English| 中文 |CSDN

OS supported (you can check other OS by yourself)

	windows	linux	mac
checked
gpu

CPU checked

	Intel i7	raspberry 3B	raspberry 4B	Jeston NX
checked

Features

• C++ Only. 3rdparty blas lib is optional, also you can use OpenBlas.
• OS supported: Windows, Linux(Ubuntu checked) and Mac os(unchecked).
• CPU supported: Intel X86, AMD(unchecked) and ARM(checked: armv7 armv8 arrch64).
• x86 avx2 supported.(Working....)
• arm neon supported.(Working....)。
• NNPack supported.(arm)。
• Keras to Msnhnet is supported. (Keras 2 and tensorflow 1.x)
• GPU cuda supported.(Checked GTX1080Ti, Jetson NX)
• GPU cudnn supported.(Checked GTX1080Ti, Jetson NX)
• GPU fp16 mode supported.(Checked GTX1080Ti, Jetson NX.)
• ps. Please check your card wheather fp16 full speed is supported.
• c_api supported.
• keras 2 msnhnet supported.(Keras 2 and tensorflow 1.x, part of op)
• pytorch 2 msnhnet supported.(Part of op, working on it)
• MsnhnetSharp supported.
• A viewer for msnhnet is supported.(netron like)
• Working on it...(Weekend Only (╮（╯＿╰）╭))

Tested networks

• lenet5
• lenet5_bn
• alexnet(torchvision)
• vgg16(torchvision)
• vgg16_bn(torchvision)
• resnet18(torchvision)
• resnet34(torchvision)
• resnet50(torchvision)
• resnet101(torchvision)
• resnet152(torchvision)
• darknet53(Pytorch_Darknet53)
• googLenet(torchvision)
• mobilenetv2(torchvision)
• yolov3(u版yolov3)
• yolov3_spp(u版yolov3)
• yolov3_tiny(u版yolov3)
• yolov4(u版yolov3)
• fcns(pytorch-FCN-easiest-demo)
• unet(bbuf keras)
• deeplabv3(torchvision)
  ==============================================================
• mobilenetv2_yolov3_lite (cudnn does not work with GTX10** Pascal Card, please use GPU model only)
• mobilenetv2_yolov3_nano (cudnn does not work with GTX10** Pascal Card, please use GPU model only)
• yoloface100k (cudnn does not work with GTX10** Pascal Card, please use GPU model only)
• yoloface500k (cudnn does not work with GTX10** Pascal Card, please use GPU model only)
• Thanks: https://github.com/dog-qiuqiu/MobileNetv2-YOLOV3 ==============================================================
• pretrained models 链接：https://pan.baidu.com/s/1mBaJvGx7tp2ZsLKzT5ifOg 提取码：x53z

Yolo Test

• Win10 MSVC 2017 I7-10700F

  net	yolov3	yolov3_tiny	yolov4
  time	380ms	50ms	432ms

• ARM(Yolov3Tiny cpu)

  cpu	raspberry 3B	raspberry 4B	Jeston NX
  without NNPack	6s	2.5s	1.2s
  with NNPack	2.5s	1.1s	0.6s
  with neon asm	?	0.432s	?

Yolo GPU Test

• Ubuntu16.04 GCC Cuda10.1 GTX1080Ti

  net	yolov3	yolov3_tiny	yolov4
  time	30ms	8ms	30ms

• Jetson NX

  net	yolov3	yolov3_tiny	yolov4
  time	200ms	20ms	210ms

Yolo GPU cuDnn FP16 Test

• Jetson NX

  net	yolov3	yolov4
  time	115ms	120ms

Mobilenet Yolo GPU cuDnn Test

• Jetson NX

  net	yoloface100k	yoloface500k	mobilenetv2_yolov3_nano	mobilenetv2_yolov3_lite
  time	7ms	20ms	20ms	30ms

DeepLabv3 GPU Test

• Ubuntu18.04 GCC Cuda10.1 GTX2080Ti

  net	deeplabv3_resnet101	deeplabv3_resnet50
  time	22.51ms	16.46ms

Requirements

• OpenCV4 https://github.com/opencv/opencv
• yaml-cpp https://github.com/jbeder/yaml-cpp
• Qt5 (optional. for Msnhnet viewer) http://download.qt.io/archive/qt/
• cuda10+ cudnn 7.0+.(for GPU)

Video tutorials(bilibili)

• Build on Linux
• Build on Windows
• Pytorch Params to msnhbin

How to build

• With CMake 3.10+

• Viewer can not build with GPU.

• Options
  
  ps. You can change omp threads by unchecking OMP_MAX_THREAD and modifying "num" val at CMakeLists.txt:52
  

• Windows

1. Compile opencv4 and yaml-cpp.
2. Config environment. Add "OpenCV_DIR" and "yaml-cpp_DIR"
3. Get qt5 and install. http://download.qt.io/ (optional)
4. Add qt5 bin path to environment.
5. Then use cmake-gui tool and visual studio to make or use vcpkg.

• Linux(Ubuntu)

ps. If you want to build with Jetson, please uncheck NNPACK, OPENBLAS, NEON.

sudo apt-get install qt5-default      #optional
sudo apt-get install libqt5svg5-dev   #optional
sudo apt-get install libopencv-dev

# build yaml-cpp
git clone https://github.com/jbeder/yaml-cpp.git
cd yaml-cpp
mkdir build 
cd build 
cmake .. -DYAML_BUILD_SHARED_LIBS=True -DYAML_CPP_BUILD_TESTS=False
make -j4
sudo make install 

#config 
sudo echo /usr/local/lib > /etc/ld.so.conf.d/usrlib.conf
sudo ldconfig

# build Msnhnet
git clone https://github.com/msnh2012/Msnhnet.git
mkdir build 

cd Msnhnet/build
cmake -DCMAKE_BUILD_TYPE=Release ..  
make -j4
sudo make install

vim ~/.bashrc # Last line add: export PATH=/usr/local/bin:$PATH
sudo ldconfig

Test Msnhnet

• 1. Download pretrained model and extract. eg.D:/models.
• 2. Open terminal and cd "Msnhnet install bin". eg. D:/Msnhnet/bin
• 3. Test yolov3 "yolov3 D:/models".
• 4. Test yolov3tiny_video "yolov3tiny_video D:/models".
• 5. Test classify "classify D:/models".
     

View Msnhnet

• 1. Open terminal and cd "Msnhnet install bin" eg. D:/Msnhnet/bin
• 2. run "MsnhnetViewer"

PS. You can double click "ResBlock Res2Block AddBlock ConcatBlock" node to view more detail
ResBlock

Res2Block

AddBlock

ConcatBlock

How to convert your own pytorch network

• pytorch2msnhnet
• ps. ultralytics yolov3 is not supported. You can trans backbone first and then add some OPs manually.

About Train

• Just use pytorch to train your model, and export as msnhbin.
• eg. yolov3/v4 https://github.com/ultralytics/yolov3

Enjoy it! :D

Acknowledgement

Msnhnet got ideas and developed based on these projects：

• DarkNet
• NCNN
• ACL

加群交流