Questions about how to deploy yolov4_tiny in deepstream

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

I referenced @kayccc YOLOV4 example in DeepStream - #4

This link cannot handle the problem of yolov4_tiny. I want to use the darknet model of v4_tiny or the pytorch model in deepstream. I found that the darknet model of yolov4_tiny and the model trained by pytorch cannot be converted to onnx. How do I deal with it?

3

Sorry for the late response, we will do the investigation soon.

4

Hi,
Regarding exporting the model to onnx, you could consult in the TRT forum or the related github.

Thanks!

5

Thank you for your attention, if you find a good solution, please let me know, thank you

6

Sure~

7

@18981275647

i find a repo can work, from darknet to onnx to tensorrt , could you try to verify it?
we find a github repo about yolov4-tiny in https://github.com/linghu8812/tensorrt_inference/tree/master/Yolov4.
then we download the repo :

git clone https://github.com/linghu8812/tensorrt_inference.git
then download the weight of yolov4-tiny follow the readme.md of the repo:

wget https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-tiny.weights
follow the readme,convert the model to onnx:
python3 export_onnx.py --cfg_file cfg/yolov4-tiny.cfg --weights_file yolov4-tiny.weights --output_file yolov4-tiny.onnx --strides 32 16 --neck FPN
use trtexec to convert the onnx model to trt plan:
trtexec --onnx=yolov4-tiny.onnx --saveEngine=yolov4-tiny.plan --verbose

8

@18981275647
and i found another repo about this, you can also try it.

Tianxiaomo/pytorch-YOLOv4: PyTorch ,ONNX and TensorRT implementation of YOLOv4 (github.com)

9

Thanks, but what I need is the v4-tiny conversion of pytorch training