oliver408i / rknn-object-detection

python scripts for rknn object detection for Orange pi 5

rknn-object-detection

python scripts for rknn object detection for Orange pi 5

How to make a model

Setup yolov5 toolkit

Python3.12 required

1. Clone the yolov5 github repo
2. Setup a venv to use with the toolkit
3. Install all the requirements pip install -r requirements.txt
4. Verify that it works (or just wait until you obtain the dataset to train)

Obtain dataset

You can manually make a dataset, or use roboflow. Make sure the image dimensions match how you are going to train it

1. Get the correct dataset format downloaded and in the same folder as the yolov5 toolkit
2. Find the .yaml file for that dataset. You might want to rename it.
3. In that yaml file, update the paths for the train and valid folders so that they are correct

Train

1. Run the following command (make sure you are using your inputs. For testing, do a 3 epoch training first) python train.py --img 640 --epochs 30 --data data.yaml --weights yolov5.pt
2. If you are on a silicon mac using m1 or m2 (m3 not supported yet), add --device mps. If you are on a cuda device (Nvidia GPU) add --device cuda. If you encounter issues, remove the flag to train on cpu
3. Once the training is finished, verify that you have a runs/train/exp/weights/best.pt file. Note that the exp folder may become exp2 or exp3 if you ran multiple trainings. The highest number is the latest training
4. You can upload the model to roboflow using upload.py. Make sure to change the version and expNum variables. After uploading, use the visualize tool on roboflow to test your model

Convert to ONNX

1. Once training is finished, run python export.py --weights /runs/train/exp/weights/best.pt --include onnx
2. Find the output onnx file in the same directory was your best.pt file

Converting to RKNN

1. Setup a linux x86 device with python3.10. Install the RKNN toolkit python package from here
2. Upload your best.onnx file to that device
3. Run convert.py to get the output model.rknn file

Run on rockpi

This is a mystery. Yet to be resolved