Skip to content

README.md

Latest commit

 

History

History
99 lines (80 loc) · 4.19 KB

README.md

File metadata and controls

99 lines (80 loc) · 4.19 KB
Raw

Private-Encoder: Enforcing Privacy in Latent Space for Human Face Images

Description

Official pytorch impementation of our private-encoder paper for both training and inferencing.

Usage

Preparation

  • Install required packages
wget https://github.com/ninja-build/ninja/releases/download/v1.8.2/ninja-linux.zip
sudo unzip ninja-linux.zip -d /usr/local/bin/
sudo update-alternatives --install /usr/bin/ninja ninja /usr/local/bin/ninja 1 --force
  • Download required pretrained models

The auxiliary models which are needed for training your own model from scratch or inference.

Path Description
FFHQ StyleGAN StyleGAN model pretrained on FFHQ taken from rosinality with 1024x1024 output resolution.
IR-SE50 Model Pretrained Arcface for use in our ID loss during training.
MTCNN Weights for MTCNN model taken from TreB1eN for use in ID similarity metric computation. (Unpack the tar.gz to extract the 3 model weights.)

Please download and save these models to the directory pretrained_models.

Training

  • The main training script can be found in scripts/train.py.
  • Before training, please define your data path at --train_source_root and --test_source_root.
  • The training results are saved to exp_dir=./experiment, so you can check the checkpoints, tensorboard logs, intermediate train output and test outputs in that directory.
  • To acquire different anonymization effects, you can adjust --target_id.
python scripts/train.py \
--exp_dir=./experiment \
--dataset_type=celeba_encode \
--checkpoint_path=./pretrained_models/psp_ffhq_encode.pt \
--train_source_root=/content/drive/MyDrive/Image_datasets/celeba_align_20k_21k_aligned \
--test_source_root=/content/drive/MyDrive/Image_datasets/celeba_align_20k_21k_aligned \
--workers=8 \
--batch_size=4 \
--test_workers=8 \
--test_batch_size=4 \
--encoder_type=GradualStyleEncoder \
--start_from_latent_avg \
--lpips_lambda=0.8 \
--id_lambda=0.1 \
--l2_lambda=1.0 \
--max_steps=100000 \
--image_interval=5000 \
--board_interval=50 \
--val_interval=1000 \
--save_interval=5000 \
--target_id=0.0

You can also directly run training colab notebook example in Open In Colab

Inference

  • Pretrained Models

Please download the pre-trained Private-Encoder models from the following links.

Path Description
Pretrained Models for FFHQ trained with the FFHQ dataset for different level of identity anonymization.
  • If you wish to use one of the pretrained models for training or inference, please change the directory of --checkpoint_path.
  • Before inference, please define your data path at --data_path.
python scripts/inference.py \
--exp_dir=./image_datasets \
--checkpoint_path=./best_models/privacy_0.0/best_model.pt \
--data_path=./image_datasets/ffhq5k_6k_related/ffhq5k_6k_aligned \
--test_batch_size=5 \
--test_workers=4

You can also directly run inference colab notebook example in Open In Colab

Citation

If you use this code for your research, please cite our paper Private-Encoder: Enforcing Privacy in Latent Space for Human Face Images:

@article{zhao2021private,
  title={Private-encoder: Enforcing privacy in latent space for human face images},
  author={Zhao, Yuan and Liu, Bo and Zhu, Tianqing and Ding, Ming and Zhou, Wanlei},
  journal={Concurrency and Computation: Practice and Experience},
  pages={e6548},
  year={2021},
  publisher={Wiley Online Library}
}