Pose-based-Conditional-BEGAN

A pytorch implementation of paper "GAN-Based Realistic Face Pose Synthesis with Continuous Latent Code", https://aaai.org/ocs/index.php/FLAIRS/FLAIRS18/paper/view/17643

It uses modified BEGAN to generate faces conditioned by face poses.

Requirements

Enviroment

• Python 3.6
• PyTorch 1.0.0

And common libraries such as numpy and matplotlib.

Dataset

Download in the wild CelebA dataset and uncompressed it to have a directory structure like below.

data_dir
├── img_celeba
|   ├── 000001.jpg
|   ├── 000002.jpg
|   └── ...
└── Anno
    ├── list_bbox_celeba.txt
    └── list_landmarks_align_celeba.txt

Usage

Train

To train a model use default parameters:

$ python train.py --data_dir path/to/data_dir

Inference

A pretrained model of generator is included in checkpoints dir, to use it generate faces:

$ python inference.py --checkpoint checkpoints/checkpoint_final_G

Results

• Zero pose faces generated randomly.
• Evenly spaced faces, fixed noise z for each row. For first 4 rows, the yaw varies evenly from -90 to 90. For next 2 rows, the roll varies evenly from -90 to 90. For last 2 rows, the pitch varies evenly from -90 to 90.

Note

• I didn't try image size with 128, nor do I have enough resources to train one. It should be easy to modify the code to train a 128x128 model, just be careful with the draw debug image part, there may be some hard-code for image size 64.
• In the paper I got, the learning rate was set to 8e-4, I failed to find converge with that learning rate. The results are given by a learning rate of 8e-5 and it worked like a charm.
• In the paper I got, the beta2 for Adam was set to 0.99, and I modified it to 0.999. Have no idea which one is right.
• In the paper, the auther flip the images and add it to dataset. While I just filp at random, and increase training epoch from 20 to 40.

Author

Li Yizhuo / @liyz15