hyperbolic_wrapped_distribution

Code for reproducing results in "A Wrapped Normal Distribution on Hyperbolic Space for Gradient-Based Learning".

Setup Environment with Docker

$ docker build . -t <image-name>
$ docker run --runtime nvidia --rm -it \
  -v $HOME/data:/root/data \
  -v $PWD:/work -w /work \
  <image-name> bash

Usage

train Hyperbolic VAE with synthetic dataset:

$ python3 -m scripts.train --recipe-path recipes/mlp_synthetic.yml \
  --p-z nagano <experiment-name>

train Hyperbolic VAE with MNIST dataset:

$ python3 -m scripts.train --p-z nagano <experiment-name>

train Hyperbolic VAE with Breakout dataset (you have to place the dataset for explored trajectories of pretrained agent in Breakout to $HOME/data/breakout/state_samples):

$ python3 -m scripts.train --recipe-path recipes/cnn_breakout.yml \
  --p-z nagano <experiment-name>

train Hyperbolic word embedding model with WordNet dataset:

$ python3 -m scripts.train_embedding --p-z nagano <experiment-name>