| Train | Inference | Model | Public LB | CV | Comment |
|---|---|---|---|---|---|
| v1-train | v1-inf | resnext50_32x4d | 0.894 | 0.88568 | image size: 256 -> 384, batch size: 32 -> 16 |
| v2 | v2 | resnext50_32x4d | 0.890 | 0.88674 | add 2019 dataset, batch size: 16 -> 24 |
| v3 | v3 | resnext50_32x4d | 0.891 | 0.88772 | image size: 384 -> 512, batch size: 24 -> 14 |
| same above | v3-TTA | resnext50_32x4d | 0.889 | same above | TTA x3 mean() |
| same above | v3-TTAx7 | resnext50_32x4d | 0.891 | same above | TTA x7 mean() |
| same above | v3-TTAx10 | resnext50_32x4d | 0.893 | same above | TTA x10 mean() |
| same above | v3-TTAx15 | resnext50_32x4d | 0.894 | same above | TTA x15 mean() |
| v4-train | v4-inf | resnext50_32x4d | 0.895 | 0.88822 | add augmentation functions, no TTA |
| same above | v4-TTAx3 | resnext50_32x4d | 0.893 | same above | TTA x3 mean() |
| same above | v4-TTAx10 | resnext50_32x4d | 0.895 | same above | TTA x10 mean() |
| same above | v4-TTAx10-simple | resnext50_32x4d | 0.894 | same above | TTA x10 mean() simplify |
| b4-v1-train | b4-v1-inf-noTTA | tf_efficientnet_b4_ns | 0.893 | 0.88879 | Use EfficientNet B4 Noisy Student based on v4-train |
| v4-train b4-v1-train | ensemble-v1 | resnext50_32x4d, tf_efficientnet_b4_ns | 0.895 | - | Ensemble ResNext50, EfficientNet B4 |
| same above | ensemble-v2 | resnext50_32x4d, tf_efficientnet_b4_ns | 0.892 | - | Ensemble ResNext50, EfficientNet B4, TTA x3 |
| same above | ensemble-v3 | resnext50_32x4d, tf_efficientnet_b4_ns | 0.894 | - | Ensemble ResNext50, EfficientNet B4, TTA x8 |
| v5-train | v5-inf | resnext50_32x4d | 0.882 | 0.89365 | Oversampling 0, 1, 2 and 4 |