EncodeGAN-PyTorch

A PyTorch implementation of EncodeGAN

this project is used to achieve face image fusion and attribute change tasks

Requirements

• Python 3.6
• PyTorch
• TensorboardX

pip install -r requirements.txt

• Dataset
  • CelebA dataset
    • Images should be placed in data/CelebA/img/img_align_celeba/*.jpg
    • Attribute labels should be placed in data/CelebA/anno/list_attr_celeba.txt
  • test_data:you can use CelebA's test dataset, or select some face image from CelebA and add them in test_data/*.jpg.

data
├── CelebA
    ├── img
            ├── img_align_celeba
    ├── anno
            ├── list_attr_celeba.txt
├── test_data
    ├── *.jpg

Usage

To train an EncodeGAN on CelebA 128x128

CUDA_VISIBLE_DEVICES=0 \
python train.py --gpu=True --data_save_root=output --experiment_name=Encode_GAN --total_steps=100000 --latent_dim=128 --batch_size=32 --b1=0 --b2=0.999 --data_path='data/CelebA/img/img_align_celeba' --attr_path='data/CelebA/anno/list_attr_celeba.txt' --data_save_root='output' --E_mode='enc' --n_e=2

To test EncodeGan in interpolation capabilities on image

CUDA_VISIBLE_DEVICES=0 \
python face_latent_space_explore_test.py --gpu=True --data_save_root=output --experiment_name=face_latent_space_explore_test --weight_path=output/Encode_GAN/checkpoint/weights.99999.pth --setting_path=output/Encode_GAN/setting.txt --test_data_path=test_data

To test EncodeGan's face attribute editing ability

CUDA_VISIBLE_DEVICES=0 \
python face_attr_change_test.py --gpu=True --data_save_root=output --experiment_name=face_attr_change_test --data_path=data/CelebA/img/img_align_celeba --attr_path=data/CelebA/anno/list_attr_celeba.txt --data_save_root=output --weight_path=output/Encode_GAN/checkpoint/weights.99999.pth --setting_path=output/Encode_GAN/setting.txt --test_data_path=test_data

To visualize training details

tensorboard \
--logdir output/your_experiment_name/summary