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
ディープラーニングでコーデを提案/FashionTechMeetup#4
Search
tn1031
June 07, 2017
Technology
0
2.4k
ディープラーニングでコーデを提案/FashionTechMeetup#4
tn1031
June 07, 2017
Tweet
Share
More Decks by tn1031
See All by tn1031
Outfit Generation and Style Extraction via Bidirectional LSTM and Autoencoder
tn1031
0
130
インタラクティブな属性操作が可能なファッションアイテム検索/attribute manipulation survey
tn1031
0
1.2k
Autoencoderを用いたOutfitからのスタイル抽出/style auto encoder
tn1031
0
13k
fashion_workshop_survey/Size Recommendation System for Fashion E-commerce
tn1031
0
290
画像を用いたファッションアイテム検索/Image Retrieval for Fashion
tn1031
0
5.5k
ファッションアイテム検索における深層学習の活用/Fashion Item Retrieval using Deep Learning
tn1031
0
2.3k
KDD 2016勉強会/Images Don’t Lie: Transferring Deep Visual Semantic Features to Large-Scale Multimodal Learning to Rank
tn1031
0
1k
ファッションのコーディネートを自動生成してみた/FashionTech Talks Tokyo #1 LT
tn1031
2
1.2k
Fashion Tech x Machine Learning/twm_fashion_ml
tn1031
5
5.7k
Other Decks in Technology
See All in Technology
Introduction to Sansan for Engineers / エンジニア向け会社紹介
sansan33
PRO
5
43k
Performance Insights 廃止から Database Insights 利用へ/transition-from-performance-insights-to-database-insights
emiki
0
320
組織改革から開発効率向上まで! - 成功事例から見えたAI活用のポイント - / 20251016 Tetsuharu Kokaki
shift_evolve
PRO
1
150
20251014_Pythonを実務で徹底的に使いこなした話
ippei0923
0
210
研究開発部メンバーの働き⽅ / Sansan R&D Profile
sansan33
PRO
3
20k
Digitization部 紹介資料
sansan33
PRO
1
5.6k
Claude Code Subagents 再入門 ~cc-sddの実装で学んだこと~
gotalab555
10
16k
Liquid AI Hackathon Tokyo プレゼン資料
aratako
0
110
ソースを読むプロセスの例
sat
PRO
15
9.3k
Zephyr(RTOS)にEdge AIを組み込んでみた話
iotengineer22
0
190
20251007: What happens when multi-agent systems become larger? (CyberAgent, Inc)
ornew
1
460
今この時代に技術とどう向き合うべきか
gree_tech
PRO
2
2.1k
Featured
See All Featured
Done Done
chrislema
185
16k
Design and Strategy: How to Deal with People Who Don’t "Get" Design
morganepeng
132
19k
"I'm Feeling Lucky" - Building Great Search Experiences for Today's Users (#IAC19)
danielanewman
230
22k
How to train your dragon (web standard)
notwaldorf
97
6.3k
Art, The Web, and Tiny UX
lynnandtonic
303
21k
Templates, Plugins, & Blocks: Oh My! Creating the theme that thinks of everything
marktimemedia
31
2.5k
Product Roadmaps are Hard
