jnbraun / bcnn

A minimalist Deep Learning framework for embedded Computer Vision

BCNN

Introduction

BCNN (Bare Convolutional Neural Networks) is a minimalist framework designed to prototype, train and deploy convolutional neural networks for embedded computer vision applications.

Features

• Written in C99. Clean C API designed to be integrated in C or C++ codebase.
• Lightweight: the minimal build requires no external dependency.
• Modular: Can leverage a Blas library such as OpenBLAS on CPU. Can also run on Nvidia's GPU. CuDNN is supported to offer maximal speed.
• Fast: Optimized CPU inference speed using AVX and ARM Neon vectorizations and OpenMP multithreading.
• Flexible: Supports multi inputs / outputs / branches. Provides the commonly used operators to build state-of-the-art CNN architectures (ResNet, DenseNet, MobileNet, Yolo ...)
• Command line tool to train / evaluate models via simple configuration file.
• Online data augmentation via bip, a fast image processing library (usable as standalone module).
• (Experimental) Model converters from Caffe -> bcnn and bcnn -> TensorFlow Lite.

Getting started

Download or clone the repository:

git clone https://github.com/jnbraun/bcnn.git

You need to have cmake installed in order to build the library.

[Optional] Dependencies

CPU

• Minimal build: no external dependency.
• Build with Blas: requires a blas library (OpenBLAS is recommended).

GPU

Requires CUDA libraries (cudart, cublas, curand) and a GPU with compute capability 2.0 at least. CuDNN is optional but supported.

Build

• User configuration: Depending on you system, you may want to edit the following lines of the CMakeLists.txt:

# User configuration settings
option(USE_AVX "Build with AVX instructions" ON)
option(USE_CUDA "Build with CUDA libraries" OFF)
option(USE_CUDNN "Build with CuDNN library" OFF)
option(USE_BLAS "Build with BLAS library" ON)
option(USE_NEON "Build with Neon instructions" OFF)
option(USE_OPENMP "Enable OpenMP multithreading" ON)

• [Optional] When building with CUDA and / or CuDNN, you may need to adjust the following line depending on the compute capability of your GPU:

# Uncomment the proper line according to the system cuda arch
set(CUDA_ARCH 
    #"-gencode arch=compute_30,code=sm_30;"
    #"-gencode arch=compute_35,code=sm_35;"
    "-gencode arch=compute_50,code=sm_50;"
    "-gencode arch=compute_50,code=compute_50;"
    "-gencode arch=compute_52,code=sm_52;"
    #"-gencode arch=compute_60,code=sm_60;"
    #"-gencode arch=compute_61,code=sm_61;"
)

• Build

cd path/to/bcnn
mkdir build
cd build/
cmake ../
make

How to use it

• Use the command line tool bcnn-cl with configuration file: see an example here.

• Or use the static library and write your own code: see an example there.

License

Released under MIT license.