Open-source projects
Libraries and frameworks built outside of papers, with the Intelligent Systems group at MIPT and on my own.
Just Relax It: discrete variable relaxation
PyTorch library of relaxation methods for discrete random variables in neural networks. Gumbel-Softmax Top-K, Hard Concrete, Straight-Through Bernoulli, Invertible Gaussian, REBAR and more, behind a Pyro-style distribution API.
HippoTrainer: gradient-based hyperparameter optimization
A PyTorch library that tunes hyperparameters by differentiating through the training loop. Implements T1-T2, Neumann-series implicit differentiation, HOAG and DrMAD behind one trainer interface.
Garage: generative augmentation framework
Replaces objects in images with newly generated ones. Grounded-SAM finds and masks the object, an Augmenter model proposes replacements and prompts, PowerPaint inpaints the result.
Epidemic spread models: COVID-19 as stochastic chemical kinetics
Various approaches to modeling the spread of epidemics, in particular COVID-19, through differential equations and Markov processes. Deterministic and stochastic SIR models fitted to real multi-wave data.