This is a re-implementation of Auto-Encoding Variational Bayes in MATLAB.
I use the MNIST from: https://github.com/y0ast/VAE-Torch/tree/master/datasets.
Please install my fork of MatConvNet, where I implemented some new layers, including:
KLD.m
: handles forward and backward propagation of KL DivergenceNLL.m
: handles forward and backward propagation of Negative Log-Likelihood (works for multi-variate Bernoulli distribution)LB.m
: combine KLD and NLL into a lower boundSampler.m
: sampling operationTanh.m
: tanh non-linearitySplit.m
: split one variable into multiple while keeping the same spatial size
For training, please see train_script.m
on how I trained models. I
implemented four stochastic gradient descent algorithms:
- SGD with momentum
- ADAM
- ADAGRAD
- RMSPROP
For demo, I have four demo scripts for visualization under demo/
,
which are:
manifold_demo.m
: visualize the manifold of a 2d latent space in image space.sample_demo.m
: sample from latent space and visualize in image space.reconstruct_demo.m
: visualize a reconstructed version of an input image.walk_demo.m
: randomly sample a list of images, and compare the morphing process done in both image space and latent space.
To learn about how VAE works under the hood, refer to the original paper or my writeup.