Model implementation for Two-Step Disentanglement for Financial Data by Naama Hadad, Lior Wolf and Moni Shahar from Tel Aviv University. This work address the problem of disentanglement of factors that generate a given data into those that are correlated with the labeling and those that are not. The model employs adversarial training in a straightforward manner.
Jupyter notebook, Python 3.5, numpy, pytorch 0.4, Matplotlib is also used to plot results
Current results:
- Implement S encoder + S classifier and achieve 99%+ accuracy in MNIST data set [Done.But only get 97% accuracy]
- Implement Decoder and Adversarial net and whole net on MNIST[Done. But unstable]
- Test it on Sprites data set [Done (see the results explanation in betterResults.pptx)]
- Extend the method to 3D point cloud [TO DO]