https://github.com/karenullrich/discretecatchup
Experiments for continuous to binary and back
https://github.com/karenullrich/discretecatchup
Last synced: about 2 months ago
JSON representation
Experiments for continuous to binary and back
- Host: GitHub
- URL: https://github.com/karenullrich/discretecatchup
- Owner: KarenUllrich
- Created: 2020-02-01T15:19:40.000Z (over 5 years ago)
- Default Branch: master
- Last Pushed: 2020-02-01T15:52:39.000Z (over 5 years ago)
- Last Synced: 2025-02-07T18:51:16.257Z (4 months ago)
- Language: Python
- Size: 6.84 KB
- Stars: 0
- Watchers: 4
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
## From NormalVAE to BinaryVAE with intermediate steps
## for Emiel
### Short note on the implemented models
* NVAE: vanilla VAE with Gaussian Normal distribution
python experiment.py --latent_dist normal
* BNVAE: vanilla VAE, gets exactly same results but is implemented with
explicit Bernoulli samples this is more like a sanity check
python experiment.py --latent_dist bnormal
* BVAE: BinConcrete VAE as implemented in \[Maddison, 2017\]
python experiment.py --latent_dist binary
* BCVAE: BinConcrete latent distribution, but samples are turned into
continious samples
python experiment.py --latent_dist bcontinuous
### Requirements
This code has been tested with
- `python 3.6`
- `tensorflow 2.1.0`
- `tensorflow-probability 0.9.0`
- `matplotlib 3.1.2`
Install conda environment via
conda env create -f environment.yml
source activate binary_vae
### Approximate Bernoulli implementation
Tensorflow implmentation of Bininary Concrete (BinConcrete) latent distribution, based on:
["The concrete distribution: A continuous relaxation of discrete random variables"](https://arxiv.org/pdf/1611.00712.pdf)
Maddison, Chris J., Andriy Mnih, and Yee Whye Teh, ICLR, 2017