retinopathy database trainval.csv and test.csv file

[liminghu]

I am interested in the ordinal classifier, and are studying it using the retinopathy database.
I have downloaded https://www.kaggle.com/c/diabetic-retinopathy-detection/data

according to the source code: src/datasets2.py, we also need:
self._df_trainval = pd.read_csv(os.path.join(DATASETS_DIR, 'retinopathy/data128/trainval.csv'))
self._df_test = pd.read_csv(os.path.join(DATASETS_DIR, 'retinopathy/data128/test.csv'))

Do you mind sharing these two files?

Thanks.

[victormvy]

Hello,

I'm glad to see that you are interested in this ordinal classifier. You can find attached those csv files along with the python script that we used to create them.
retinopathy.zip

[liminghu]

Thanks a lot. I checked the *.py script you shared, I noticed that it was looking for *.bmp files, but the original image files are in *.jpeg format. I also read your paper again: The images are resized to 128 by 128 pixels and rescaled to [0, 1] range.
Can you share other scripts? I just want to follow the steps to duplicate your result.

Thanks.

[victormvy]

I have been looking for the script that we used to resize the images to 128x128 but I could not find it. However, I can provide you the resized images that I used (2.73GB zip). The [0,1] rescaling is done in the generator class that is used to load the images. I hope that you can replicate the experiments. Let me know if you need something else.

Dataset download: https://mega.nz/file/TSZDTKCT#jm4I5s-xahLJ0C3vsbtmswPcCJpaW69g4xJar5I02Ew

[liminghu]

Thanks. I will try it again.

[liminghu]

I just noticed that the repo used tensorflow v1.3 which needs lower version of CUDA:
https://stackoverflow.com/questions/50622525/which-tensorflow-and-cuda-version-combinations-are-compatible
My env is higher CUDA version, hard to make it work.

[liminghu]

I checked the Diabetic Retinopathy Detection leadboard, the leading QWK score is around 0.85~0.86 which is far above the scores reported in the paper: 0.58.

[victormvy]

You can achieve higher QWK score using higher resolution images. The original images are high resolution but we rescaled them to 128x128 due to our hardware limitations. In our work, we compared our method with other methods that used 128x128 images too.

[liminghu]

Thanks. I will try it. The winning solution is using a resolution of 300*~.
https://github.com/btgraham/SparseConvNet/blob/kaggle_Diabetic_Retinopathy_competition/competitionreport.pdf