iamctodd
PRO
54
11k
Connecting the Dots Between Site Speed, User Experience & Your Business [WebExpo 2025]
tammyeverts
10
600
ReactJS: Keep Simple. Everything can be a component!
pedronauck
667
120k
Agile that works and the tools we love
rasmusluckow
331
21k
I Don’t Have Time: Getting Over the Fear to Launch Your Podcast
jcasabona
33
2.5k
jQuery: Nuts, Bolts and Bling
dougneiner
65
7.9k
Transcript
σΟʔϓϥʔχϯάͰίʔσΛఏҊ !UO7"4*-: JOD 'BTIJPO5FDI.FFUVQ
ίʔσΛఏҊ͢Δ w ϑΝογϣϯͰѻ͏ը૾ w Γ͍ͨ͜ͱ ࣮ݱํ๏ w $//Λ༻͍ͨಛநग़ w εφοϓը૾ͷදݱ
w ݕࡧͷΈ ࣮ݧ ·ͱΊ "HFOEB
w தଜຏ!UO w σʔλαΠΤϯςΟετ w લ৬ɿ4* w ࣄɿػցֶशɾը૾ೝࣝ w ΞϧΰϦζϜͷ։ൃ
ࣗݾհ !UO
3&4&"3$)ˍ%&7&-01.&/5 ࣗࣾͰഓͬͨ։ൃٕज़ͷఏڙ ɾΞϓϦ։ൃ ɾΫϩʔϦϯά ɾػցֶश ɾσΟʔϓϥʔχϯά ɾը૾ղੳ
։ൃٕͨ͠ज़ Ϟσϧண༻ը૾ εφοϓը૾ ΛΫΤϦͱͯ͠ը૾Λݕࡧ͢Δ ΫΤϦը૾ ݕग़ ݕࡧ
ίʔσΛఏҊ͢Δ w ϑΝογϣϯͰѻ͏ը૾ w Γ͍ͨ͜ͱ ࣮ݱํ๏ w $//Λ༻͍ͨಛநग़ w εφοϓը૾ͷදݱ
w ݕࡧͷΈ ࣮ݧ ·ͱΊ "HFOEB
ϑΝογϣϯͰѻ͏ը૾ ը૾ͱεφοϓը૾ͱ͍͏छྨͷυϝΠϯ͕ଘࡏ͢Δ w ஔ͖ࡱΓϚωΩϯͷը૾͕ଟ͍ w ਖ਼໘͔ΒΈͨ࣌ͷσβΠϯ͕Θ͔Γ͍͢ w ண༻Πϝʔδ͕༙͖ʹ͍͘
ը ૾ ε φ ο ϓ ը ૾ ը૾ ಛ w Ϟσϧ͕ண༻ͨ͠ը૾ w ண༻࣌ͷҹίʔσΟωʔτͷࢀߟʹͳΔ w ϙʔζഎܠʹΛڽΒ͍ͯ͠Δ
Γ͍ͨ͜ͱ ͷண༻ΠϝʔδΛఏڙ͍ͨ͠ w ண༻Πϝʔδͷఏڙར༻γʔϯͷى w ίʔσΟωʔτͷఏҊ ৄࡉΛΈ͍ͯΔϢʔβʔʹͱࣅ͍ͯΔΞΠςϜΛ ͬͨεφοϓը૾Λදࣔ͢Δ
ίʔσΛఏҊ͢Δ w ϑΝογϣϯͰѻ͏ը૾ w Γ͍ͨ͜ͱ ࣮ݱํ๏ w $//Λ༻͍ͨಛநग़ w εφοϓը૾ͷදݱ
w ݕࡧͷΈ ࣮ݧ ·ͱΊ "HFOEB
$//Λ༻͍ͨಛநग़ ΞΠςϜͷಛΛ$//ͰϕΫτϧԽ͢Δ w ը૾͔ΒΞΠςϜͷಛΛநग़ͯ͠ϕΫτϧԽ͢Δ w ϕΫτϧಉ࢜ͷҐஔؔྨࣅͱΈͳ͢͜ͱ͕Ͱ͖Δ ಛྔۭؒ f(x) ͍ۙ(ࣅ͍ͯΔ)
ԕ͍(ࣅ͍ͯͳ͍) ը૾σʔλ ॎԣ480pixelͷ߹ɺ࣍ݩ 480x480x3 = 691200 dim ը૾ಛྔ ը૾σʔλΛදݱ͢Δ࣍ͷϕΫτϧ ѹॖ ؔʹCNNΛ࠾༻
$//Λ༻͍ͨಛநग़ ͷྨࣅͷؔUSJQMFUMPTTͰධՁ͢Δ Anchor Positive Negative CNN CNN CNN w
ը૾ͷυϝΠϯʹؔͳ͘ڞ௨ͷωοτϫʔΫΛ͏ w ࣅ͍ͯΔը૾ಉ࢜ۙͮ͘Α͏ʹʗࣅ͍ͯͳ͍ը૾ಉ࢜ԕ͔͟ΔΑ͏ʹ Embedding margin ֶश Embedding
εφοϓը૾ͷදݱ εφοϓը૾ʹؚ·ΕΔΞΠςϜͷಛྔΛΧςΰϦຖʹܭࢉ͢Δ w εφοϓը૾͔Β֤ΧςΰϦͷΞΠςϜΛݕग़ͯ͠ύʔε w ͦΕͧΕͷಛྔΛ·ͱΊͯεφοϓը૾ͷදݱͱ͢Δ UPQTUPQTྖҬ͔Βநग़ͨ͠ಛྔ QBOUTQBOUTྖҬ͔Βநग़ͨ͠ಛྔ CBHTCBHTྖҬ͔Βநग़ͨ͠ಛྔ
GPPUXFBSGPPUXFBSྖҬ͔Βநग़ͨ͠ಛྔ \ εφοϓը૾ ݕग़ εφοϓը૾ͷදݱ
ݕࡧͷΈ छྨͷϞδϡʔϧ͔ΒͳΔ w ݕग़Ϟδϡʔϧεφοϓը૾Λύʔεͯ͠ಛྔΛܭࢉ͓ͯ͘͠ w ݕࡧϞδϡʔϧը૾͔ΒύʔεࡁΈεφοϓը૾Λݕࡧ͢Δ ݕग़ ݕࡧ εφοϓը૾
ΞΠςϜྖҬ ݕग़ Ϟδϡʔϧ ݕࡧ Ϟδϡʔϧ IUUQBSYJWPSHBCT
ίʔσΛఏҊ͢Δ w ϑΝογϣϯͰѻ͏ը૾ w Γ͍ͨ͜ͱ ࣮ݱํ๏ w $//Λ༻͍ͨಛநग़ w εφοϓը૾ͷදݱ
w ݕࡧͷΈ ࣮ݧ ·ͱΊ "HFOEB
࣮ݧ݁Ռ1/2 ΫΤϦը૾ ݕࡧ݁Ռ
࣮ݧ݁Ռ2/2 ΫΤϦը૾ ݕࡧ݁Ռ
ίʔσΛఏҊ͢Δ w ϑΝογϣϯͰѻ͏ը૾ w Γ͍ͨ͜ͱ ࣮ݱํ๏ w $//Λ༻͍ͨಛநग़ w εφοϓը૾ͷදݱ
w ݕࡧͷΈ ࣮ݧ ·ͱΊ "HFOEB
w ը૾͔Βεφοϓը૾Λݕࡧ͢ΔΈΛఏҊ w ண༻Πϝʔδͷىʗར༻γʔϯͷى w ίʔσΟωʔτͷఏҊ w σΟʔϓϥʔχϯάΛ༻͍࣮ͯݱ w ݕग़ͱݕࡧͷΈ߹Θͤ
w ྨࣅͷධՁʹUSJQMFUMPTTΛ༻͍Δ w ༻ײ w େࡶͳಛଊ͑ΒΕ͍ͯΔ w ৎײͷΑ͏ʹࡉ͔͍ಛνϡʔχϯάޙॲཧͰٵऩ w ݕग़ͷਫ਼͕ݕࡧਫ਼ʹӨڹ͢Δ ·ͱΊ
͝ਗ਼ௌ͋Γ͕ͱ͏͍͟͝·ͨ͠