cutz-j / T-GD

T-GD: Transferable GAN-generated Images Detection Framework. (ICML 2020)

T-GD: Transferable GAN-generated Images Detection Framework.

Hyeonseong Jeon, Youngoh Bang, Junyaup Kim, and Simon S. Woo. "T-GD: Transferable GAN-generated Images Detection Framework.". Thirty-seventh International Conference on Machine Learning (ICML). 2020.

Overview of our framework.

Clone

git clone https://github.com/cutz-j/T-GD

Dataset

The dataset for each result condition can be downloaded by running the file in dataset or here.

CelebA, CelebA-HQ, FFHQ, LSUN-Bedroom, LSUN-Bird.

PGGAN_CelebA, StarGAN, StyleGAN1, StyleGAN2, PGGAN_Bedroom, PGGAN_Bird.

A example script for downloading the testset is as follows:

# Download the dataset
cd dataset
bash download_PGGAN.sh
bash download_StarGAN.sh
bash download_StyleGAN.sh
bash download_StyleGAN2.sh
bash download_PGGAN_lsun_bedroom.sh
bash download_PGGAN_lsun_bird.sh
cd ..

For the PGGAN dataset, we have contacted with the dataset provider about opening the dataset. For now we only uploaded data for the test set. PGGAN-images from LSUN-bedroom and LSUN-bird will be added to dataset.

Our datasets are heavily from here.

Download pre-trained model weights

The pretrained weights can be downloaded by running the file in dataset or here.

# Download the pre-trained weights
cd weights
bash download_weights.sh
cd ..

Setup

pip install -r requirements.txt

• warmup scheduler

pip install git+https://github.com/ildoonet/pytorch-gradual-warmup-lr.git

Evaluation

# Dataset and model weights need to be downloaded.
# source and target dataset dir. i.e., StarGAN --> StyleGAN2
# pretrained weight. i.e., efficientnet/stargan.pth.tar
# t-gd pretrained weight. i.e., t-gd/efficientnet/star_to_style2.pth.tar
python eval.py --source_dataset dataset/StarGAN_128 \
                --target_dataset dataset/StyleGAN2_256 \
                --pretrained_dir weights/pre-train/efficientnet/stargan.pth.tar \
                --resume weights/t-gd/efficientnet/star_to_style2.pth.tar

Pre-train

# Dataset needs to be downloaded.
# source dataset

python pretrain.py --source_dataset dataset/StarGAN_128

Transfer

# Dataset and model weights are needed.
# source and target dataset dir. i.e., StarGAN --> StyleGAN2
# pretrained weight. i.e., efficientnet/stargan.pth.tar

python transfer.py --target style2
                --source_dataset dataset/StarGAN_128 \
                --target_dataset dataset/StyleGAN2_256 \
                --pretrained_dir weights/pre-train/efficientnet/stargan.pth.tar