https://github.com/jnbraun/bcnn
A minimalist Deep Learning framework for embedded Computer Vision
https://github.com/jnbraun/bcnn
arm c99 computer-vision convolutional-neural-networks deep-learning edge-ai embedded-vision gpu high-performance
Last synced: 6 months ago
JSON representation
A minimalist Deep Learning framework for embedded Computer Vision
- Host: GitHub
- URL: https://github.com/jnbraun/bcnn
- Owner: jnbraun
- License: mit
- Created: 2016-09-25T18:23:31.000Z (about 9 years ago)
- Default Branch: master
- Last Pushed: 2019-12-31T01:34:41.000Z (almost 6 years ago)
- Last Synced: 2025-02-02T17:42:42.364Z (8 months ago)
- Topics: arm, c99, computer-vision, convolutional-neural-networks, deep-learning, edge-ai, embedded-vision, gpu, high-performance
- Language: C
- Homepage:
- Size: 1.56 MB
- Stars: 45
- Watchers: 7
- Forks: 9
- Open Issues: 3
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# BCNN
[](https://travis-ci.org/jnbraun/bcnn/)
[](LICENSE)
## 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](https://github.com/jnbraun/bcnn/tree/master/examples/yolo) ...)
* [Command line tool](https://github.com/jnbraun/bcnn/tree/master/examples/mnist_cl) to train / evaluate models via simple configuration file.
* Online data augmentation via [bip](https://github.com/jnbraun/bcnn/tree/master/src/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](https://github.com/jnbraun/bcnn/tree/master/examples/mnist_cl).
* Or use the static library and write your own code: see an example [there](https://github.com/jnbraun/bcnn/tree/master/examples/mnist).
## License
Released under MIT license.