kingofprank / SGF

SGF

Unofficial PyTorch implementation of Surrogate Gradient Field for Latent Space Manipulation

Required

• PyTorch
• dlib
• OpenCV

Pretrained Checkpoints

• StyleGAN2 model (from rosinality/stylegan2-pytorch)
• shape_predictor_68_face_landmarks.dat (link)
• Face Parsing model (from zllrunning/face-parsing.PyTorch)
• Face Attributes Classifier (my checkpoint)
• Auxiliary Mapping (my checkpoint)

Inference

Run python translate_keypoints.py. This script aligns the face keypoints of a randomly generated image with another randomly generated image.

Train on your own dataset

1. Create a Classifier Network on models/classifier.py.
2. Run make_dataset.py to create a dataset by generating images and classifiering their properties.

   python make_dataset.py \
        --size [images size of stylegan2] \
        --g_ckpt [the pretrained checkpoint of stylegan2] \
        --data_dir [where you want to save training data] \
        --n_sample 200000 \
        --batch 8 \
        --truncation 0.8
   

3. Run train.py.

   python train.py \
        --total_iter 500000 \
        --data_dir data \
        --batch 8 \
        --n_layer 15
   

License

StyleGAN2 codes and model are from rosinality/stylegan2-pytorch.

Face-Parsing codes and checkpoint are from Zllrunning/face-parsing.PyTorch.

senet.py is from cydonia999/VGGFace2-pytorch, and the face attributes classifier model was transfer trained from pretrained weights provided by the above repository.