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
画像ディープラーニングコンペの基本
Search
Jun Koda
August 07, 2025
5
1.7k
画像ディープラーニングコンペの基本
上位Kagglerに学ぶ~画像コンペの戦い方~
2025-08-07
Jun Koda
August 07, 2025
Tweet
Share
More Decks by Jun Koda
See All by Jun Koda
第3回関東kaggler会 🤔 妙だな... (Jun Koda)
junkoda
10
3.2k
Featured
See All Featured
Raft: Consensus for Rubyists
vanstee
140
7.1k
Dealing with People You Can't Stand - Big Design 2015
cassininazir
367
26k
実際に使うSQLの書き方 徹底解説 / pgcon21j-tutorial
soudai
PRO
183
54k
Fantastic passwords and where to find them - at NoRuKo
philnash
51
3.4k
[RailsConf 2023] Rails as a piece of cake
palkan
56
5.8k
ピンチをチャンスに:未来をつくるプロダクトロードマップ #pmconf2020
aki_iinuma
126
53k
Code Review Best Practice
trishagee
70
19k
Java REST API Framework Comparison - PWX 2021
mraible
33
8.8k
Embracing the Ebb and Flow
colly
87
4.8k
Building Adaptive Systems
keathley
43
2.7k
Statistics for Hackers
jakevdp
799
220k
Scaling GitHub
holman
462
140k
Transcript
ը૾σΟʔϓϥʔχϯάίϯϖͷجຊ ্Ґ KagglerʹֶͿɹʙը૾ίϯϖͷઓ͍ํʙ 2025-08-07 Jun Koda ᅳాɹ३ʢίμɹδϡϯʣ https://hakuhodo-technologies.connpass.com/event/361499/
ཧ GMɻը૾ GM Ͱͳ͍ ⇧ ᠳΜͰ࡛ۄͷʮ࡛ۄʹॅΜͰΔͷʹ౦ژͬͯॻౕ͘ʯͰக໋ইΛෛͬͨ AI ʹΑͬͯ Notebook ͕ෆཁʹͳΔલʹ
Notebook GM ʹͳΖ͏͔ͳ ϙευΫͱͯ͠ӉཧΛ͖ͬͯͨ ࠓۀҕୗͰػցֶशݚڀ։ൃ
ඈߦػӢ segmentation. U-Net 70×1000 ͱ͍͏αΠζ͚ͩͲ U-Net ͓ೃછΈͷҩྍը૾εϥΠε ৴߸ॲཧཁૉ͋Γͷը૾ྨ ͡Ίͯͷը૾ίϯϖ ଟνϟϯωϧ͚ͩͲ2Dը૾ྨ
RTX2080Ti Λങͬͯઓ 3D ݂ segmentation ը૾ίϯϖ
G2Net (2021) ͷࢥ͍ग़ ϒϥοΫϗʔϧ߹ମ৴߸͕͋Δ͔Ͳ͏͔Λఆ͢Δ ࣌ܥྻ৴߸Λ spectrogram ͬΆ͘ը૾ʹ͢Δ “࠷େͷϒϨʔΫεϧʔ͕ learning rate
Λ্͛Δͩͬͨ” • Public notebook Λࣸܦ • ৴߸ॲཧΛ͢Δϝμϧݍ͔Β΄ Ͳԕ͍ • ৽͍͠ public notebook ΛΈΔ • lr 1e-5 → 1e-4 ͰείΞര্͕Γ
https://medium.com/@junkoda/kaggle-ॏྗͰۜϝμϧ-1c7135e69817 ࣌ͷϒϩάΑΓ খ͍͞ lr local minimum ʹϋϚΔ G2Net (2021)
ͷࢥ͍ग़
ɾɾɾ ɾɾɾ 0.005 େࠩ ↑ prize ↓ ݍ֎
Hyperparameter tuning Ͱ Kaggle Λউͭ͜ͱͰ͖ͳ͍ ͕ͩෛ͚Δ͜ͱͰ͖Δ
