StyleGAN2_tf2pytorch

Convert checkpoint from offical tensorflow StyleGAN2 to pytorch StyleGAN2

Requirements

• Python 3.6

• Pytorch >= 1.4.0.

• Tensorflow-gpu==1.14.0 (Offical StyleGAN2 recommended)

• CUDA 10.0 needed to compile some ops from dnnlib

Usage

python tf2pytorch.py --tf_cp *tensorflow checkpoint path* 
                     --output_path *output pytorch checkpoint path*

You can download pytorch checkpoint from google drive provided by editGAN and tensorflow checkpoint (stylegan2-car-config-f.pkl) from google drive provided by StyleGAN2, and compare the weight diff (should be zero) between official tensorflow checkpoint and pytorch checkpoint.

python tf2pytorch.py --compare true

To load the converted checkpoint, you can

import pickle
with open(*checkpoint path*, 'rb') as f
    checkpoint_data = pickle.load(f)
net.load_state_dict(checkpoint_data, strict=False)