Triton face recogition

End-to-end deploy face detection, face align, face recogition in triton inference server

Prerequisites

• Docker
• Nvidia-driver
• Nvidia-docker2

Export .engine

Face detection model

• Export checkpoint to onnx model:
  • Clone yolov7-face-detection and cd into yolov7-face-detection folder
  • Download weight and save into weights/yolov7-tiny33.pt
  • Export to onnx: python3 export.py --weights ./weights/yolov7-tiny33.pt --img-size 640 --batch-size 1 --dynamic-batch --grid --end2end --max-wh 640 --topk-all 100 --iou-thres 0.5 --conf-thres 0.2 --device 1 --simplify --cleanup --trt
• Or download onnx file from from github.com/hiennguyen9874/yolov7-face-detection/releases/tag/v0.1 and save as models/FaceDetection/1/yolov7-tiny41-nms-trt.onnx
• Export to TensorRT: docker-compose run --rm triton /usr/src/tensorrt/bin/trtexec --onnx=/models/FaceDetection/1/yolov7-tiny41-nms-trt.onnx --saveEngine=/models/FaceDetection/1/model.plan --fp16 --minShapes=images:1x3x640x640 --optShapes=images:1x3x640x640 --maxShapes=images:4x3x640x640 --shapes=images:1x3x640x640 --workspace=12288

Face recognition model

• Download webface_r50.onnx from deepinsight/insightface and cleaning onnx file: python3 scripts/onnx_clean.py --onnx-path samples/models/Secondary_Recognition/ms1mv3_r50_pfc.onnx --image-size 112,112 --batch-size 1 --simplify --dynamic --cleanup
• Or download onnx file from from: github.com/hiennguyen9874/deepstream-face-recognition/releases/tag/v0.1 and save as models/FaceRecognition/1/webface_r50_dynamic_simplify_cleanup.onnx
• Export to TensorRT: docker-compose run --rm triton /usr/src/tensorrt/bin/trtexec --onnx=/models/FaceRecognition/1/webface_r50_dynamic_simplify_cleanup.onnx --saveEngine=/models/FaceRecognition/1/model.plan --fp16 --minShapes=input.1:1x3x112x112 --optShapes=input.1:1x3x112x112 --maxShapes=input.1:16x3x112x112 --shapes=input.1:1x3x112x112 --workspace=12288

Run demo

• Up docker: docker-compose up --build
• main.ipynb