Recent work in natural language processing usually requires the training and inference of a large, high dimensional neural network, often a transformer. Despite its efficacy for natural language processing / text processing, the field still has relatively immature methods for illustrating how activations of particular network units influence the output.
This project aims to create a visualiser for neural network architectures, especially the basic transformer. Given training or testing data, we wish to visualise how the data changes the weights of the network over batches of data, over epochs of training and over individual input sentences during inference.