habout632/stargan2

Unofficial Pytorch version StarGAN v2

---

Unofficial PyTorch implementation of StarGAN-v2.

This repository is forked from Official StarGAN PyTorch v1.0 and add new features to the model according to Paper StarGAN v2. StarGAN can flexibly translate an input image to any desired target domain using only a single generator and a discriminator.

Paper

StarGAN v2: Diverse Image Synthesis for Multiple Domains

Yunjey Choi, Jung-Woo Ha   
Clova AI Research (NAVER Corp.)
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020

Dependencies

• Python 3.5+
• PyTorch 1.0.0+
• TensorFlow 1.3+ (optional for tensorboard)

Usage

1. Cloning the repository

$ git clone https://github.com/habout632/StarGAN2.git
$ cd StarGAN/

2. Dataset Structure

you need to arrange your images as torchvision Suggested ImageFolder in the following way , for example, afhq：

root/ants/xxx.png
root/ants/xxy.jpeg
root/ants/xxz.png
.
.
.
root/bees/123.jpg
root/bees/nsdf3.png
root/bees/asd932_.png

where ‘ants’, ‘bees’ etc. are class labels.

AFHQ Baidu Disk Link extraction code: ptm8 Useful for China Mainland Developers

3. Training

To train StarGAN on dataset(CelebA-HQ, AFHQ or any other your own custom dataset, run the training script below. See here for a list of selectable attributes in the CelebA dataset. If you change the selected_attrs argument, you should also change the c_dim argument accordingly.

$ python main.py --mode train --batch_size 2 --image_size 256 --num_domains 2 --beta1 0 \
                 --beta2 0.99 --g_conv_dim 32 --num_iters 900000 --resume_iters 600000 \
                 --sample_dir /data2/starganv2/celeba/samples --model_save_dir /data2/starganv2/celeba/models \
                 --result_dir /data2/starganv2/celeba/results --image_dir /data/datasets/celeba-hq/train                  

Hardware

we train StarGAN v2 model on Nvidia GTX 1080 ti GPU, and the number of iterations is more or less 900K, about one week.

4. Testing

To test StarGAN v2 on dataset:

$ python main.py --mode test --dataset CelebA --image_size 128 --c_dim 5 \
                 --sample_dir stargan_celeba/samples --log_dir stargan_celeba/logs \
                 --model_save_dir stargan_celeba/models --result_dir stargan_celeba/results \
                 --selected_attrs Black_Hair Blond_Hair Brown_Hair Male Young

5. Pretrained model

Once finished training on celebhq, we will publish the pretrained model.The Pretrained model will be provided in two following ways: Google Drive Baidu Disk To download a pretrained model checkpoint, run the script below. The pretrained model checkpoint will be downloaded and saved into ./stargan_celeba_256/models directory.

$ bash download.sh pretrained-celeba-256x256

To translate images using the pretrained model, run the evaluation script below. The translated images will be saved into ./stargan_celeba_256/results directory.

$ python main.py --mode test --dataset CelebA --image_size 256 --c_dim 5 \
                 --selected_attrs Black_Hair Blond_Hair Brown_Hair Male Young \
                 --model_save_dir='stargan_celeba_256/models' \
                 --result_dir='stargan_celeba_256/results'

Results

1. Latent Guided Image Generation on CelebA-HQ

2. Latent Guided Image Generation on CelebA-HQ

3. Reference Guided Image Generation on CelebA-HQ, work still in progress

Acknowledgement

This work was mainly inspired by StarGAN_v2-Tensorflow, appreciate helping of the author Junho Kim

Contribution

we still need your help, as we stuck on reference guided image synthesis and computing resources(multiple gpus training), any contribution(demos, code and the docs)is welcomed

Contributers

habout
ryan

stargan2's Issues

the result of celebA-HQ maybe is wrong ?

Can you upload the log for all loss

Excuse me,is the code of multi-FC layer of model wrong?

also, i do not how to make this FC code according to the paper....5555