Variational Inference with Normalizing Flows

Installation

git clone https://github.com/vrvrv/Variational-Inference-with-Normalizing-Flows.git
cd Variational-Inference-with-Normalizing-Flows

# pip install -r requirements.txt

Train

You can find configuration files at configs/experiment/. In our code, wandb is the default logger. So, before running code, please sign up wandb.

Training

If you want to control the number of hidden dimension, add model.D=<hidden_dim>.

python train.py experiment=mnist_nfvae model.D=10

Experimental Result

We compared Vanila-VAE and NFVAE on several different settings.

MNIST dataset

We sampled from the model by taking decoder after generating gaussian noise.

VAE

NFVAE

Model	Negative log-likelihood
VAE	~ 110.98
NFVAE (D=10)	~ 109.73

1D Simulation

Coupla Dataset

References

• Variational Inference with Normalizing Flows, ICML 2015