In this branch, it's written in Keras and mostly copy from: https://github.com/decouples/Unet
Because my research direction is not Biomedical Image Segmentation, I do this just for interest and I repaire a mis-match problem between test and predict images. So, I re-produce it in Pytorch 1.0.0(pytorch branch), and use ISBI dataset(http://brainiac2.mit.edu/isbi_challenge/). You can also download from my link below in How to run.
ps: I don't use any data augmentation methods. My chinese blog is here.
- ubuntu18.04 + GTX TITAN XP
- keras2.2.4 + tensorflow-gpu1.13.1 + cuda10.0 + cudnn7.4.2
- torch1.0.0 + torchvision0.2.0
In this branch, you first need to create a new folder in this directory named results for saving model predicted pictures, like in Linux, you can run command in terminal:
mkdir results
Next step is: you need download ISBI dataset(baidu drive pswd:jwix or google drive) and extract it in this directory named raw in which there has three sudfolders: train, label and test.
Then you can run the code.
Now I display a 7-th results from keras branch:
In windows, it has some small mistakes with glob.glob, but in pytorch branch it has no this bug.