This is a collection of Jupyter notebooks for image generation topics written in pytorch. This git currently includes notebooks on
The level of explanation provided between notebooks varies. Links are included within the notebooks to sources for those interested in learning more. The notebooks are listed in reverse chronological order. The pictures below are generated by the model after training. They all feature MNIST images at the moment for comparative purposes.
Note: if the notebooks below won't load on github, you can load them using the Jupyter nb viewer or run them locally instead.
If you want to run these notebooks on your own computer, then the easiest way is to install conda, and use the following commands after cloning this repository:
conda env create -f environment.yaml
conda activate pytorch
jupyter notebook
This will open a browser window that can be used to load the notebook files.
Note: The environment created will be quite large. It will take some time to create.
Also: If you are not going to run this on a GPU, then add the cpuonly package to the environment.yaml file. This will reduce the environment size as
it will skip installing the CUDA libraries.
The Lagrangian VAE is a Variational Autoencoder that optimizes the Lagrangian dual equation to provide both consistency and meaningful encodings.
Lagrangian VAE Notebook (draft)
Image reconstruction of MNIST. It's worth noting how much sharper these images are as opposed to the basic VAE, even after interjecting noise.
Image reconstruction using sampling of latent space. These images are not as sharp, but I don't know how to fix that.
Distribution of encoded means. While not conclusive, this seems to suggest that the model learns to group similar observations near eachother.
Masked Autoencoders for Distribution Estimation (MADE) is an autoencoder that uses masks to provide an autoregressive property. The use of multiple masks effectively give us one set of parameters that can be used to form an ensemble.
Image reconstruction of MNIST using an ensemble of masks.
A Variational Autoencoder is a type of deep generative model. Below we can see some example image reconstructions.
Image Reconstructions of MNIST
Image Generation using sampling of latent space
