• Tensorflow implement of U-Net: Convolutional Networks for Biomedical Image Segmentation..[Paper]
• Borrowed code and ideas from zhixuhao's unet: https://github.com/zhixuhao/unet.

First ensure that you have installed the following required packages:

• TensorFlow1.4.0 (instructions). Maybe other version is ok.
• Opencv (instructions). Here is opencv-2.4.9.

• In this implementation of the Unet, we use Carvana Image Masking Challenge data.[download] We download train.zip and train_masks.zip. You can put all the datasets in datasets folder.
• Run scripts/transform_images.py to transform all the image to gray JPEG image.
• Run scripts/build_tfrecords.py to generate training data, data format is tfrecords.

• Run the following script to train the model, in the process of training, will save the training images every 500 steps. See the model/unet.py for details.

sh train.sh

You can change the arguments in train.sh depend on your machine config.

• Run the following script to test the trained model. The test.sh will transform the datasets.

sh test.sh

The script will load the trained StarGAN model to generate the transformed images. You could change the arguments in test.sh depend on your machine config.

• Pretrained model: [download]. The model-8500 is better.