FakeFace

Introduction

This repository holds the PyTorch-based code for training a fake (face) image classifier.

We used this code to participate in the 2018 AI RnD Challenge, hosted by the Ministry of Science and ICT, Government of South Korea. Our team(Lomin) won the 2nd place of the challenge.

In this repository, we provide

• Training/Test code
• Our 'Photoshop dataset', which we specially made for this task
• Demo website for the test: visit https://faceforensic.lomin.ai

Dependencies

• Python packages (We recommend using Conda environment)
  • tensorflow-gpu==1.12.0
  • torch==1.0.0
  • torchvision==0.2.1
  • Pillow==5.2.0
  • matplotlib==2.2.2
  • scipy==1.1.0
  • tqdm==4.24.0
  • opencv-python==3.4.3.18
  • pyyaml==4.2b1
  • addict==2.2.0
  • tensorboardX==1.4
  • mtcnn==0.0.8
• An NVIDIA GPU & CUDA 9.0

Quick start

• To test our models, visit our demo website: https://faceforensic.lomin.ai (Korean)
• If you want to manually test and get scores for each images, follow the steps:

  1. Place your images at any directory.

  2. Download our models from this link.

  3. Run the script "test.py". Use at least one of —inference or —avg option.

     python test.py --root_dset (root directory for your test images) --root_model (root directory where you downloaded models) --batch_size (batch size for inference) --inference (if you want to run inference) --avg (if you want to ensemble results)

Data preparation

Photoshop dataset: In order to train a more accurate classifier, we created 3,000 synthetic face images with the help of image editing experts. You can download it from this link.

Other images used for real/fake samples are from public datasets. We use

• UMDFaces
• CelebA-HQ
• PGGAN

Locate datasets (or create softlinks to directories) under fakeface/dataset.

Training

git clone https://github.com/lomin-ai/fakeface.git

python train.py
	--tag         (tag, or identifier of training)
	--preset      (choose one of 'gan', 'syn', 'mod', 'gan+syn')
	--clear_cache (do not use cache for dataset)
	--set         (additional settings)

Contact

Bee Lim ([email protected])