Variational Autoencoders are a class of generative modelling techniques that aim to learn a good representation of the distribution of the observed data x by imposing an underlying latent space z. By accurately approximating the posterior p(z|x) it aims to maximize the lower bound to p(x).
However, VAEs make assumption that the posterior is approximately factorial which limits the capacity of the model. IWAEs use multiple samples from the approximate posterior to get a tighter lower bound on p(x), which allows you to fit posteriors that can potentially overcome the assumptions VAE lays out.
IWAE has been introduced and explained in the paper Importance Weighted Autoencoders by Yuri Burda, Roger Grosse, and Ruslan Salakhutdinov.
The experiments and results from this work are summarized in this poster (PDF version here)
Note - In the "Conclusion and Insights" section for the poster, the third point of interest is incorrect and was the result of a misinterpretation of the scatter plots in Fig. 3. Note that the scatter plots in Fig. 3 show the distribution of the aggregate posterior Exp(z|x)p(x) and not the posterior conditioned on a particular input. Since the aggregate posterior is pushed closer to the prior by design, Fig. 3 doesnot emphasize on the multimodality of the posterior distributions conditioned on a particular input. A corrected and more clearer plot is coming up soon!
Load the required datasets by running datasets_download.py
tensorflow, tensorflow.python.debug, numpy, progressbar, sys, os .. etc.
This code allows you to train, evaluate and compare VAE and IWAE architectures on the mnist dataset. Additionally you can also plot samples from the posterior from the model. All this can be done by changing the parameters of the main code in experiments.py and running it.