YOLOv3-Torch2TRT

Introduction

Convert YOLOv3 and YOLOv3-tiny (PyTorch version) into TensorRT models, through the torch2trt Python API.

Installation

Clone the repo

git clone https://github.com/DocF/YOLOv3-Torch2TRT.git

Download pretrained weights

$ cd weights/
$ bash download_weights.sh

Requirements

Two special Python packages are needed:

• tensorrt

• torch2trt

Due to the upsampling operation in YOLO, according to torch2trt API introduction, you need to install the version with plugins.

Installation reference: https://github.com/NVIDIA-AI-IOT/torch2trt

Check torch2trt API

python3 check.py

Inference Acceleration

Acceleration Techs：

• FP16

• TensorRT

Here are some results on TITAN xp:

Entire Model* = Backbone + Feature Net + YOLO Head

python3 detect.py

Statement

This repo is based on PyTorch-YOLOv3. Thx for the great repo.