Constrained domain adaptation for segmentation. Pytorch implementation of our method for adapting Semantic Segmentation using constrained optimization. We compare to adversarial adaptation, using the method proposed here : https://github.com/wasidennis/AdaptSegNet.
Install PyTorch from http://pytorch.org with Python 2 and CUDA 8.0
The dataset is available at https://ivdm3seg.weebly.com/data.html. We used the Water (Wat) modality as the Source domain, and the In-Phase (IP) modality as the Target domain. Note: in our experiments we rotated the data from the sagittal plane to the transverse plane.
Download the data and put in the data/sagittal folder, then rotate and save into the data/transverse folder, both for the Wat and the IP modality:
python rotate.py --base_folder='./data/transverse/IP/' --folders=['train','val'] --save_folder='./data/sagittal/IP/' --rot=’rot’ --grp_regex="Subj_\\d+_"
Train the model with constrained domain adaptation. Create the Results folder results/IP. Results will be saved in the result folder results/IP/Constraint/
python main.py --dataset='data/transverse/' --target_dataset='data/transverse/'
--folders="[('Wat', png_transform, False), ('GT', gt_transform, False)]"
--target_folders="[('IP', png_transform, False), ('GT', gt_transform, False)]+"
"[('GT', gt_transform, False), ('GT', gt_transform, False)]"
--grp_regex='Subj_\\d+_\\d+'
--losses="[('CrossEntropy', {'idc': [0,1], 'weights':[1,1]}, None, None, None, 0),"
" ('NaivePenalty', {'idc': [1]},'PreciseBounds', {'margin': 0.1, 'mode': 'percentage'}, 'soft_size', 2.5)]"
--losses_source="[('CrossEntropy',{'idc': [0,1],'weights':[1,1]},None,None,None, 1)]"
--model_weights='pretrained_source.pkl' --n_class=2 --in_memory=False,
--network='ENet' --resize=36 --scheduler='DummyScheduler'
scheduler_params='{}', workdir='results/Constraint'