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

SIIM-ISIC Melanoma Classification 2020

Inoichan
September 01, 2020
Technology
0
92

SIIM-ISIC Melanoma Classification 2020

2020 SIIM-ISIC Melanoma Classification
Competition link: https://www.kaggle.com/c/siim-isic-melanoma-classification

Inoichan

September 01, 2020
Tweet

More Decks by Inoichan

See All by Inoichan
[CV勉強会@関東 CVPR2023] UniAD: Planning-oriented Autonomous Driving
inoichan
2
960
[CV関東3D勉強会] TPVFormer ~マルチカメラを用いた自動運転の3D Occupancy Prediction~
inoichan
0
740
Teslaはカメラを使ってどのように世界を認識しているか
inoichan
5
8k
[CVPR2022読み会] Time3D: End-to-End Joint Monocular 3D Object Detection and Tracking for Autonomous Driving
inoichan
0
680
[第4回 全国医療AIコンテスト] Kaggleから学んだ医療画像データ解析の取り組み方
inoichan
5
3.5k
Use of Data Science in the field of Life Science
inoichan
1
410
Lyft Motion Prediction for Autonomous Vehicles 2020
inoichan
0
55
RSNA STR Pulmonary Embolism Detection Solution Overview
inoichan
0
81
論文紹介「ResNeSt: Split-Attention Networks」
inoichan
0
940

Other Decks in Technology

See All in Technology
LOWYAビジネスダッシュボード：DevとBizの共通認識が切り拓く未来
kazumax55
0
110
UIコンポーネントライブラリをうまく使うためにできること / components-with-designer
mottox2
3
1.7k
クラウドだからできた 地方主導のJAWS DevOps
matsuihidetoshi
2
170
リスキリングの特効薬！「技術コミュニティ」のススメ
mamohacy
1
650
Jetpack Glanceではじめる Material 3のColor
mochico
0
570
Setting up Monitoring for Kubernetes
prathamesh
0
160
asken リアーキテクチャを行う背景と DDD導入戦略
murahigeas
1
140
自作WASMランタイムをKernelに改造する試み.pdf
riru
2
470
kanto_kaggler_senkin13
senkin13
0
350
IoTだからこそ、サーバーレスを活用すべき３つの理由
soracom
PRO
0
110
様々なユースケースに利用できる "パスキー" の 導入事例の紹介とUXの課題解説 @ DroidKaigi 2023
ritou
1
1.7k
変化する組織の中で運用するフロントエンドエコシステム / Frontend Ecosystem in Organizational Changes
i524
1
360

Featured

See All Featured
Agile that works and the tools we love
rasmusluckow
323
20k
Rebuilding a faster, lazier Slack
samanthasiow
71
7.8k
Dealing with People You Can't Stand - Big Design 2015
cassininazir
352
21k
StorybookのUI Testing Handbookを読んだ
zakiyama
9
3.9k
BBQ
matthewcrist
76
8.4k
Rails Girls Zürich Keynote
gr2m
90
12k
Stop Working from a Prison Cell
hatefulcrawdad
265
18k
"I'm Feeling Lucky" - Building Great Search Experiences for Today's Users (#IAC19)
danielanewman
219
21k
Fontdeck: Realign not Redesign
paulrobertlloyd
74
4.6k
Designing with Data
zakiwarfel
93
4.5k
RailsConf & Balkan Ruby 2019: The Past, Present, and Future of Rails at GitHub
eileencodes
123
30k
Practical Orchestrator
shlominoach
179
9.3k

Transcript

  1. SIIM-ISIC Melanoma Classification
    いのうえ

    View Slide

  2. Overview
    → →

    View Slide

  3. Last year’s solution (Link)
    ↓

    View Slide

  4. My approach summary
    Region model
    EfficientNetB3
    TPU models
    EfficientNets
    TPU models
    EfficientNets
    TPU models
    EfficientNets
    TPU models
    EfficientNets
    TPU models
    EfficientNets
    TPU models
    EfficientNets
    TPU models
    EfficientNets
    TPU models
    Effnets
    Crop-model
    RegNet16, 80
    Crop model
    RegNet16, 80
    Patch-Attention-model
    RegNets
    Patch-Attention-model
    RegNets
    Patch-Attention-model
    RegNets
    Patch Attention model
    Effnets, RegNets
    Ridge Stacking
    TPU models
    EfficientNets
    TPU models
    EfficientNets
    TPU models
    EfficientNets
    TPU models
    EfficientNets
    TPU models
    Effnets
    Ridge Stacking

    View Slide

  5. Interesting Discussion (Link)
    ↓

    View Slide

  6. Region model
    ↓

    View Slide

  7. Patch-Based Attention model (link)

    View Slide

  8. Postprocessing?

    View Slide

  9. Postprocessing?
    以上の予測値を 倍
    →

    View Slide

  10. Key discussion (link)

    View Slide

  11. 1st place (link)
    2018+2019+2020's data or 2020 only
    →

    View Slide

  12. 2nd place (link)
    →
    EfficientNet-B6, 512x512, BS64, no metadata (CV 0.9336 / public 0.9534)
    EfficientNet-B7, 640x640, BS32, gradient accumulation 2, no metadata (CV 0.9389 / public 0.9525)
    Model 1, trained on combined training and pseudolabeled test data (CV 0.9438 / public 0.9493)

    View Slide

  13. 2nd place (link)
    →
    EfficientNet-B6, 512x512, BS64, no metadata (CV 0.9336 / public 0.9534)
    EfficientNet-B7, 640x640, BS32, gradient accumulation 2, no metadata (CV 0.9389 / public 0.9525)
    Model 1, trained on combined training and pseudolabeled test data (CV 0.9438 / public 0.9493)

    View Slide

  14. Others

    View Slide