This folder contains scripts to setup Torchserve for Yolov5, popular open-source object detection method.
1. Create env
sh create_env.sh
Note: assume os is linux with cuda 11.0
2. Download the model from YoloV5 releases
cd yolov5_mar
sh download_model.sh
3. Archive the model files
cd yolov5_mar
zip -r ../yolo5.mar .
cd ..
4. Host the yolov5 model
mkdir model_store
mv yolo5.mar model_store/.
5. Run the server
nohup sh start_server.sh &
6. Check the healthy
curl "http://localhost:8081/models/yolo5"
or check logs/ts_log.log
7. Test object detection API
curl -X POST http://127.0.0.1:8080/predictions/yolo5 -T kitten.jpg
Coco 2017 Val Images(5K) is used to test throughput/latency in the above setup.
1. Server machine: Intel xeon Gold 6159 [email protected], 2 Tesla P4 with CentOS 7
2. Test setup
a. 10 http clients in parallel. Each client which runs in other machine sends 500 images sequentially.
Throughput: 26 images/second Average latency: 384 millisec/request
b. Native torch program (yolov5_mar/simple_batch_test.py) without torchserve. Run two processes (batch_size: 10). Each one has own gpu and dataset (#images: 2500).
Throughput: 26 images/second.