Skip to content

Latest commit

 

History

History
161 lines (136 loc) · 15.4 KB

02-baseline.md

File metadata and controls

161 lines (136 loc) · 15.4 KB
Raw

Cassava-Leaf-Disease-Classification

Score

Single Model

Train Inference Model Public LB CV Comment
resnext-v1 resnext-inf-v1 resnext50_32x4d 0.894 0.89069 -
efficient-v1 efficient-inf-v1 efficientnet_b4_ns 0.900 0.89103 CutMix, freeze batch normalization
efficient-v2 efficient-inf-v3 efficientnet_b4_ns 0.897 0.88814 gradient accumulation, CosineAnnealingWarmupRestarts
efficient-v3 efficient-inf-v4 efficientnet_b4_ns 0.898 0.89137 MixUp
vt-v1 vt-inf-v1 vit_base_patch16_384 0.897 0.88958 based on efficient-v3
deit-v1 deit-inf-v1 deit_base_patch16_384 0.895 0.89019 based on efficient-v1
efnet-b3-v1 efnet-b3-inf-v1 tf_efficientnet_b3_ns 0.895 0.89255 based on deit-v1, batch update, increase min lr
seres-v1 seres-inf-v1 seresnext50_32x4d 0.900 0.89422 based on efnet-b3-v1, MixUp
seres-v2 seres-inf-v3 seresnext50_32x4d 0.899 0.89532 label smoothing
vt-v2 vt-inf-v2 vit_base_patch16_384 0.899 0.89220 label smoothing, freeze BN, etc

Distillation

Train Inference Model Public LB CV Comment
efnet-dist-v1 efnet-dist-inf-v1 efficientnet_b4_ns 0.898 0.94961 label ensemble-v3, soft label only
efnet-dist-v2 efnet-dist-inf-v1 efficientnet_b4_ns 0.894 0.94848 label ensemble-v3, soft label 0.3
efnet-dist-v3 efnet-dist-inf-v1 efficientnet_b4_ns 0.897 0.94980 label ensemble-v3, soft label 0.9
efnet-dist-v4 efnet-dist-inf-v2 efficientnet_b4_ns 0.892 0.94510 label ensemble-tta-v2 TTA x7, soft label 0.9
deit-dist-v1 deit_base_patch16_384 label ensemble-tta-v2 TTA x7, soft label 0.9

TTA

Train Inference Model Public LB CV Comment
efficient-v1 efficient-inf-v2 efficientnet_b4_ns 0.898 - TTA x10
efficient-v1 efficient-inf-v5 efficientnet_b4_ns 0.901 - TTA x7
seres-v1 seres-inf-v2 seresnext50_32x4d 0.900 - TTA x7 (flip)

Ensemble

Train Inference Model Public LB CV Comment
- ensemble-v2 seresnext50_32x4d 0.903 - seres-v1, seres-v2
- ensemble-v3 efficientnet_b4_ns, seresnext50_32x4d 0.903 - efficient-v1, seres-v1, seres-v2
- ensemble-v4 efficientnet_b4_ns, seresnext50_32x4d 0.902 - efficient-v1, seres-v1
- ensemble-v5 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 0.903 - efficient-v1, seres-v1, seres-v2, vt-v2

Ensemble + TTA

Train Inference Model Public LB CV Comment
- ensemble-tta-v1 efficientnet_b4_ns, seresnext50_32x4d 0.904 - efficient-v1, seres-v1, seres-v2, TTA x9
- ensemble-tta-v2 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 Time out - efficient-v1, seres-v1, seres-v2, vt-v2, TTA x8
- ensemble-tta-v2 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 0.906 - efficient-v1, seres-v1, seres-v2, vt-v2, TTA x7
- ensemble-tta-v2 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 0.906 - efficient-v1, seres-v1, seres-v2, vt-v2, TTA x6
- ensemble-tta-v2 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 0.905 - efficient-v1, seres-v1, seres-v2, vt-v2, TTA x5
- ensemble-tta-v2 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 0.905 - efficient-v1, seres-v1, seres-v2, vt-v2, TTA x4
- ensemble-tta-v3 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 Time out - efficient-v1, seres-v1, vt-v2, TTA x10
- ensemble-tta-v3 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 0.906 - efficient-v1, seres-v1, vt-v2, TTA x9
- ensemble-tta-v3 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 0.904 - efficient-v1, seres-v1, vt-v2, TTA x8
- ensemble-tta-v3 *1 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 0.907 - efficient-v1, seres-v1, vt-v2, TTA x7
- ensemble-tta-v3 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 0.906 - efficient-v1, seres-v1, vt-v2, TTA x6
- ensemble-tta-v4 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 Time out - efficient-v1, seres-v2, vt-v2, TTA x10
- ensemble-tta-v4 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 0.906 - efficient-v1, seres-v2, vt-v2, TTA x9
- ensemble-tta-v4 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 0.906 - efficient-v1, seres-v2, vt-v2, TTA x8
- ensemble-tta-v4 *2 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 0.907 - efficient-v1, seres-v2, vt-v2, TTA x7

Weighted ensemble + TTA

Train Inference Model Public LB CV Comment
- w-ensemble-v1 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 0.907 - based on *1, weight 1 :1 :1.1
- w-ensemble-v1 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 0.906 - based on *1, weight 1 :1 :0.9
- w-ensemble-v2 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 0.907 - based on *2, weight 1 :1 :1.1
- w-ensemble-v2 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 0.907 - based on *2, weight 1 :1 :0.9
- w-ensemble-v1 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 0.907 - based on *1, weight 1 :1.1 :1
- w-ensemble-v1 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 0.907 - based on *1, weight 1 :0.9 :1
- w-ensemble-v1 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 0.907 0.9328321372973384 based on *1, weight 1 :0.8 :1
- w-ensemble-v1 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 0.907 - based on *1, weight 1 :0.7 :1
- w-ensemble-v2 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 0.907 - based on *2, weight 1 :1.2 :1
- w-ensemble-v2 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 0.907 - based on *2, weight 1 :1.1 :1
- w-ensemble-v2 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 0.907 - based on *2, weight 1 :0.9 :1
- w-ensemble-v1 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 0.907 - based on *1, weight 1.1 :1 :1
- w-ensemble-v1 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 0.907 - based on *1, weight 0.9 :1 :1
- w-ensemble-v2 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 0.907 - based on *2, weight 1.1 :1 :1
- w-ensemble-v2 efficientnet_b4_ns, seresnext50_32x4d, vit_base_patch16_384 0.907 - based on *2, weight 0.9 :1 :1
  • w-ensemble-v1
    • efficientnet_b4_ns: <1
    • seresnext50_32x4d: 0.7<<1 -> 0.8
    • vit_base_patch16_384: >1
  • w-ensemble-v2
    • efficientnet_b4_ns: =1
    • seresnext50_32x4d: 1.2>>1 -> 1.1
    • vit_base_patch16_384: =1

Validation

efficient-v1, seres-v1, vt-v2 + TTA x7 + weight 1:0.8:1

  • CV: 0.9328321372973384
  • LB: 0.907

              precision    recall  f1-score   support

           0       0.82      0.79      0.81      1492
           1       0.93      0.89      0.91      3476
           2       0.91      0.87      0.89      3017
           3       0.96      0.98      0.97     15462
           4       0.86      0.85      0.85      2890

    accuracy                           0.93     26337
   macro avg       0.90      0.88      0.89     26337
weighted avg       0.93      0.93      0.93     26337