Has anyone succeeded in DAGM2007 dataset？

[a03ki]

Has anyone succeeded in DAGM2007 dataset？

[A03ki]

Hi, @18813223994. Thank you for using this repository.
This image looks selecting a small epoch number. Could you try to select big epoch number.

Example:

$ python train_wgangp.py "DAGM2007/Class6/train" --seed 1 --n_epochs 1000 --channels 3 --lr 0.0001
$ python train_encoder_izif.py "DAGM2007/Class6/train" --seed 1 --n_epochs 200  --channels 1 --lr 0.0001
$ python test_anomaly_detection.py "DAGM2007/Class6/test" --channels 3
$ python save_compared_images.py "DAGM2007/Class6/test" --n_grid_lines 10  --channels 3

We can check no scratches on the reconstructed images.
However, this abnormal detection score isn't high.

The existing model seems weak for tasks such as a carpet from my experience.

---

Training GAN and Encoder setting grayscale maybe give you a little more score because images in DAGM2007 are gray color.

Example:

$ python train_wgangp.py "DAGM2007/Class6/train" --seed 1 --n_epochs 1000 --channels 1 --lr 0.0001
$ python train_encoder_izif.py "DAGM2007/Class6/train" --seed 1 --n_epochs 200  --channels 1 --lr 0.0001
$ python test_anomaly_detection.py "DAGM2007/Class6/test" --channels 1
$ python save_compared_images.py "DAGM2007/Class6/test" --n_grid_lines 10  --channels 1

Before run the above, you should rewritten as follows in your_own_dataset/train_wgangp.py and your_own_dataset/train_encoder_izif.py:

transform = transforms.Compose([transforms.Resize([opt.img_size]*2),
                                                  transforms.RandomHorizontalFlip(),
                                                  transforms.RandomVerticalFlip(),
                                                  transforms.Grayscale(num_output_channels=1),
                                                  transforms.ToTensor(),
                                                  transforms.Normalize([0.5], [0.5])])

Moreover, rewritten in your_own_dataset/test_anomaly_detection.py and your_own_dataset/save_compared_images.py:

transform = transforms.Compose([transforms.Resize([opt.img_size]*2),
                                                  transforms.Grayscale(num_output_channels=1),
                                                  transforms.ToTensor(),
                                                  transforms.Normalize([0.5], [0.5])])

---

DL the Class6 dataset and split data.

$ wget https://conferences.mpi-inf.mpg.de/dagm/2007/Class6.zip
$ wget https://conferences.mpi-inf.mpg.de/dagm/2007/Class6_def.zip
$ mkdir -p DAGM2007/Class6
$ unzip -d DAGM2007/Class6/train Class6.zip
$ unzip -d DAGM2007/Class6/test Class6_def.zip

import glob
import shutil
import os

output_dir = "DAGM2007/Class6/test/Class6/"
os.makedirs(output_dir, exist_ok=True)

path = "DAGM2007/Class6/train/Class6/*"
file_paths = sorted(glob.glob(path))

for file_path in file_paths[len(file_paths)//2:]:
    new_path = shutil.move(file_path, output_dir)