Forked from https://github.com/digitalbrain79/darknet-nnpack, where NNPACK was used to optimize Darknet without using a GPU. It is useful for embedded devices using ARM CPUs.
Comparing with original version, the modifications/improvements in this version are:
- Reducing inference memory footprint by removing unnecessary memory allocations.
- Fixing bugs on loading weight on 32-bit OS (For example, Raspbian)
- Improving C++ compatibility.
Log in to Raspberry Pi using SSH.
Install PeachPy and confu
sudo pip install --upgrade git+https://github.com/Maratyszcza/PeachPy
sudo pip install --upgrade git+https://github.com/Maratyszcza/confu
Install Ninja
git clone https://github.com/ninja-build/ninja.git
cd ninja
git checkout release
./configure.py --bootstrap
export NINJA_PATH=$PWD
Install clang
sudo apt-get install clang
Install NNPACK-darknet
git clone https://github.com/thomaspark-pkj/NNPACK-darknet.git
cd NNPACK-darknet
confu setup
python ./configure.py --backend auto
$NINJA_PATH/ninja
sudo cp -a lib/* /usr/lib/
sudo cp include/nnpack.h /usr/include/
sudo cp deps/pthreadpool/include/pthreadpool.h /usr/include/
Build darknet-nnpack
git clone https://github.com/thomaspark-pkj/darknet-nnpack.git
cd darknet-nnpack
make
The weight files can be downloaded from the YOLO homepage.
YOLOv2
./darknet detector test cfg/coco.data cfg/yolo.cfg yolo.weights data/person.jpg
Tiny-YOLO
./darknet detector test cfg/voc.data cfg/tiny-yolo-voc.cfg tiny-yolo-voc.weights data/person.jpg
| Model | Build Options | Prediction Time (seconds) |
|---|---|---|
| YOLOv2 | NNPACK=1,ARM_NEON=1 | 8.2 |
| YOLOv2 | NNPACK=0,ARM_NEON=0 | 156 |
| Tiny-YOLO | NNPACK=1,ARM_NEON=1 | 1.3 |
| Tiny-YOLO | NNPACK=0,ARM_NEON=0 | 38 |