ybigta-hair-styling

This repository is for recording YBIGTA 18th freshman project(Feb, 2021)

Requirements

python>=3.6
numpy==1.19.4
matplotlib==3.3.3
opencv-python==4.5.1.48
torch==1.17.0
torchvision==0.8.1

Train

There are 3 ways to train your hair styling application.

Note that you should change your argument --image_path in configs.py correctly!

For detailed descriptions, we recommend you to check configs.py in each directory.

1. Pix2Pix-Hair

cd pix2pix-hair
python main.py

sampled images during training

2. Pix2Pix-Skin and Hair

cd pix2pix
python main.py

sampled images during training

3. MichiGAN

cd michigan
python train.py --name [name_experiment] --batchSize 8 --no_confidence_loss --gpu_ids 0,1,2,3,4,5,6,7 --no_style_loss --no_rgb_loss --no_content_loss --use_encoder --wide_edge 2 --no_background_loss --noise_background --random_expand_mask --use_ig --load_size 568 --crop_size 512 --data_dir [pah_to_dataset] ----checkpoints_dir ./checkpoints

Evaluate

1. PixPix-Hair

cd pix2pix-hair
python inference.py --your_pic {your_img_path} --celeb_pic {celeb_img_path}

Sample

2. PixPix-Skin and Hair

cd pix2pix
python inference.py --your_pic {your_img_path} --celeb_pic {celeb_img_path}

3. MichiGAN

cd michigan
python inference.py --name MichiGAN --gpu_ids 0 --inference_ref_name 60429 --inference_tag_name 56024 --inference_orient_name 60429 --netG spadeb --which_epoch 50 --use_encoder --noise_background --expand_th 5 --load_size 512 --crop_size 512 --add_feat_zeros --data_dir ./datasets/FFHQ_single --expand_tag_mask

---

For pretrained resnet and generator, you can download them here.