β-VAE to train from frozen or untrained weights

This is a fork of the implementation by Konny1

Dependencies

python 3.6.4
pytorch 0.3.1.post2
visdom

Datasets

[same with here]

Usage

initialize visdom

python -m visdom.server

you can reproduce results below by

sh run_celeba_H_beta10_z10.sh
sh run_celeba_H_beta10_z32.sh
sh run_3dchairs_H_beta4_z10.sh
sh run_3dchairs_H_beta4_z16.sh
sh run_dsprites_B_gamma100_z10.sh

or you can run your own experiments by setting parameters manually.
for objective and model arguments, you have two options H and B indicating methods proposed in Higgins et al. and Burgess et al., respectively.
arguments --C_max and --C_stop_iter should be set when --objective B. for further details, please refer to Burgess et al.

e.g.
python main.py --dataset 3DChairs --beta 4 --lr 1e-4 --z_dim 10 --objective H --model H --max_iter 1e6 ...
python main.py --dataset dsprites --gamma 1000 --C_max 25 --C_stop_iter 1e5 --lr 5e-4 --z_dim 10 --objective B --model B --max_iter 1e6 ...

check training process on the visdom server

localhost:8097

Reference

1. [β-VAE: Learning Basic Visual Concepts with a Constrained Variational Framework, Higgins et al., ICLR, 2017]
2. [Understanding disentangling in β-VAE, Burgess et al., arxiv:1804.03599, 2018]
3. Repo from Konny1