LORD

Demystifying Inter-Class Disentanglement
Aviv Gabbay and Yedid Hoshen
International Conference on Learning Representations (ICLR), 2020.
Official tensorflow implementation [Pytorch re-implementation]

Content transfer between classes

Cars3D	SmallNorb	KTH

CelebA

Usage

Dependencies

• python >= 3.6
• numpy >= 1.15.4
• tensorflow-gpu >= 1.12.0
• keras >= 2.2.4
• keras-lr-multiplier >= 0.7.0
• opencv >= 3.4.4
• dlib >= 19.17.0

Getting started

Training a model for disentanglement requires several steps.

Preprocessing an image dataset

Preprocessing a local copy of one of the supported datasets can be done as follows:

lord.py --base-dir <output-root-dir> preprocess
    --dataset-id {mnist,smallnorb,cars3d,shapes3d,celeba,kth,rafd}
    --dataset-path <input-dataset-path>
    --data-name <output-data-filename>

Splitting a preprocessed dataset into train and test sets can be done according to one of two configurations:

lord.py --base-dir <output-root-dir> split-classes
    --input-data-name <input-data-filename>
    --train-data-name <output-train-data-filename>
    --test-data-name <output-test-data-filename>
    --num-test-classes <number-of-random-test-classes>

lord.py --base-dir <output-root-dir> split-samples
    --input-data-name <input-data-filename>
    --train-data-name <output-train-data-filename>
    --test-data-name <output-test-data-filename>
    --test-split <ratio-of-random-test-samples>

Training a model

Given a preprocessed train set, training a model with latent optimization (first stage) can be done as follows:

lord.py --base-dir <output-root-dir> train
    --data-name <input-preprocessed-data-filename>
    --model-name <output-model-name>

Training encoders for amortized inference (second stage) can be done as follows:

lord.py --base-dir <output-root-dir> train-encoders
    --data-name <input-preprocessed-data-filename>
    --model-name <input-model-name>

Citing

If you find this project useful for your research, please cite

@inproceedings{gabbay2020lord,
  author    = {Aviv Gabbay and Yedid Hoshen},
  title     = {Demystifying Inter-Class Disentanglement},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2020}
}