Macrograd
Built a custom machine learning framework (like Pytorch) in C++
Overview
Implemented a custom machine learning framework in C++ by understanding backpropagation in deep learning. Prototyped the framework in python, followed by porting over the code to C++, which helped speed up the computation while training. Future work involves benchmarking and comparing with pytorch, followed by implementing matrix multiplications in CUDA to optimize when GPU resources are available.
Technologies Used
Python
Numpy
C++
Pytorch
CMake
GMock
Key Features
- Can train any number of connected linear layers and train models seamlessly.
- Can use any of the following activation functions: Tanh, sigmoid, ReLU.
- Weights are saved in human readable format, making it easy to import weights into an existing pytorch model.