Skip to content

Latest commit

 

History

History
22 lines (16 loc) · 1.26 KB

README.md

File metadata and controls

22 lines (16 loc) · 1.26 KB
Raw

CollisionGAN

Collision image translation

Dear GC3D teammates,

For a simplied CycleGAN implementation, there are only three places you need to look at

  1. cycleGan.ipynb that has the implementaion of CycleGAN class, data handling, training options, and training;
  2. models/base_model.py (base model for CycleGAN), models/resnet.py (for generator), models/alexnet.py (for discriminator) I didn't use the original discriminator because it is sensitive to dimension, and Dmitrii's AlexNet works very well.
  3. utils/network.py that has a few utility functions.

I hack the original CycleGAN code and make it simply because

  1. I want to force myself to look as close as possible into it;
  2. The original implementation is very comprehensive but some parts of it is not very necessary for us and distractive;
  3. The simplification help me to compartmentalize my understanding of the building blocks of CycleGAN.

CLAIM: I haven't give my simplification a thorough check. So please don't trust it.

TO-DO:

  1. I will use the PIL Python package to convert our npz files to images, and run the original CycleGAN in order to check whether my simplification works properly; I will post the image dataset when it is done.
  2. Use paired test dataset to see how the translation works.