MVtec AD dataset anomaly detection (Image Classification)

The following codes are the solutions (1st place, private score: 0.92708) for the dacon competition.
If you would like to know more about the competition, please refer to the following link:
https://dacon.io/competitions/official/235894/overview/description

Briefly, the task is to classify an extremely imbalanced images into 88 classes in which the label is composed of class-state pairs.
To solve this problem, we used Efficientnet and Resnext as a backbone with different training methods (e.g. one-class self-supervised learning, arcFace loss, and label smoothing).
We also blended the model weights and prediction results based on the validation results and hypothesis.
Please refer to ensemble.py to see how we blended the model weights and prediction results.

Note that we were not able to reproduce our best private score (0.9270) perfectly, but got the close private score (0.9264).

Blending strategy

• Baseline1 (LB 0.896): EfficientNet-B6 (66) & EfficientNet-B6 (85, ugmentation including rotation of 45 degree)
• soft_ensemble (LB 0.899): EfficientNet-B6 (66) & EfficientNet-B6 (85, rotate 45) & EfficientNet-B6 (39, arcFace loss) & EfficientNet-B6 (92, arcFace & label smoothing loss
• Baseline2 (LB 0.894): EfficientNet-B6 (146, label smoothing loss)
• One-class based self-supervised learning model: EfficientNet-B6 (87, excluding arcFace will be better)
• We drop the filp augmentations for some cases (e.g., metal_nut) on the one-class based SSL

---

HARD ENSEMBLE WITH HYPOTHESIS 1

Observation 1-1: On the overall validation results, the anomaly class-combined was relatively hard to predict.
Observation 1-2: Baseline1 has relatively better performance for the normal and anomaly-combined on the validation results.

Hypothesis 1: Trust the results from the baseline1 and blend the new perspective from the soft_ensemble result except for both normal and anomaly-combined.

Result: LB 0.902

Priority 1: Trust a new perspective for a anomaly class except for the combined class.
Priority 2: Trust normal class and anomaly class-combined from the baseline1.

---

HARD ENSEMBLE WITH HYPOTHESIS 2

Observation 2-1: The baseline2 showed better performance for the classes tile, carpet, and zipper on the validation compared to the baseline1.
Observation 2-2: The baseline2 exhibited a different view for the anomaly class-combined on the validation compared to the baseline1.

Hypothesis 2-1: Trust the new perspective from the baseline2 results including the classes tile, carpet, and zipper.
Hypothesis 2-2: Trust the new perspective for the anomaly class-combined from the baseline2 results.

Result: LB 0.9099

Priority 1: Trust a new perspective for the anomaly class-combined (modified).
Priority 2: Trust a new perspective for anomaly classes tile, carpet, and zipper from the baseline2.
Priority 3: Trust normal and anomaly class-combined from the baseline1.

---

HARD ENSEMBLE WITH HYPOTHESIS 3

Hypothesis 3: One-class based self supervised learning model has a different view compared with the baseline1 and 2.

The combination of labels, 'cable', 'grid', 'metal_nut', 'pill', and 'wood', showed the best performance.

Result: LB-Private 0.926

Priority 1: Trust a new perspective for the anomaly class-combined.
Priority 2: Trust a new perspective for anomaly classes tile, carpet, and zipper from the baseline1.
Priority 3: Trust a new perspective for anomaly classes cable, grid, metal_nut, pill, and wood from the one-class based self supervised learning model.
Priority 4: Trust normal class from the baseline1 (modified).

---

Used Models

The models we used to make the best prediction are as follows:

• features_66: EfficientNet-B6, 10-Fold, baseline, inference img_size=640, inference batch_size=64

• features_85: EfficientNet-B6, 10-Fold, baseline, added augmentation (rotate 45 degree), inference img_size=640, inference batch_size=64

• features_146: EfficientNet-B6, 10-Fold, baseline, label smoothing applied

• features_39: EfficientNet-B6, 10-Fold, arcFace loss applied, inference img_size=640, inference batch_size=64

• features_92: EfficientNet-B6, 5-Fold, label smoothing with arcFace applied, inference img_size=640, inference batch_size=64

• features_100: ResNext, 5-Fold, resnext101_32x8d, batch_size=16, fold=5

• validation_E6_87.csv: EfficientNet-B6, one-class self-supervised learning strategy, arcFace loss applied

Data Directory

You can download the dataset at here: https://dacon.io/competitions/official/235894/data

anomaly
├── data
│   ├── 5-Fold_idx.npy
│   ├── 10-Fold_idx.npy
│   ├── train
│   │   ├── images
│   │   ├── train_df.csv
│   ├── test
│   │   ├── images
│   ├── sample_submission.csv
      .
      .
      .

Requirements

• albumentations >= 1.1.0
• opencv-python >=4.5.5.64
• pandas >= 1.3.5
• scikit-learn >= 1.0.2
• timm >= 0.5.4
• torch >= 1.8.2
• torchvision >= 0.9.2
• tqdm >=4.64.0

Run

• To run the codes, use the following commands:
  

# EfficientNet-B6 with label smoothing
python main.py train --arch 6 --img_size 576 --criterion smoothing

# Different backbone model (example of ResNest)
python main.py train --model resnest --model_name resnest50d_4s2x40d

# One-class training with ArcFace
python main.py train --train_method one_class --arcloss arcface

• You can choose either 10- or 5-fold to train the model.
• We used the fixed training and validation index for each fold for solid experiment.

Configurations

• arch: EfficientNet backbone scale (E0 ~ E7)
• model: backbone model (e.g. efficientnet and resnest)
• model_name: name of the pretrained model that is available in timm
• train_method: training strategy
• arcloss: ArcFace usage (default: False)
• fold: number of folds (different number of folds other than 5 or 10 is not available due to the fixed index)
• multi_gpu: multi-gpu learning options