Rethinking SWATS (Switching from Adam to SGD) Optimizer
A variant of SWATS optimizer which outperforms vanilla SWATS in terms of test accuracy by 1.3 % for ResNet-18 and 1.4 % for DenseNet-121 on Cifar-10 dataset.
Link
messup: A new regularization method
A new regularization technique by encountering samples through exponential smoothing inspired by mixup regularization. This reduces generalization error by 1.6 % when compared to ERM.
Link
Symm-PPO
A variant of OpenAI’s popular reinforcement learning algorithm - Proximal Policy Optimization (PPO). Symm-PPO inculcates the analogy of new policy not being too different from the old one for the entropy term. Median rewards from Symm-PPO outperforms/is on-par with vanilla PPO.
Link
MusePlay
A Spotify like music streaming website built entirely with PHP, JavaScript, CSS and HTML with PostgreSQL done as part of Database Management Systems coursework. MusePlay provides features such as user creation, playlist creation, search music, personalized music recommendation, top songs local/global.
Link
pytorch-LapGAN
Laplacian pyramid framework to generate images in PyTorch.
Link
pytorch-Group-Normalization
Empirical justification on why Group Normalization works.
Link