AttGAN-PyTorch

A PyTorch implementation of AttGAN - Arbitrary Facial Attribute Editing: Only Change What You Want

Test on the CelebA validating set

Test on my custom set

Inverting 13 attributes respectively. From left to right: Input, Reconstruction, Bald, Bangs, Black_Hair, Blond_Hair, Brown_Hair, Bushy_Eyebrows, Eyeglasses, Male, Mouth_Slightly_Open, Mustache, No_Beard, Pale_Skin, Young

The original TensorFlow version can be found here.

Requirements

• Python 3
• PyTorch 0.4.0
• TensorboardX

pip3 install -r requirements.txt

If you'd like to train with multiple GPUs, please install PyTorch v0.4.0 instead of v1.0.0 or above. The so-called stable version of PyTorch has a bunch of problems with regard to nn.DataParallel(). E.g. pytorch/pytorch#15716, pytorch/pytorch#16532, etc.

pip3 install --upgrade torch==0.4.0

• Dataset
  • CelebA dataset
    • Images should be placed in ./data/img_align_celeba/*.jpg
    • Attribute labels should be placed in ./data/list_attr_celeba.txt
  • HD-CelebA (optional)
    • Please see here.
  • CelebA-HQ dataset (optional)
    • Please see here.
    • Images should be placed in ./data/celeba-hq/celeba-*/*.jpg
    • Image list should be placed in ./data/image_list.txt
• Pretrained models: download the models you need and unzip the files to ./output/ as below,

  output
  ├── 128_shortcut1_inject0_none
  ├── 128_shortcut1_inject1_none
  ├── 256_shortcut1_inject0_none
  ├── 256_shortcut1_inject1_none
  ├── 256_shortcut1_inject0_none_hq
  ├── 256_shortcut1_inject1_none_hq
  ├── 384_shortcut1_inject0_none_hq
  └── 384_shortcut1_inject1_none_hq
  

Usage

To train an AttGAN on CelebA 128x128

CUDA_VISIBLE_DEVICES=0 \
python train.py \
--img_size 128 \
--shortcut_layers 1 \
--inject_layers 1 \
--experiment_name 128_shortcut1_inject1_none \
--gpu

To train an AttGAN on CelebA-HQ 256x256 with multiple GPUs

CUDA_VISIBLE_DEVICES=0 \
python train.py \
--data CelebA-HQ \
--img_size 256 \
--shortcut_layers 1 \
--inject_layers 1 \
--experiment_name 256_shortcut1_inject1_none_hq \
--gpu \
--multi_gpu

To visualize training details

tensorboard \
--logdir ./output

To test with single attribute editing

CUDA_VISIBLE_DEVICES=0 \
python test.py \
--experiment_name 128_shortcut1_inject1_none \
--test_int 1.0 \
--gpu

To test with multiple attributes editing

CUDA_VISIBLE_DEVICES=0 \
python test_multi.py \
--experiment_name 128_shortcut1_inject1_none \
--test_atts Pale_Skin Male \
--test_ints 0.5 0.5 \
--gpu

To test with attribute intensity control

CUDA_VISIBLE_DEVICES=0 \
python test_slide.py \
--experiment_name 128_shortcut1_inject1_none \
--test_att Male \
--test_int_min -1.0 \
--test_int_max 1.0 \
--n_slide 10 \
--gpu

To test with your custom images (supports test.py, test_multi.py, test_slide.py)

CUDA_VISIBLE_DEVICES=0 \
python test.py \
--experiment_name 384_shortcut1_inject1_none_hq \
--test_int 1.0 \
--gpu \
--custom_img

Your custom images are supposed to be in ./data/custom and you also need an attribute list of the images ./data/list_attr_custom.txt. Please crop and resize them into square images in advance.