PyTorch implementation of parametric t-SNE.
While the code was initially based on a TensorFlow implementation by jsilter, everything is now implemented from scratch, including PCA pretraining and different metrics for comparison of probability distributions.
This version of parametric t-SNE makes use of pynndescent
to approximate nearest neighbors for an efficient calculation of an approximated distance matrix.
This approach as well as the PCA pretraining are borrowed from the openTSNE package.