/betaVAE

TF beta-VAE

Primary LanguagePython

beta-VAE

Latent traversal

Latent traversal replication using the 2D shapes VAE architecture from A.2 Table 1 of beta-VAE by Higgins et al. (https://openreview.net/references/pdf?id=Sy2fzU9gl).

Results after 40 epochs:

Note that this code is older, and has not been updated to the tf.Dataset API.

Understanding disentangling

Attempt at reproducing the Figure 3 (top left) of Understanding disentangling in $\beta$-VAE by Burgess et al. (https://arxiv.org/pdf/1804.03599.pdf). The preset hyperparameters definitely lead to an increase in latent factor capacity in stages as presented in the paper (i.e. first towards x and y, then scale etc.), though not at the same rate.