This is the official codebase of the course project report for IFT6269, lead by Jerry Huang, Xinyu Yuan and Le Zhang.
We reproduce the methods presented by Aneja et al. (2021) to train autoencoder priors for high-quality image generation. It is common for variational autoencoders (VAEs) to generate poor images when sampling from the prior without tempering. To tackle this issue, the authors proposed an energy-based prior defined by the product of a base prior distribution and a reweighting factor, to bridge the gap between the base prior and the aggregate posterior.
For our project, we re-implement the training of this re-weighting factor from scratch and reproduce a subset of the results presented in the original paper.
This codebase is based on PyTorch and Pythae. It supports training and inference with CPUs or GPUs.
conda install pytorch torchvision torchaudio torchmetrics cudatoolkit=11.1 -c pytorch-lts -c nvidia -c conda-forge -c pyg
conda install matplotlib-base
conda install numpy
pip install -U scikit-learn
pip install pythae
pip install tqdmThe purpose of the preprocessor is to ensure the data is not corrupted (no nan), reshape
it in case inconsistencies are detected, normalize it and converted it to a format handled by the
pythae.trainers.Trainer. In particular, an input data is converted to a
torch.Tensor and all the data is gather into a pythae.data.datastest.BaseDatset
instance.
By choice, we do not provided very advanced preprocessing functions (such as image registrations) since the augmentation method should be robust to huge differences in the data and be able to reproduce and account for this diversity. More advanced preprocessing is up to the user.
This notebook provides the training of the three stages mentioned in the report on both MNIST and CIFAR10. Some visualization results are also included, say reconstructed images, generated images and interplorations of images.