problem about dataset

[rezashaemi]

Thanks for your great project! I downloaded REFUGE dataset but the name and number of image files are different from names in this repo! For example, training part has 360 files with names 0001.jpg, ...

[askerlee]

Hi @rezashaemi , you can download the preprocessed data from baidu netdisk: https://pan.baidu.com/s/1HAB5AMV8gqgoi5joUFdpVQ?pwd=5iud

[rezashaemi]

Thanks a lot! It seems it doesn't allow someone out of china download data!

[askerlee]

Sorry for that. Could you pls try the dropbox link: https://www.dropbox.com/s/fd8hty3tgwwsl1i/refuge2020.tar?dl=0

[rezashaemi]

Great! Thanks again,

[mahdiyehbakhshi822]

hi Thank you for great code I also have the same problem with the rim database. The name of the images I downloaded is different from the name of the data code.

[askerlee]

hi @mahdiyehbakhshi822 , the preprocessed RIM data can be downloaded at: https://www.dropbox.com/s/8osan2fuon7nb0q/rim.zip?dl=0

[mahdiyehbakhshi822]

Thank you

[Lemonweier]

Sir，can you tell me something about the dataset you give? Thank u. 1.refuge2020g0001_576_597,0``n0001_576_712,18 they looks like 5 parts, I guess "g" and "n" are glaucoma or not."576" is the size, the other parts I don't knew. drishtitrain-089_576_32,32``` what's the meaning of "32" "32" 2. the lists show images/g0002_560_697,47.png images/g0002_576_689,39.png`,why there are two of each one image? Sorry to bother you again.

[askerlee]

@Lemonweier the two numbers in the filename are the (x,y) coordinates that each image is cropped at: https://github.com/askerlee/segtran/blob/master/code/MNet_DeepCDR/Step_1_Disc_Crop.py#L124

[Lemonweier]

Thank you for your reply!

[OvO1111]

Thanks for the project, but is there any way to train the REFUGE2020 model without the help of drishti and rim datasets? I downloaded your cropped refuge dataset and removed the 'drishti,rim' from attribute ds_names inside fundus in train2d.py and trained&tested the model accordingly, but the dice for both cup and disk is tested to be 0, i fear there is something i did wrong, can you help me? thanks in advance!

[askerlee]

Hi @OvO1111 yes of course you can train without drishti and rim. I guess maybe when you modified ds_names, some kind of bug was introduced? Could you try an alternative approach: do not change the code, but only remove the image and mask files in the drishti and rim folders, and do the training?

[OvO1111]

ok thanks for your help, i'll try

EDIT: after running ./train2d.sh --task fundus --split all --translayers 3 --layercompress 1,1,2,2 --net segtran --bb eff-b4 --maxiter 10000 --bs 6 --noqkbias, it prompts FileNotFoundError: [Errno 2] No such file or directory: '../data/fundus/drishti/images', this trace back to db_train = init_training_dataset(...) in train2d.py, i guess i have to modify theds_names somehow...

[OvO1111]

Oh i'm sorry, removing the dataset names from ds_names should work, i accidentally made a mistake in the original code, thanks for your support

[geek-APTX4869]

Hi @rezashaemi , you can download the preprocessed data from baidu netdisk: https://pan.baidu.com/s/1HAB5AMV8gqgoi5joUFdpVQ?pwd=5iud

REFUGE20验证集有400张图片，但链接只有120张？

[askerlee]

I've checked. There are four folders, each folder having 400 images.

[geek-APTX4869]

I've checked. There are four folders, each folder having 400 images.

all the test.list have 120 lines