جຊతͳσΟʔϓϥʔχϯά܇࿅ͷ ίίϩॻ͍ͯ͋Δ learning rate warm up, batch size, ͳͲͳͲ খ͍͞
batch size over fi t Λ͙͚ͲɺͦͷͨΊʹখ͘͢͞Δͷअಓͳؾ͕͢Δ
https://www.kaggle.com/competitions/hms-harmful-brain-activity-classi fi cation/discussion/488083 େGrandmater ҙ֎ͱࡉ͔͘ௐͯ͠Δ
Grid search ʹ͢Δʁ Optunaʁ ͦΕͱ G • S • Dʁ
ͱ͜ΖͰ Hyperparameter optimization Ͳ͏ͯ͠·͔͢ʁ
Kaggle Ͱ graduate student ࢲͨͪࣗɻࢲखͰͪ·ͪ·ௐͯ͠Δ
Augmentation େࣄ
Albumentatations RandomRain: https://explore.albumentations.ai/transform/RandomRain ˚ ܇࿅σʔλʹͳ͍ʹରԠͰ͖ΔΑ͏ʹͳΔʢe.g. Ӎͷࣸਅʣ ˕ σʔλΛ૿ͯ͘͠܇࿅Ͱ͖ΔΑ͏ʹ͢Δ Augmentation ͷޮՌ
Augmentation ͍͢͝ͷਤ ඈߦػӢίϯϖ ࠷ॳͷ 10 epoch ͚ͩͩͱҧ͍গͳ͍ ͘܇࿅Ͱ͖Δͷ͕ϙΠϯτ
b d ͜Ε b HFlip ͜Ε d Augmentation Λ͍͚ͯ͠ͳ͍߹ ҰจࣈΞϧϑΝϕοτྨͰ
Horizontal fl ip Ϟσϧࠔ
ΠϯυਓΛӈʹ ϋϯυϧΛࠨʹ ߹ੑͷऔΕΔΑ͏ʹస - atmacup ंͷيಓ༧ଌ hakubisin ͞Μ 1st place
solution https://speakerdeck.com/hakubishin3/turing-x-atmacup-number-18-1st-place-solution ը૾ɾϋϯυϧ֯ɾΟϯΧʔͳͲΛҰ؏ͯ͠స͢Δ ※ ͜ΕͰं͕ӈଆΛΓͩͨ͠Βେมͳ߹μϝ
ରশੑΛճ෮ͤ͞Δ - ճసɾస͕͏·͍͔͘ͳͯ͘मਖ਼Ͱ͖Δ͜ͱ https://www.kaggle.com/competitions/waveform-inversion/writeups/ruby-14th-place-solution ίϯϖɾసͤ͞Δͱݯ͕̍ͭζϨΔ ͳΒɺసͤͯ͞1 pixel ͣΒ͍͍ͤ͡Όͳ͍ tascj ͞ΜͷඈߦػӢ
solution 0.5 pixel ͣΒ͢ ڥք͕ؾʹͳΔ͚ͲେৎΒ͍͠ ← సͯͣ͠Β͢ ճసɾస͕Ͱ͖ͳͯ͘ͻͱͰͰ͖Δ߹
σʔλ͕গͳ͍ͱ͖ Augmentation Λڧ͘ɺϞσϧখ͞Ί Theo Viel https://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/writeups/on-strike-2nd-place-solution Cut Mix Yun et
al (2019) https://arxiv.org/pdf/1905.04899 ϥϯμϜʹը૾ΛࠞͥΔ ϥϕϧ໘ੵൺ soft label ͦΜͳཚͳ! RSNA 2023 ೣ͕4ʹݘ͕6ʂ
Augmentation ڧ͍ Ճσʔλͬͱڧ͍ G2Net (2021) ѹత1Ґ https://www.kaggle.com/competitions/g2net-gravitational-wave-detection/writeups/kdl-top-1-solution-deep-learning-part (Geophysical Waveform
Inversion) https://www.kaggle.com/competitions/waveform-inversion/leaderboard 13TB Ҏ্ͷσʔλΛੜ On-the- fl y σʔλੜͷͨΊʹ CUDA Λॻ͍ͨ Augmentation ٖࣅతͳՃσʔλͳͷ͔ͩΒͦΕͦ͏ ࢲ: 660 GB ͷσʔλ͕େ͖ͯ͘ਏ͍ → ශऑ!
