Quantitative Evaluation of Disentangled Representations

Code to reproduce the results in our ICLR 2018 paper: A Framework for the Quantitative Evaluation of Disentangled Representations.

Prerequisites

• Python 2.7.5+/3.5+, NumPy, TensorFlow 1.0+, SciPy, Matplotlib, Scikit-learn

Data

• Download here.
  • If RAM < 10GB, convert .npz to .jpeg before training to load batches of images into memory (rather than entire dataset)
    • python npz_to_jpeg.py (after editing paths)
• Generated using this renderer.

Models

• PCA
• InfoGAN
• VAE / \beta-VAE

Train

• PYTHONPATH=[/path/to/qedr/] python main.py

Save codes

• PYTHONPATH=[/path/to/qedr/] python main.py --save_codes

Quantitative Evaluation

• quantify.ipynb