BigGAN-AM

This code release accompanies the following paper:

A cost-effective method for improving and re-purposing large, pre-trained GANs by fine-tuning their class-embeddings [Video] [arXiv]

Qi Li, Long Mai, Michael A. Alcorn, and Anh Nguyen. Asian Conference on Computer Vision (ACCV). 2020. Oral presentation. :star: Huawei Best Application Paper Honorable Mention. :star:

Setup

Download the pre-trained BigGAN model weights (see here for an explanation of the commands):

cd biggan-am/pretrained_weights

# 128x128 weights.
fileid="1nAle7FCVFZdix2--ks0r5JBkFnKw8ctW"
curl -c ./cookie -s -L "https://drive.google.com/uc?export=download&id=${fileid}" > /dev/null
curl -Lb ./cookie "https://drive.google.com/uc?export=download&confirm=`awk '/download/ {print $NF}' ./cookie`&id=${fileid}" -o biggan_weights.zip
unzip biggan_weights.zip
rm biggan_weights.zip

# 256x256 weights.
fileid="1FEAXaUjRcV8mb0sHIwuS2EFyAcAITDTw"
curl -c ./cookie -s -L "https://drive.google.com/uc?export=download&id=${fileid}" > /dev/null
curl -Lb ./cookie "https://drive.google.com/uc?export=download&confirm=`awk '/download/ {print $NF}' ./cookie`&id=${fileid}" -o biggan_256_weights.pth

Install the Python packages (we used Python 3.6):

cd biggan-am
pip3 install -r requirements.txt

Alternatively, you can use our provided Dockerfile to run the model in a container:

cd biggan-am
docker build -t biggan-am:v1 .
nvidia-docker run \
    -a stdin \
    -a stdout \
    -it \
    --name biggan-am \
    --mount type=bind,source=$(pwd),target=/biggan-am \
    biggan-am:v1 \
    /bin/bash

Running BigGAN-AM

Edit opts.yaml:

cd biggan-am
nano opts.yaml

Run BigGAN-AM:

python3 biggan_am.py

Generating samples from a BigGAN-AM optimized class embedding

cd biggan-am
python3 sample_biggan_am_embedding.py [directory] [target]

e.g.:

python3 sample_biggan_am_embedding.py "places" "266_pier"

Citation

If you find this work useful for your research, please consider citing:

@article{li2020improving,
  title={A cost-effective method for improving and re-purposing large, pre-trained GANs by fine-tuning their class-embeddings},
  author={Li, Qi and Mai, Long and Alcorn, Michael A. and Nguyen, Anh},
  journal={Asian Conference on Computer Vision},
  year={2020}
}

Acknowledgments

This work is supported by the National Science Foundation under Grant No. 1850117 and a donation from Adobe Inc.

Thanks to BigGAN-PyTorch by Andy Brock

robustness package by MadryLab

Places365-CNNs by MIT CSAIL

anguyen8 / biggan-am