/pytorch.ALICE.toydata

An unofficial PyTorch implementation of ALICE (NIPS2017) on GMM toy dataset.

Primary LanguagePython

OFFICIAL REPOSITORY IS HERE.
This is not official and does not reproduce/implement all the experiments and results.

ALICE

Adversarially Learned Inference with Conditional Entropy.

This repository has 2 kinds of experiments on GMM dataset.

Requirements

  • Python 3.6.5
  • matplotlib
  • pytorch 0.4.1
  • tqdm (progress bar)

How to use

  1. python3 train_ALICE_toydata.py: Train using explicit cyclic consistency
  2. python3 train_ALICE_toydata.py --adv: Train using implicit cyclic consistency

--easy option reduces the number of modes in dataset X.

Results and the used dataset saved under args.results_dir/{YYMMDD-HMS}ALICE_unsupervised_{MSE or adversarial}_reconstruction

Datasets are saved [x, z]_trn.pkl using torch.save method.

  • Default dataset:
    figure1
  • --easy:
    figure2

Results

Some results are under results.

Explicit Cyclic Consistency

figure3

Implicit Cyclic Consistency

I emprically confirmed --n_dis had some effects on results though, training in this setting was not stable in general. figure4

What I implement

  • Experiments on toy-dataset (GMM)