进入docker环境

docker run --gpus all -it --rm -v /root/ws/:/root/ws/ nvcr.io/nvidia/pytorch:20.11-py

进到工程目录

cd ~/ws/tensorrtHack2_1

模型导出

不同设备环境，需要重新导出engine文件 python3 resnet50_demo.py -s

yolov5模型导出

./yolov5 -s yolovs-hand.wts yolov5s-hand.engine s

yolov5-int8优化模型导出

./yolov5 -s yolov5s-hand-int8.wts yolov5s-hand-int8.engine s

测试结果

测试没有优化的模型 python yolov5_trt12.py 测试int8量化的模型 python yolov5_trt12_int8.py

模型文件地址

链接: https://pan.baidu.com/s/1QaKLq4q-FRbuCl0asEJ9hg 提取码: qrtc

Benchmark

• 运行Pytorch-FP16测试基准

# Pytorch - FP16 - batch_size = 1
python3 test.py --weights weights/best-hand.pt --data hand.yaml --batch-size=1 --img-size 1
# Pytorch - FP16 - batch_size = 8
python3 test.py --weights weights/best-hand.pt --data hand.yaml --batch-size=1 --img-size 8
# Pytorch - FP16 - batch_size = 16
python3 test.py --weights weights/best-hand.pt --data hand.yaml --batch-size=1 --img-size 16
# Pytorch - FP16 - batch_size = 32
python3 test.py --weights weights/best-hand.pt --data hand.yaml --batch-size=1 --img-size 32

• 运行Tensorrt-FP16测试基准

# Tensorrt - FP16 - batch_size = 1
python yolov5_trt12-test-batch.py 1 /root/ws/datasets_1100/
# Tensorrt - FP16 - batch_size = 8
python yolov5_trt12-test-batch.py 8 /root/ws/datasets_1100/
# Tensorrt - FP16 - batch_size = 16
python yolov5_trt12-test-batch.py 16 /root/ws/datasets_1100/
# Tensorrt - FP16 - batch_size = 32
python yolov5_trt12-test-batch.py 32 /root/ws/datasets_1100/

• 运行Tensorrt-FP8测试基准

# Tensorrt - INT8 - batch_size = 1
python yolov5_trt12_int8-test-batch.py 1 /root/ws/datasets_1100/
# Tensorrt - INT8 - batch_size = 8
python yolov5_trt12_int8-test-batch.py 8 /root/ws/datasets_1100/
# Tensorrt - INT8 - batch_size = 16
python yolov5_trt12_int8-test-batch.py 16 /root/ws/datasets_1100/
# Tensorrt - INT8 - batch_size = 32
python yolov5_trt12_int8-test-batch.py 32 /root/ws/datasets_1100/

结果输出如下： Excel文件下载