Hi, I am
Bhushan Sonawane

I make Machine Learning run on device

Resume

Experience

with Deep Learning, Computer Vision, Compilers, Teaching

Contribute for Fun

Apr 2018 - Ongoing

Implemented isInf, isFinite,
Negative indices to tensor.narrow,
Return partial keys in load_state_dict

CoreML Frameworks

June 2019 - Ongoing

Working as SE in Machine Learning - CoreML
Working on CoreML converters and CoreML infrastructure

Intern

Santa Clara, USA

May 2018 - Aug 2018

Worked on LLVM based compiler in Vulkan ecosystem to tune and re-order optimizations. Debugging and Compile time infrastructure.

Graduate Researcher, Computer Vision Lab

Dec 2017 - May 2019

Domain Adaptation, Facial Lighting Estimation, Deep Learning

Compiler Developer

Pune, India

Jun 2015 - Jul 2017

Optimizing backend compiler for Graphics and CUDA

Intern

Pune, India

Jun 2014 - Apr 2015

PBQP Based register allocator

Senior Research Aid, Research Foundation for SUNY

Jan 2018 - March 2018

Image conversion tool for Bio-Medical high resolution images

Visiting Instructor, Vishwakarma Institute of Technology, Pune

Jan 2017 - May 2017

Instructed undergraduate course Problem Solving and Programming

Basics

projects in Deep Learning, Computer Vision, Frameworks, Networks, Mobile & Web Apps

Lighting Esimation using GANs

For Face lighting estimation, we lack ground truth lighting.
1. Used Spherical Harmonics(SH) to model Lighting
2. Used SIRFS methods to generate noisy SH for Synthetic and Real images.
3. Implemented domain adaption method and implemented LDAN paper Source code
4. Implemented denoising AutoEncoder to denoise noisy SH and trainined on single neural network which did not perform better than domain adaptation.
5. Report

Co-Operative GANs

Training GANs is difficult, Generator can face following problems- Mode Collapsing, Saddle point and Local Minima
Updated GANs training procedure as follows to solve these problems-
1. Use multiple Generator with different configuration
2. At the end of each epoch copy weights of best performing Generator to all other Generators
3. Step 2 makes sure we start all-over again and progress towards better resolution
4. Results- Results
5. Source - Source

ADMM Optimizer in PyTorch

1. Implemented ADDM Optimizer for Lasso solver and Ridge Regression problem


2. Outperformed Scikit-Learn's state of the art Lasso and Ridge solves. 3. Report

TO BE ADDED

TO BE ADDED