Yolo5 for Jetson Devices

This repo contains optimized yolo5 for jetson devices.

• All the dependencies including ones to train and validate in a compact docker container.
• On the fly TensorRT engine build with FP32, FP16, INT8 with calibration.
• No need for any custom plugins.
• At least 4X inference speed up...could be more with different devices and configs.
• Dynamic shape input is also supported

Usage

Clone the repo from github.com:

git clone https://github.com/deepsworld/Yolo5Optimized
cd Yolo5Optimized

Pull and run pre-built docker image

• DETECT: run sample detection on images in a directory. The output rendered images will be stored in outputs dir.

make run
cd yolo5
# run inference on images in assets 
python3 detect.py --fp16 --inputs assets --outputs outputs

• EVALUATE: run evaluation on COCO validation set (2017)

make run
cd yolo5
# start evaluation script
python3 test.py --data coco.yaml --fp16

Build docker image

make build 

Evaluation on COCO

• Device: Jetson Nano 4GB
• Arch: yolo5s
• Batch_Size: 2
• Scale: 640
• FP16: True

Class	Images	Labels	P	R	mAP@.5	mAP@.5:.95
all	5000	36335	66.6	49.8	54.9	35.3

Speed: 1.3/52.4/53.7 ms inference/NMS/total per 640x640 image at batch-size 2

Notes

• It will take ~5-6 mins to build fp16 engine on Jetson Nano
• Consider disabling GUI to free up some extra memory
• Run in 10W power mode with sudo nvpmodel -m 0 or make max_power
• The validation script can be used to validate over some other test set with minor changes if needed.
• If using a Jetson Nano with 2GB memory, set DEFAULT_WORKSPACE_SIZE = 2 << 20 or 1 << 20.

Common Issues and Resolutions

• Docker build fails with make build: This could happen if the default docker runtime is not set to nvidia. To fix this do as follow:

# edit docker daemon json
sudo vi /etc/docker/daemon.json

# Make sure it looks like this
{
    "default-runtime": "nvidia",
    "runtimes": {
        "nvidia": {
            "path": "nvidia-container-runtime",
            "runtimeArgs": []
        }
    }
}

# restart docker daemon
sudo systemctl restart docker

• Disable GUI to free up some extra memory

sudo systemctl set-default multi-user
sudo reboot now

References

• https://github.com/necla-ml/ML-Vision
• https://github.com/necla-ml/ML

Contributing

• Please raise an issue if you find any bugs or problems in running the code.