CV_From_Scratch

CV_From_Scratch is a project where I implemented a PyTorch-like library in C++ for Computer Vision tasks. This library includes layers, a CNN architecture, a Learner class, data loaders, and basic image transforms, all from scratch. The project uses OpenCV for image processing and LibTorch (temporarily) for tensor operations. My colleague later replaced LibTorch with a custom tensor implementation, this final version can be seen in the final branch.

Features

• Implemented CNN architecture and layers from scratch in C++.
• DataLoader and Datasets with basic transformations using OpenCV.
• A simple demonstration app for training and testing a CNN model.
• Unit tests for verification (optional).

Usage

While this project doesn't support CUDA and isn't designed for high-performance usage, it includes a basic terminal-based demonstration app. This app allows you to:

• Train a CNN.
• Run inference on the trained model.

To use the app, compile the project and run the CV_From_Scratch executable.

There are also unit tests located in the tests/ directory that can be executed by running tests/tests after compilation.

Setup

To set up the project locally, follow these steps:

1. Clone the repository:

   git clone <repository-url>
   cd CV_From_Scratch
   

2. Download the MNIST dataset: Run the following Python script to download the MNIST dataset and place the images in the data/ directory: python get_data.py

3. Download required libraries: You need to download the following libraries and place them in the lib/ directory:

• LibTorch

• OpenCV

• Catch2 (for tests)

  The details regarding setting up all the 3 above mentioned libraries is in the README.md file inside lib/

Note: If you do not want to use tests, you can skip Catch2 by disabling tests in the CMakeLists.txt file:

# Add tests directory if tests are enabled
option(ENABLE_TESTS "Enable unit tests" ON) # Set ON -> OFF to disable tests
if (ENABLE_TESTS)
    add_subdirectory(tests)
endif()

4. Build and compile the project: Follow the standard CMake build process:

mkdir build
cd build
cmake ..
make

5. Run the demonstration app: After compilation, run the executable: ./CV_From_Scratch

Project Structure

Here’s a quick overview of the project structure:

CV_From_Scratch/
├── build/             # A place for compiling the project
├── data/              # MNIST dataset
├── include/           # Header files for CNN, layers, data loading, etc.
├── lib/               # External libraries (LibTorch, OpenCV, Catch2)
├── src/               # Source code for CNN, layers, data loading, etc.
├── tests/             # Unit tests (optional)
├── CMakeLists.txt     # CMake configuration
├── get_data.py        # Download MNIST dataset
└── README.md          # Project description

License

This project is licensed under the MIT License. See the LICENSE file for more information.