͠·ͱΊΒΕͨσʔλ(ߴ࣍ݩۭؒʹ͓͚Δॏ ৺) Λ࢖ͬͨํ͕ޯ഑Λಘ΍͔ͬͨ͢ͷͰ͸ • MMD GAN͸ൺֱతେ͖ͳόοναΠζΛඞཁͱ͢Δ͕ɺࠓճ͸લड़ ͷ௨Γαϯϓϧ͋ͨΓͷσʔλαΠζ͕খ͘͞όοναΠζΛՔ͗΍ ͔ͬͨͨ͢Ίɺར఺Λ׆͔͠΍͔ͬͨ͢
18. Ҏ্ɺϑΝογϣϯΛ Պֶͯ͠Έ·ͨ͠ ઈࢍΤϯδχΞืूதͰ͢ʂ
19. ࢀߟࢿྉɾग़య • Generative Adversarial Networks: Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, Yoshua Bengio. abs/ 1406.2661, 2014 • Wasserstein GAN: Martin Arjovsky, Soumith Chintala, Léon Bottou. abs/1701.07875, 2017 • The Cramer Distance as a Solution to Biased Wasserstein Gradients: Marc G. Bellemare, Ivo Danihelka, Will Dabney, Shakir Mohamed, Balaji Lakshminarayanan, Stephan Hoyer, Rémi Munos. abs/1705.10743, 2017 • MMD GAN: Towards Deeper Understanding of Moment Matching Network: Chun-Liang Li, Wei-Cheng Chang, Yu Cheng, Yiming Yang, Barnabás Póczos. abs/1705.08584, 2017 • ը૾ఏڙ • ΞΠΩϟονϟʔ (https://ai-catcher.com/) • #CBK stocks (https://stock.cubki.jp/)