YOLOv2 for Annie (AMD NN Inference Engine)

This project shows how to run tiny yolov2 (20 classes) with AMD's NN inference engine(Annie):

• A python convertor from yolo to caffe
• A c/c++ implementation and python wrapper for region layer of yolov2
• A sample for running yolov2 with Annie

---

Preliminaries

Please install amdovx modules and modelcompiler from https://github.com/GPUOpen-ProfessionalCompute-Libraries/amdovx-modules.git.

Step 1. Compile Python Wrapper

make

Step 2. Convert Caffe to Annie python lib as shown below using NNIR ModelCompiler (amdovx-modules/utils/model_compiler/)

First convert caffe to NNIR format and compile NNIR to deployment python lib using the following steps

% python caffe2nnir.py ./models/caffemodels/yoloV2Tiny20.caffemodel <nnirOutputFolder> --input-dims 1,3,416,416
% python nnir2openvx.py [OPTIONS] <nnirInputFolder> <outputFolder> (details are in ModelCompiler page of amdovx_modules git repository)

There will be a file libannpython.so (under build) and weights.bin

Step 3. Run tests

python3 ./detectionExample/Main.py --image ./data/dog.jpg --annpythonlib <libannpython.so> --weights <weights.bin>
python3 ./detectionExample/Main.py --capture 0 ./data/dog.jpg --annpythonlib <libannpython.so> --weights <weights.bin> (live Capture)

This runs inference and detections and results will be like this:

Run Other YoloV2 models

Convert Yolo to Caffe

Install caffe and config the python environment path.
sh ./models/convertyo.sh

Tips:

Please ignore the error message similar as "Region layer is not supported".

The converted caffe models should end with "prototxt" and "caffemodel".

Update parameters

Please update parameters (biases, object names, etc) in ./src/CRegionLayer.cpp, and parameters (dim, blockwd, targetBlockwd, classe, etc) in ./detectionExample/ObjectWrapper.py.

Please read ./src/CRegionLayer.cpp and ./detectionExample/ObjectWrapper.py for details.

References

• caffe
• yolo
• caffe-yolo

Contributors

• ichigoi7e
• nathiyaa
• cpagravel

---

License

Research Only

Author

[email protected] [email protected]