Variational AutoEncoder
- This is a tensorflow implementation for paper Auto-Encoding Variational Bayes, the basic version come from hwalsuklee/tensorflow-mnist-CVAE
- Experiment for different dimensions of z: 2,4,8,10,20,100
- High light results:
- sampling manifold and mappings of train and test data(for mnist and dimension 2)
- reconstruction results among all dimensions(that means a mapping from x to x_hat)
- Something interesting: plot sum loss, binary cross entropy loss(BCE) and kl-divergence(KLD). To my surprised, the results of BCE and KLD are logical, but the dim20 get the lowest sum loss, not the dim100.
Visualizations of data manifold. Note: in manifold.png the range of z is [-2,2]
The distribution of labeled data(including train and test).
| train mappings | test mappings |
| 
|
| Sum Loss | BCE | KLD |
| 
| 
|
Reconstruction for images. Dimension is a very import factor for reconstruction.
| 2-Dim latent space | 4-Dim latent space | 8-Dim latent space |
| 
| 
|
| 10-Dim latent space | 20-Dim latent space | 100-Dim latent space |
| 
| 
|
[1] hwalsuklee/tensorflow-mnist-VAE
[2] kvfrans/variational-autoencoder