Personal Projects

Here is a summary of a couple of personal projects available on my github:

• A reproducibility study of an ICLR 2020 paper about an energy-based approach to training hybrid (simultaneously generative and discriminative) deep learning models, and submitted a report to the 2020 ML reproducibility challenge. The reprocibility study code can be found here.

• An implementation of the original transformer architecture from Vaswani et al for machine translation (German-English). Architecture implemented from scratch, data loaded using torchtext. Code available here.

• A fastAPI deployment of a named entity recognition model (backend BERT fined tuned on this dataset) on an AWS EC2 instance, code for finetuning and deploying can be found here. See deployment tab in side panel for a tutorial on bare-bones web-deployment using AWS.

• A two-latent layer ladder variational autoencoder, implemented in PyTorch. This is an extension of a homework assignment for Stefano Ermon's Deep Generative Models course. The original sourcecode is written using Theano, Lasagne and Parmesan and is available here. See also this paper on training ladder variational autoencoders.