Code and Colab Notebook
The code is provided in the form of Jupyter Notebook in Google Colab. Together with model files and Python Pickle data, they are available via Google Drive. I suggest keeping directory and file structure there to rerun everything with minimum efforts.
All the model files (55.6 GB) of VAE generated by TensorFlow have been saved and shared here, so you can reproduce the results without rerunning VAE.
There are two special folders:
Eliminated for the unlucky runs,
Eliminated (Technical) for the runs with some technical issues that prevent the visualisation in Tensorboard.
Python Pickle data
Python Pickle data are provided to reproduce the results quickly. If you would like to process the data from scratch, you need to download the datasets and upload them into Google Drive by yourself.
The parameters used in the study. ("r" means reconstruction space; "l", latent space)