Skip to content

A single stage object detection toolbox based on PyTorch

License

Notifications You must be signed in to change notification settings

deepdrivepl/vedadet

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

77 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Introduction

vedadet is a single stage object detector toolbox based on PyTorch.

Features

  • Modular Design

    We re-design MMDetection based on our taste and needs. Specifically, we decompose detector into four parts: data pipeline, model, postprocessing and criterion which make it easy to convert PyTorch model into TensorRT engine and deploy it on NVIDIA devices such as Tesla V100, Jetson Nano and Jetson AGX Xavier, etc.

  • Support of several popular single stage detector

    The toolbox supports several popular single stage detector out of the box, e.g. RetinaNet, FCOS, etc.

  • Friendly to TensorRT

    Detectors can be easily converted to TensorRT engine.

  • Easy to deploy

    It's simple to deploy the model accelerate by TensorRT on NVIDIA devices through Python front-end or C++ front-end.

License

This project is released under the Apache 2.0 license.

Installation

Requirements

  • Linux
  • Python 3.7+
  • PyTorch 1.6.0 or higher
  • CUDA 10.2 or higher

We have tested the following versions of OS and softwares:

  • OS: Ubuntu 16.04.6 LTS
  • CUDA: 10.2
  • PyTorch 1.6.0
  • Python 3.8.5

Install vedadet

a. Create a conda virtual environment and activate it.

conda create -n vedadet python=3.8.5 -y
conda activate vedadet

b. Install PyTorch and torchvision following the official instructions, e.g.,

conda install pytorch torchvision -c pytorch

c. Clone the vedadet repository.

git clone https://github.com/Media-Smart/vedadet.git
cd vedadet
vedadet_root=${PWD}

d. Install vedadet.

pip install -r requirements/build.txt
pip install -v -e .

Train

a. Config

Modify some configuration accordingly in the config file like configs/trainval/retinanet/retinanet.py

b. Multi-GPUs training

tools/dist_trainval.sh configs/trainval/retinanet/retinanet.py "0,1"

c. Single GPU training

CUDA_VISIBLE_DEVICES="0" python tools/trainval.py configs/trainval/retinanet/retinanet.py

Test

a. Config

Modify some configuration accordingly in the config file like configs/trainval/retinanet/retinanet.py

b. Test

CUDA_VISIBLE_DEVICES="0" python tools/test.py configs/trainval/retinanet/retinanet.py weight_path

Inference

a. Config

Modify some configuration accordingly in the config file like configs/infer/retinanet/retinanet.py

b. Inference

CUDA_VISIBLE_DEVICES="0" python tools/infer.py configs/infer/retinanet/retinanet.py image_path

Deploy

a. Convert to Onnx

Firstly, install volksdep following the official instructions.

Then, run the following code to convert PyTorch to Onnx. The input shape format is CxHxW. If you need the onnx model with constant input shape, please remove --dynamic_shape in the end.

CUDA_VISIBLE_DEVICES="0" python tools/torch2onnx.py configs/trainval/retinanet/retinanet.py weight_path out_path --dummy_input_shape 3,800,1344 --dynamic_shape

Here are some unsupported operations for model conversion.

  • GN
  • Deformable Conv

Please see more details in this.

b. Inference SDK

Firstly, install flexinfer following the official instructions.

Then, see the example for details.

Contact

This repository is currently maintained by Yanjia Zhu (@mike112223), Hongxiang Cai (@hxcai), Yichao Xiong (@mileistone).

Credits

We got a lot of code from mmcv and mmdetection, thanks to open-mmlab.

About

A single stage object detection toolbox based on PyTorch

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages

  • Python 87.0%
  • Cuda 7.1%
  • C++ 5.8%
  • Shell 0.1%