https://dacon.io/competitions/official/235894/overview/description
Classification on MVtec AD Dataset
Macro f1 score
- Preprocessing
- Resize to (512 x 512)
- Data augmentation: flip left-right or up-down
- Training
- Model:
efficientnet_b0
- Optimizer:
Adam
(lr=1e-3) - Loss:
CEE
- Gradient scaler
- Model:
- Evaluation
- Training metric:
0.96759
- Test metric:
0.66579
- Training metric:
- FastFlow (Detection AUROC:
99.4
)- Rank 1
- Unsupervised anomaly detection
- Early stopping 적용
- TensorFlow porting
- Validation
- validation_on: optimal epochs 결정 -> 더 좋은 성능
- validation_off: train_full로 학습
- Sample weight 적용
- Model:
efficientnet_b0
- 2-level classification
- Classification(
class
)
model1
: supervised - Classification(
label
)
model2
: supervised
- Classification(
- 3-level classification
- Classification(
class
)
model1
: supervised (efficientnet_b0
) - Anomaly detection
model2
: unsupervised (PatchCore
) - Classification(
label
)
model3
: supervised (efficientnet_b0
)
- Classification(