GenerativeSkinLesion

Using progressive gorwing of GANs to augmentation skin lesion dataset.

This project was originated as SFU CMPT-726 (18 Fall) course project. Collaborator: Yongjun He, Ruochen Jiang.

I wrote an article (in Chinese) to discuss some tricks and implementation details. Zhihu article

Be careful to use the code!

• This may not be full implementation of progressive GANs, and we haven't run the code on that same dataset as the original paper.
• Unsolved issue: the training fails when using PyTorch 1.0. I'm still not sure what difference between 1.0 and 0.4 causes this issue...

Dependences

• PyTorch 0.4 with torchvision
• tensorboardX
• Pillow

Data

ISIC 2018 classification task.

Training

python main.py --mode train

The training taks about 4 days on 4 NVIDIA P100 Pascal GPUs.

(After training) Generating

python main.py --mode test --num 1000

Acknowledgement

• tkarras/progressive_growing_of_gans (official Tensorflow implementation)
• akanimax/pro_gan_pytorch
• nashory/pggan-pytorch