"SPatchGAN: A Statistical Feature Based Discriminator for Unsupervised Image-to-Image Translation" (ICCV 2021)

by CJWBW

• CUDA 10.0
• Python 3.6
• pip install -r requirements.txt

• Dataset structure (dataset_struct='plain')

- dataset
    - <dataset_name>
        - trainA
            - 1.jpg
            - 2.jpg
            - ...
        - trainB
            - 3.jpg
            - 4.jpg
            - ...
        - testA
            - 5.jpg
            - 6.jpg
            - ...
        - testB
            - 7.jpg
            - 8.jpg
            - ...

• Supported extensions: jpg, jpeg, png
• An additional level of subdirectories is also supported by setting dataset_struct to 'tree', e.g.,

- trainA
    - subdir1
        - 1.jpg
        - 2.jpg
        - ...
    - subdir2
        - ...

• Selfie-to-anime:

  • The dataset can be downloaded from U-GAT-IT.

• Male-to-female and glasses removal:

  • The datasets can be downloaded from Council-GAN.
  • The images must be center cropped from 218x178 to 178x178 before training or testing.
  • For glasses removal, only the male images are used in the experiments in our paper. Note that the dataset from Council-GAN has already been split into two subdirectories, "1" for male and "2" for female.

• Set the suffix to anything descriptive, e.g., the date.
• Selfie-to-Anime

python main.py --dataset selfie2anime --augment_type resize_crop --n_scales_dis 3 --suffix scale3_cyc20_20210831 --phase train

• Male-to-Female

python main.py --dataset male2female --cyc_weight 10 --suffix cyc10_20210831 --phase train

• Glasses Removal

python main.py --dataset glasses-male --cyc_weight 30 --suffix cyc30_20210831 --phase train

• Find the output in ./output/SPatchGAN_<dataset_name>_<suffix>
• The same command can be used to continue training based on the latest checkpoint.
• For a new task, we recommend to use the default setting as the starting point, and adjust the hyperparameters according to the tips.
• Check configs.py for all the hyperparameters.

• Replace --phase train with --phase test

• Replace --phase train with --phase freeze_graph
• Find the saved frozen model in ./output/SPatchGAN_<dataset_name>_<suffix>/checkpoint/pb

cd frozen_model
python test_frozen_model.py --image <input_image_or_dir> --output_dir <output_dir> --model <frozen_model_path>

• Download the pretrained models from huggingface, and put them in the output directory.
• You can test the checkpoints (in ./checkpoint) or the frozen models (in ./checkpoint/pb). Either way produces the same results.
• The results generated by the pretrained models are slightly different from those in the paper, since we have rerun the training after code refactoring.
• We set n_scales_dis to 3 for the pretrained selfie2anime model to further improve the performance. It was 4 in the paper. See more details in the tips.
• We also provide the generated results of the last 100 test images (in ./gen, sorted by name, no cherry-picking) for the calibration purpose.

• We provide a PyTorch implementation of the SPatchGAN discriminator in spatchgan_discriminator_pytorch.py.

@InProceedings{Shao_2021_ICCV,
    author    = {Shao, Xuning and Zhang, Weidong},
    title     = {SPatchGAN: A Statistical Feature Based Discriminator for Unsupervised Image-to-Image Translation},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2021},
    pages     = {6546-6555}
}

• Our code is partially based on U-GAT-IT.