DL-Cpp

C++ lightweight library enabling the creation of Machine Learning models using configuration files

Features Supported

• Create Playground, a benchmarker to measure correctness and performance of ML frameworks vs PyTorch.
• Support cross-compatible serialization of weights between PyTorch and DL-CPP
• Provide framework for modular creation of Deep Learning models, and ablility to represent with simple cfg (libconfig++) files.
• Perceptron and MLP cfg support pre-packaged
• Support for lightweight automatic differentiation engine. Reverse mode autodiff library for general usage as well as specifically backpropogation in DL-Cpp

Features in progress

• Use pybind11 to expose Playground to DL-CPP for graphing purposes
• Support for CUDA in addition to scalar code for forward and backpropagation
• pre-packaged cfg support for CNN and RNN architectures

Get Started

The major working parts that need to be puzzled together to design and create a model and test/train workflow using DL-CPP are as listed:

• Model: The architecture of the model created. The template.
• Dataset: Wrapper for raw data
• DataLoader: Wrapper for obtaining train/test batches or "loading" from dataset
• Config: Design the parameters required and how they are read. This needs to be provided when creating a model

Example Design

Perceptron workflow using Playground

• Model: Perceptron
• Dataset: PlaygroundDataset
• DataLoader: PlaygroundDataLoader
• Config: CFG

Build

Ensure libconfig++ and catch2 are installed in your system

DL-CPP

cmake --build . --target clean; make;

Autodiff Tests

cd units/autodiff/tests && cmake --build . --target clean; make; cd ../../../

Run

DL-CPP Suggested run command

./dl_cpp <cfg> > log 2> debug

Autodiff Tests

./units/autodiff/tests/test_autodiff