TOWARDS BUILDING A GROUP-BASED UNSUPERVISED REPRESENTATION DISENTANGLEMENT FRAMEWORK

TOWARDS BUILDING A GROUP-BASED UNSUPERVISED REPRESENTATION DISENTANGLEMENT FRAMEWORK
Tao Yang, Xuanchi Ren, Yuwang Wang, Wenjun Zeng, Nanning Zheng
arXiv preprint arXiv:2102.10303
ICLR 2022

[Paper] [Arxiv] [Appendix]

Update:

🔲 Release code

Discription:

In this repo, built on the group-based definition and inspired by the n-th dihedral group, we first propose a theoretical framework towards achieving unsupervised representation disentanglement. We then propose a model, based on existing VAEbased methods, to tackle the unsupervised learning problem of the framework.The overview of our method is as follows:

Implementation details

• The Isomorphism Loss in our paper is implemented in "model/share.py"
• The model constraint in our paper is implemented in the function "forward" in "model/XXX_VAE.py"

Requirements

• build a docker environment by using "./dockerfile"

Models

The VAE-based models:

• $\beta$-VAE
• AnnealVAE
• FactorVAE
• $\beta$-TCVAE

Datasets

The evaluation datasets:

• dSprites
• Shapes3D
• Car3D
• noisy dSprites
• color dSprites

Download them to "./dataset_folder/"

Usage

The Arguements

usage: main.py [-h] --config_num CONFIG_NUM [--eval]

optional arguments:
-h, --help            show this help message and exit
--config_num CONFIG_NUM
                        the number of settings of hyperparameters and random
                        seeds
--eval                eval model or not (default: False)

Config File

The hyperparamter and random seed settings are numbered in './config.csv'. Every row in './config.csv' correspond to a setting.

Examples:

Training

Train the model under the 1000th setting:

python main.py --config_num = 1000

Evaluation

Evaluate the model we trained above:

python main.py --config_num = 1000 --eval

Qualitative evaluation

dSprites
BetaVAE	DCI
MIG	FactorVAE

NOTE: Groupified VAE achieves better performance mean with lower variance.

Latent space visualization

dSprites AnnealVAE
C_max = 10, Original	C_max = 20, Original
C_max = 10, Groupified	C_max = 20, Groupified

Qualitative results

Cars3d and Shapes3d
Original	Groupified

Cyclic latent space
Original
Groupified

Controllable meaningful dimensions in Groupified VAEs

dSprites Anneal VAE
KL divergence of dimensions	Traversal results

Downstream Task Performance

Citation

@article{Tao2022groupified,
  title   = {TOWARDS BUILDING A GROUP-BASED UNSUPERVISED REPRESENTATION DISENTANGLEMENT FRAMEWORK},
  author  = {Tao, Yang and Xuanchi, Ren and Yuwang, Wang and Wenjun, Zeng and Nanning, Zheng and Pengju,Ren},
  journal = {ICLR},
  year    = {2022}
}