nku-shengzheliu / Cassava-Leaf-Disease-Classification

This repository is used to record our code in "Cassava Leaf Disease Classification" kaggle competition.

Cassava-Leaf-Disease-Classification-NKU_428

This repository is used to record our code in this competition

Dataset analysis

• 图像数量(Number of images)：21397
• 分辨率(image size)：800*600
• 类别分布(class distribution):

Results

Loss	Trick	Code	Acc in val	Acc in test
Cross Entropy loss	warmup	-	0.867	-
Cross Entropy loss	warmup, imgaug	-	-	0.872
Focal loss	warmup, imgaug	-	-	0.882
Cross Entropy loss	label smooth, warmup, imgaug	code	-	0.891
Bi tempered logistic loss	label smooth, warmup, imgaug	code	0.8851	0.890
Bi tempered logistic loss	label smooth, warmup, imgaug, freeze bn	code	0.8917	0.890
-	-	-	-	-

2021.1.13 Methods for future improvement 可以改进的方向

• Bi-tempered-logistic-loss (Description found here).
• Frozen BN.
• Training with early stopping criterion.
• N-Fold CV

2021.1.14 Methods for future improvement 可以改进的方向

• 将backbone中的maxpool改为softpool
• use SAM Optimizer

2021.1.15 Methods for future improvement 可以改进的方向

• use label smooth and cutmix may get score of 0.903
• use 5-fold CV

2021.1.16 Methods for future improvement 可以改进的方向

• CVPR 2019 Cassava Disease Classification挑战赛: here