search

LeeJunHyun / Image_Segmentation

Pytorch implementation of U-Net, R2U-Net, Attention U-Net, and Attention R2U-Net.
2.68k stars 595 forks source link

The result is not good in ISIC 2018 #24

Closed FENGShuanglang closed 5 years ago

FENGShuanglang commented 5 years ago

Jaccard and F1 only 0.5-0.7,are you sure? On the official competition website, the result common is 0.80+ in term of jaccard. In addition, the result of Unet in my experiment is 0.78+ in term of jaccard, Yes! ISIC 2018. 2594 images.I resize it to 256*192

zz10001 commented 5 years ago

can you show how to train it?i just download the task one like this image what is the train/valid/test train_GT/valid_GT/test_GT in dataset.py. thx

FENGShuanglang commented 5 years ago

No,you can download this https://challenge.kitware.com/#phase/5abcb19a56357d0139260e53

zz10001 commented 5 years ago

i downlaod from the link but when i extract，it only the some with what i show.and can you share you path in dataset.py like this.should i change something image

zz10001 commented 5 years ago

No,you can download this https://challenge.kitware.com/#phase/5abcb19a56357d0139260e53

i can't find valid ground truth and test ground truth,when i run it .it occurs image

FENGShuanglang commented 5 years ago

I download and pre-process it in myself code

LeeJunHyun commented 5 years ago

Hi @FENGShuanglang . The score can be differ according to train strategy such as optimizer, initialization, and early stopping.

LeeJunHyun commented 5 years ago

If you don't have further issues, I will close the issue. I hope my answer was helpful :)