This repository is to implement Variational Autoencoder and Conditional Autoencoder. Notebook files for training networks using Google Colab, and evaluating results are provided. Also, trained checkpoints are included.
Variational Autoencoder is a specific type of Autoencoder. In which, the hidden representation (encoded vector) is forced to be a Normal distribution. As the result, by randomly sampling a vector in the Normal distribution, we can generate a new sample, which has the same distribution with the input (of the encoder of the VAE), in other word, the generated sample is realistic.
There are some results of VAE below:
One problem of VAE is generating samples without any conidtions (e.g., labels, ground truths). CVAE is to deal with this issue. The idea is suplementing an additional information (e.g., label, groundtruth) for the network so that it can learn reconstructing samples conditioned by the additional information.
There are some results of CVAE below:
