search

TakeruEndo / kaggle_Cassava

0 stars 1 forks source link

6. ノイジーラベルに対する対策 #6

Open TakeruEndo opened 3 years ago

TakeruEndo commented 3 years ago

discussion: \ https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/209782

code: \ https://github.com/CoinCheung/pytorch-loss/blob/master/pytorch_loss/taylor_softmax.py

TakeruEndo commented 3 years ago

@arutema47 のノイズ除去 \ https://github.com/kentaroy47/Kaggle-PANDA-1st-place-solution

TakeruEndo commented 3 years ago

Label smoothing is good in public lb. Bi-Tempered Logistic Loss and Focal Cosine Loss can be a good alternative.

TTA Choosing commonly used TTA, public lb score may get worse. This is heavily influenced by the CV Strategy. A little rotation or simple augmentation can help. However, it is sensitive to the number of TTAs. (In fact, you can reach the current gold medal area with no TTA.)

https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/207450

TakeruEndo commented 3 years ago

label smootingが重要 \ https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/209065