Research code for ICML 2019 paper on "Hierarchical Decompositional Mixtures of Variational Autoencoders"
- tensorflow > 1.8
Main Idea of Paper
Variational autoencoders (VAEs) have received considerable attention, since they allow us to learn expressive neural density estimators effectively and efficiently. However, learning and inference in VAEs is still problematic due to the sensitive interplay between the generative model and the inference network. Since these problems become generally more severe in high dimensions, we propose a novel hierarchical mixture model over low-dimensional VAE experts. Our model decomposes the overall learning problem into many smaller problems, which are coordinated by the hierarchical mixture, represented by a sum-product network. In experiments we show that our model consistently outperforms classical VAEs on all of our experimental benchmarks. Moreover, we show that our model is highly data efficient and degrades very gracefully in extremely low data regimes.
Download from Github
.npy files are stored using Git LFS.
Before cloning this directory, please install Git LFS according to the instructions
on this link.
If you are using an ubuntu system like me, a convenience script
install_gitlfs_ubuntu.sh is provided.
Without Git LFS installed,
.npy files will still be downloaded
but will not be interpretable by numpy.
If Git LFS cannot be installed, the dataset can still be obtained by running one of the
srcis where the low level logic of SPN and VAEs are implemented.
datahouses the three different datasets used in the paper in a consistent format
scriptscontains blueprints for how different components are joined together
mainis where models are run and analysed
hptune*.py hyperparameter tuning scripts that are within the
The results will be saved within the same directory and can be compiled using the
Use of GPUs is highly recommended. Otherwise the models will take too long to train. Expected duration of hyperparameter tuning on a single GPU: 1 week