Convert Yolov4 Darknet to TensorRT add BatchedNMSDynamic

Convert Yolov4 Darknet to ONNX

• If you want to convert model not Dynamic: setting batch_size > 0.
• Run:

python3 darknet2onnx.py --config config/yolov4.cfg --weightfile weights/yolov4_50000.weights --batch_size -1 --onnx_file_path save_convert/yolov4.onnx

Add NMS Plugins

• Open file add-nms.py and update attribute values to suit your model
• Then, run:

python3 add-nms.py --model save_convert/yolov4.onnx

Convert ONNX to TensorRT

/usr/src/tensorrt/bin/trtexec --onnx=save_convert/yolov4-nms.onnx \
                                --saveEngine=save_convert/yolov4-add-nms.trt \
                                --explicitBatch \
                                --minShapes=input:1x3x416x416 \
                                --optShapes=input:1x3x608x608 \
                                --maxShapes=input:1x3x608x608 \
                                --verbose \
                                --device=1

Inference TrT Model add NMS:

• Open file demo_trt_nms.py and update attribute values: model_weights, max_size, names
• Run:

python3 demo_trt_nms.py

REFERENCE

1. https://github.com/Tianxiaomo/pytorch-YOLOv4
2. https://github.com/NNDam/yolor/blob/main/object_detector_trt_nms.py
3. https://github.com/NVIDIA-AI-IOT/yolov4_deepstream