What would it be like if you appeared as a character of a Kangpool Cartoon?

• Windows server 2019
• Python 3.6, TensorFlow-gpu 1.14
• NVIDIA Tesla T4 * 2 GPUs, NVIDIA drivers, CUDA 10.0 toolkit and cuDNN 7.5.

python projector.py --outdir=~/projection --target=~/projection/target.png \
    --network=~/model/network-snapshot-001476.pkl

Datasets are stored as multi-resolution TFRecords, i.e., the same format used by StyleGAN and StyleGAN2. Each dataset consists of multiple *.tfrecords files stored under a common directory, e.g., ~/datasets/ffhq-r*.tfrecords FFHQ datasets for Transfer Learning: Download the Flickr-Faces-HQ dataset as TFRecords: Kangpool datasets:

python dataset_tool.py create_from_images ~/datasets/kangponFace ~/kangpoolFace_png
python dataset_tool.py display datasets/kangpoolFace

python train.py --outdir=~/output --gpus=2 --data=~/datasets/kangpoolFace \
    --mirror=1 --aug=ada --resume=ffhq256 --snap=10 --freezed=12 --metric=fid50k --metricdata=~/datasets

Traing Hyper parameters Arguments

• --mirror=1 amplifies the dataset with x-flips. Often beneficial, even with ADA.
• --aug=ada enables ADA (default: enabled).
• --resume=ffhq256 performs transfer learning from FFHQ trained at 256x256.
• --resume=~/models/<RUN_NAME>/network-snapshot-<KIMG>.pkl resumes where a previous training run left off.
• --freezed=12 uses FreezeD (default: 0). discriminator layers
• --snap=10 to export network pickles more frequently than usual. This is recommended, because transfer learning tends to yield very fast convergence.
• --metrics=fid50k to evaluate FID the same way as in the StyleGAN2 paper (see below).
• --metricdata to evaluate quality metrics against the original FFHQ dataset, not the artificially limited 10k subset used for training.

• Epoch = 0 Pre-trained (FFHQ256) Layer. It will be a guide-face to map the Cartoon face
• Epoches = 800 Not bad. But glasses are represented weird.
• Epoches = 1600 Good. But some images are dimmed.

pip install streamlit
streamlit run main.py

Copyright © 2020, NVIDIA Corporation. All rights reserved. This work is made available under the Nvidia Source Code License. Referenced from StyleGAN2