High-quality and Diverse Few-shot Image Generation via Masked Discrimination

Official Implementation of our paper accepted by TIP 2024.

Early Access: https://ieeexplore.ieee.org/document/10496521

Requirements

• Linux
• NVIDIA GPU (NVIDIA TITAN RTX in our experiments) + CUDA CuDNN 11.2
• PyTorch 1.7.0
• torchvision 0.8.1
• Python 3.6.9
• Install all the required libraries: pip install -r requirements.txt

Training and Evaluation

Prepare datasets

Transfer target images to lmdb format for GAN adaptation.

CUDA_VISIBLE_DEVICES='0' python prepare_data.py --out path/to/datasets --size 256 path/to/original/images

Train the adapted model and evaluate Intra-LPIPS

1.Checkpoints and samples are saved in checkpoints/exp_name and samples/exp_name automatically.

2.We implement codes to evaluate Intra-LPIPS with fixed input noise vectors and print results in training process.

3.The noise vectors used for our experiments are provided in "test_noise.pt".

CUDA_VISIBLE_DEVICES='0' python train.py --ckpt path/to/source/model --data_path path/to/datasets --exp <exp_name> --dataset <dataset_name>

FID evaluation

We follow prior works to use pytorch-fid for FID evaluation (carried out for abundant datasets).

Install pytorch-fid through pip first:

pip install pytorch-fid==0.1.1

Generate fake images for FID evaluation:

CUDA_VISIBLE_DEVICES='0' python generate.py --ckpt_target /path/to/model/ --imsave_path /path/to/fake/images

Calculate FID

CUDA_VISIBLE_DEVICES='0' python -m pytorch_fid /path/to/real/images /path/to/fake/images

Citation

If you are doing research related to our work, please consider to cite it!

@ARTICLE{zhu2024few,
  author={Zhu, Jingyuan and Ma, Huimin and Chen, Jiansheng and Yuan, Jian},
  journal={IEEE Transactions on Image Processing}, 
  title={High-quality and Diverse Few-shot Image Generation via Masked Discrimination}, 
  year={2024},
  doi={10.1109/TIP.2024.3385295}}