ը૾ίϯϖಓͳվળ͕ඞཁ • େ͖ͳΞΠσΟΞͰυϯͱείΞ͕Α͘ͳΔͷͰͳ͍ • ಉ͡Α͏ʹݟ͑ͯখ͞ͳվળͷੵΈॏͶ͕େ͖ͳࠩʹ զʑ RSNA ίϯϖϨδΣϯυͷΑ͏ʹ͍͔ͳ͍ ಓʹίπίπࢼߦࡨޡɾܦݧΛੵΉʁ ࡉ͔͍ςΫχοΫΛ
prize solution ίʔυΛಡΜͰձಘ͠Α͏ʢࢲͰ͖ͯͳ͍ʣ
ʮʓʓ͚ͨ͠Ͳޮ͔ͳ͔ͬͨʯҙຯ͕ͳ͍ • ͕݅ͦΖͬͯॳΊͯޮՌΛൃش͢Δ͜ͱ͕Α͋͘Δ • ྫ͑ը૾Λେ͖ͨ͘͠ͱ͖ɺϞσϧΛେ͖ͨ͘͠ͱ͖ɺ͘܇࿅ͨ͠ͱ͖ʹॳΊͯ ޮՌΛൃش͢Δ • ؆୯ͳྫͩͱ augmentation ͨ͠Β
epoch Λ૿͢ • ͦͷ··ͩͱɺ୯ʹσʔλ͕ྼԽ͚ͨͩ͠ tattaka ͞ΜʮҰࣺͯͨΞΠσΟΞʹҙࣝతʹཱͪฦΖ͏ͱߟ͑ͯΔʯ ؔ౦ Kaggler ձ 2025 य़ private communication ޙ͔Βߟ͑Δͱਖ਼ղͷۙ͘ΛԿ௨Γա͗ͯΔɺͱ͍͏͜ͱԿ͋Δ
Kenshin ͞Μ https://tech.preferred.jp/ja/blog/kaggle-contrails-3rd-place/ https://blog.knshnb.com/posts/journey-to-grandmaster/ ඈߦػӢίϯϖ ࢲΛؚΊͨଟ͘ͷਓʮ2.5D Ϟσϧ͏·͍͔͘ͳ͍ʯ ৴೦ͱࢼߦࡨޡʂ
G2Net Ͱҹʹͬͨ͜ͱ έϩοϐઌੜʮը૾Ͱ 0.88 ͑ΒΕͳ͍ɺ৴߸Λݟͳ͍ͱʯ ࣮ࡍɺ্Ґ solution ͦͷ௨Γͩͬͨ GMʹͳΔͱඍௐͰ͑ΒΕͳ͍ݶք͕ѲͰ͖ΔΒ͍͠ ͳͷͰ
Hyperparameter tuning ࡉ͔͍ௐ͖ͬͱྑ͍ܦݧ ͍Ζ͍Ζࢼߦࡨޡͯ͠ɺͲͷ͘Β͍·ͰվળͰ͖Δ͔Θ͔ΕҰਓલ ʢࢲΘ͔Βͳ͍ʣ
• ແ৬Λଓ͚Δ͓͕ۚ͏ͳ͍ • Grandmaster ʹͳΓ͍ͨ • ͓͕ۚͳ͍ͷͰϥετνϟϯε • Cloud GPU
ΛͬͯͰۚϝμϧ • ࠷ޙͷ1िؒʹ A100 Λ 4~5 • େ͖ͳϞσϧ → ࣦഊ • ը૾Λେ͖͘ • 256 pixel ͷೖྗը૾Λ 1024 ʹ Ryushi ͞Μ͔Βͷ͓ ্Ґೖ࣌ʹߟ͑ͨ͜ͱ ্Ґ..ೖ...? ͏͍͜ͱۚϝμϧऔͬͯͳ͍ͷͰΕ·ͨ͠ ը૾ΛͰ͔ͬ͘ 1024×1024 ඈߦػӢίϯϖ Google Research - Identify Contrails to Reduce Global Warming ৴߸͕ࡉ͍͔Βޮ͍ͨͷ͔ɻG2Net Ͱޮ͍ͨ
ίϯϐϡʔλࣄ ϩʔΧϧϚγϯ Ubuntu RTX 4090 (RAM 24 GB) Intel Core
i5 13500 (TDP 65W) ਫྫྷΛආ͚ۭͯྫྷʹ͢ΔͨΊ߇͑Ίʹ͚ͨ͠ͲɺίΞ͕ͨ͘͞Μ͋Δͱ͍͑ҙ֎ͱCPU͏ M.2 SSD WD Black 2TB (Gen 4 ~7000MB/s) σʔλಡΈࠐΈҙ֎ͱେࣄ ۚϝμϧ͕औΕͦ͏ͳΒ࠷ޙͷ1िؒʹ Cloud GPU Fractal Design "North"
औΕ·ͤΜͰͨ͠!!! Yale/UNC-CH - Geophysical Waveform Inversion
ۚϝμϧ͕औΕͦ͏ͳΒ࠷ޙͷ1िؒʹ Cloud GPU ʮۚϝμϧΛऔΔʯͳͲͱࢤ͕͍͔ΒۜϝμϧʹͳΔͷͰ͢ Cloud GPU ʹ͓ۚΛ͗ࠐΉͳΒ1ҐΛऔΓͳ͍͞
1Ґ͕औΕͳ͍ͳΒ Cloud GPU ʹ͓ۚΛ͗ࠐΉͷΊͳ͍͞ ରۮ ͦΜͳ͜ͱΑΓ prize solution Λࣸܦ͠Α͏ ϓϥΠεϨε
👍
ѱຐͷᅤ͖ ۚͰޙ͔Βิరͱ͔͍͏ ѱຐͷ༠ʹෛ͚ͳ͍ ๏Χ...ʢͦΕҎ্͍͚ͳ͍ʂ
30ສԁΛਓੜʹࢿʁ RTX 5070 (12G), 5070Ti (16G) 10ສԁલޙͷ GPU ͔Βελʔτ͢ΔͷΞϦͳؾ͕͢Δ
ࢲͷ͓͢͢ΊϩʔυόΠΫ ݈߁େࣄ ӡಈʹޮ͘ 10͑Δ 30ສԁΛਓੜʹࢿʁ RTX 5070 (12G), 5070Ti (16G)
10ສԁલޙͷ GPU ͔Βελʔτ͢ΔͷΞϦͳؾ͕͢Δ Shimano 105 ΛͬͯΔͭ