This is the source code of paper Generalizable Semantic Segmentation via Model-agnostic Learning and Target-specific Normalization).
Pytorch 1.2.0, GPU : 4 * V100,
The training procedure needs gta5(G), synthia_rand_citys(S), idd(I), mapillary(M) and cityscapes(C). Please download them and put to the same folder which can be specified in paths.py. The folder tree structure is as follows:
├── GTA5
│ ├── images
│ ├── labels
│ ├── mapping.mat
│ └── split.mat
├── IDD
│ ├── codes
│ ├── gtFine
│ └── leftImg8bit
├── SYNTHIA
│ └── RAND_CITYSCAPES
├── cityscapes
│ ├── gtFine
│ ├── leftImg8bit
│ ├── train_fine.txt
│ └── val_fine.txt
├── mapillary
│ ├── README
│ ├── config.json
│ ├── demo.py
│ ├── requirements.txt
│ ├── testing
│ ├── training
│ └── validation
We also generated some synthetic images with CycleGAN, we download the style_ukyoe.pth, style_vangogh.pth, style_cezanne.pth pretrained models and transfered GTA5 dataset with these three pretrained models. The generated images are all put in the 'style_xxx' folder directly, which is the same as the images folder of GTA5.
The model can be trained with train.py. If we want to train the MLDG with four source domains(G, S, I, M) and one target domain(C), we can parse these arguments.
python train.py --name exp --source GSIM --target C --train-num 1
For more details, please refer to train.py
The model can be trained with following command using eval.py.
python eval.py --name exp --targets C --test-size 16
python predict.py --name exp --targets C --test-size 16