bot66/yolov5_tinytrt

Deploy yolov5 on TensorRT with Libtinytrt ⚡️.Both x86 and ARM(NVIDIA Jetson).

Benchmark

Models	Device	BatchSize	Mode	Input Shape(HxW)	Pytorch	TensorRT
YOLOV5-X	RTX3070	1	FP32	640x640	38.0ms	21.1ms
YOLOV5-L	RTX3070	1	FP32	640x640	23.3ms	13.4ms
YOLOV5-M	RTX3070	1	FP32	640x640	12.0ms	7.5ms
YOLOV5-S	RTX3070	1	FP32	640x640	6.3ms	4.6ms
YOLOV5-S	Jetson Nano	1	FP32	640x640	\	6.4ms

Installation

Build on x86

Require TensorRT 8+ . Recommend use Nvidia official Docker image: nvcr.io/nvidia/pytorch:21.11-py3

Create docker container

#this docker image is tested, recommend pull this image
docker pull nvcr.io/nvidia/pytorch:21.11-py3

#create container
nvidia-docker run -it --name yolov5_tinytrt nvcr.io/nvidia/pytorch:21.11-py3 /bin/bash

Build on ARM(NVIDIA Jetson)

Recommend pull this docker image l4t-ml:r32.6.1-py3(OpenCV inside).Make sure your JetPack version support it. My Jetson Nano is JetPack 4.4, also can run this docker image.

Create docker container

#this docker image is tested, recommend pull this image
docker pull nvcr.io/nvidia/l4t-ml:r32.6.1-py3

#create container
nvidia-docker run -it --name yolov5_tinytrt nvcr.io/nvidia/l4t-ml:r32.6.1-py3 /bin/bash

Install

#clone project and submodule
git clone --recurse-submodules -j8 https://github.com/bot66/yolov5_tinytrt.git

#install dependencies
sudo apt-get update -y
sudo apt-get install cmake zlib1g-dev

#for python binding
sudo apt-get install python3 python3-pip
pip3 install numpy

#build
cd yolov5_tinytrt
mkdir build && cd build

cmake .. && make -j8

Usage

Use TensorRT to speed up your model, you need parse it to TensorRT .engine format, you can use build/tiny-tensorrt/tinyexec to parse .onnx model to create .engine model.

#default yolov5s as example

#generate engine 
./build/tiny-tensorrt/tinyexec --onnx yolov5s.onnx --model yolov5s.engine

#inference
#usage:./yolov5_tinytrt <engine_model> <input_folder> <output_folder>
./build/yolov5_tinytrt yolov5s.engine  images/ results/

Reference

zerollzeng / tiny-tensorrt

ultralytics / yolov5

yolov5_tinytrt's Issues

can support yolov5-6.0 models?

can use yolov5n6,yolov5s6 or some else v6.0 models on jetson platform?
please help!