Skip to content

(04.29, dongwoo) Unet test, DLV3 resnext50 with aug

Jump to bottom
ddooom edited this page Apr 29, 2021 · 17 revisions

Unet test


|모델|배치|시드|time/epoch|epoch|loss|train_loss|val_loss|val_mIoU|LB score|config| |------|---|---|---|---|---|---|---|---|---| |Efficientnet Unet|16|7|83s|22|CrossEntropy|0.133|0.384|0.373|0.4499|config01| |Resnext50 Unet|16|7|67s|9|0.157|CrossEntropy|0.372|0.371|0.4679|config02|

Unet은 성능이 떨어진다.


### ### ## DLV3+ resnext50 with augmentation

|모델|배치|시드|time/epoch|epoch|loss|train_loss|val_loss|val_mIoU|LB score|config| |------|---|---|---|---|---|---|---|---|---| |DLV3+ Resnext50|8|7|103s|16|IoU+CE|0.681|0.425|0.419|0.5756|config03|

현진님의 best model에서 augmentation만 추가했습니다. aug 결과 성능이 살짝 좋아졌네요. 512*512 size로 학습하면 더 좋아질 듯

used aug

  • config03 : HorizontalFlip, ShiftScaleRotate, RandomBrightnessContrast, GaussNoise
Clone this wiki locally