riron1206 / self_supervised_learning_test

Self-Supervised Learning試す

データはkaggleのキャッサバコンペ

https://www.kaggle.com/c/cassava-leaf-disease-classification

SwAV

https://arxiv.org/abs/2006.09882

埋め込みベクトルにクラスタリングを適応したSelf-Supervised Learning

Backborn Model

• ResNet50

Self-Supervised Training

• https://www.kaggle.com/anonamename/swav-tf

Supervised Training

• 実行環境: windows10, poetry (tensorflow 2.3.1 etc: pyproject.toml)

• imagenetの重みで学習: no-swav-resnet50-v2_local.ipynb

• SwAVの重みで学習: swav-tf-linear-evaluation_local.ipynb

  
  

BYOL

https://arxiv.org/abs/2006.07733

Negative sampleが不要なSelf-Supervised Learning

Backborn Model

• SeResNeXT50_32x4d

Self-Supervised Training

• 実行環境: windows10, anaconda (pytorch- lightning 1.1.0 etc: lightning.yml)

• ①画像サイズ224*224: byol-pytorch_seresnext50_32x4d.ipynb

• ➁画像サイズ512*512: byol-pytorch_seresnext50_32x4d_512.ipynb

Supervised Training

• 実行環境: windows10, anaconda (lightning.yml)
• ①のモデル+画像サイズ224*224: byol-pytorch_seresnext50_32x4d_Supervised_Training_224.ipynb
• ①のモデル+画像サイズ224*224 + Data Augmentationあり: byol-pytorch_seresnext50_32x4d_Supervised_Training_224_aug.ipynb
• ①のモデル+画像サイズ512*512: byol-pytorch_seresnext50_32x4d_Supervised_Training_512.ipynb
• ➁のモデル+画像サイズ512*512: Supervised_Training_512_aug_byol-pytorch_seresnext50_32x4d_512.ipynb