timbmg/vae-cvae-mnist

Variational Autoencoder and Conditional Variational Autoencoder on MNIST in PyTorch

Variational Autoencoder & Conditional Variational Autoenoder on MNIST

VAE paper: Auto-Encoding Variational Bayes

CVAE paper: Semi-supervised Learning with Deep Generative Models

---

In order to run conditional variational autoencoder, add --conditional to the the command. Check out the other commandline options in the code for hyperparameter settings (like learning rate, batch size, encoder/decoder layer depth and size).

---

Results

All plots obtained after 10 epochs of training. Hyperparameters accordning to default settings in the code; not tuned.

z ~ q(z|x) and q(z|x,c)

The modeled latent distribution after 10 epochs and 100 samples per digit.

VAE	CVAE

p(x|z) and p(x|z,c)

Randomly sampled z, and their output. For CVAE, each c has been given as input once.

VAE	CVAE

vae-cvae-mnist's Issues

normal distribution sampling error

I am afraid there exists a bug in line 42 of models.py
eps = torch.randn_like(std)
torch.randn_like sample from uniform distribution, should have been normal distribution

Version of Pytorch used.

Can you please reveal the Pytorch version used to code CVAE?

About the KLD loss term

In the original paper， loss=recon_loss - KLD. But in your code, loss=recon_loss + KLD. I aslo found your code work in both cases. Can you please tell the idea about the loss in your code?

Hello! About the latent space figure

I also evaluated the similar phenomena about the Diffuse latent space by CVAE, It is obviously different from the VAE model. Do you know which published paper researched on it or had similar figures with yours?

It seems doesn't work well.

I use the default params but it doesn't look structured. So I change params: more latent size and more layer, but it looks the same.