Kaggle - Carvana Image Masking Challenge1の6位解法2の調査記事です.
Title: Overview of JbestDeepGooseFlops solution (3rd public)
- Name: n01z3
- Kaggle Discussion: https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/40119#latest-225659
開発環境
- GPU x 40個
Models
- 全てのモデルに対してCV=5でクロスバリエーションを実施
フォーラム2より引用
Valeriy Babushkin (VENHEADs)の試み
- U-Netを様々なOptimizer(Adam, RMSProp, SGD)で学習
Artur Kuzin (n01z3)の試み
- mxnetを利用(Deconvolution部分の不具合のため,LinkNet3を構築できず)
Evgeny Nizhibitsky (nizhib)の試み
- pytorchで作った最良のモデルを作成.
- incnet: inception-v3 + linknet decoders(Residual Blockを組み込んだU-Net)
- dinknet: densenet + linknet decoders
- Adam, learning rate = 0.0001
Roman Trusov (lextal)の試み
- pytorchでpspnet4を3つの中間層を一つのDeconv層に修正したモデルを構築.
- ネットワークの重みはResNetの18層と34層で初期化
[参考]Steven Nguyen(28位)の試み
- DeepUNet5を利用.
References
-
Kaggle, Carvana Image Masking Challenge, 2017. ↩
-
n01z3, Overview of JbestDeepGooseFlops solution (3rd public), 2017. ↩
-
Abhishek Chaurasia et al., LinkNet: Exploiting Encoder Representations for Efficient Semantic Segmentation, 2017. ↩
-
Yangqing Jia, et al., Caffe: Convolutional Architecture for Fast Feature Embedding, 2014. ↩
-
Ruirui Li, et al., DeepUNet: A Deep Fully Convolutional Network for Pixel-level Sea-Land Segmentation, 2017. ↩