| Classifer | Preprocessing | Best Acc | Macro F1 score | Parameters |
|---|---|---|---|---|
| SVC | Normalization&Standardization | 0.8999 | 0.8998 | ~28.8M |
| 2-layer CNN with Dropout | RandomHorizontalFlip&VerticalFlip,Normalization&Standardization | 0.9335 | 0.9333 | ~824K |
| ResNet-7 | RandomHorizontalFlip&VerticalFlip,Normalization&Standardization,Random Affine,Random Crop | 0.9300 | \ (Deleted before Cal) | ~184K |
| WRN-28-2 | RandomHorizontalFlip&VerticalFlip,Normalization&Standardization,Random Affine,Random Crop | 0.9564 | 0.9564 | ~1.42M |
| WRN-40-4 | RandomHorizontalFlip&VerticalFlip,Normalization&Standardization,Random Affine,Random Crop | 0.9578 | 0.9577 | ~8.53M |
| SE-WRN-40-4 [BEST] | RandomHorizontalFlip,Normalization&Standardization,Random Affine,Random Crop,RandomErasing,RandomApply-ColorJitter | 0.9609 | 0.9609 | ~8.97M |
| SE-WRN-64-8 | RandomHorizontalFlip,Normalization&Standardization,Random Affine,Random Crop,RandomErasing,RandomApply-ColorJitter | 0.9595 | 0.9595 | ~60.9M |
以下是2-layer CNN with Dropout 和 SE-WRN-40-4 (BEST) 的训练过程中 loss-epoch 与acc-epoch的可视化图
以下是SE-WRN-40-4的训练过程中 学习率变化 的可视化图
python run_best.py
即可跑最好的CNN与RNN模型
github上的版本所有模型参数都使用LFS （这背后是一个教训，本仓库为重置版）