Easy and Efficient Object Detector

EOD (Easy and Efficient Object Detection) is a general object detection model production framework. It aim on provide two key feature about Object Detection:

• Efficient: we will focus on training VERY HIGH ACCURARY single-shot detection model, and model compression (quantization/sparsity) will be well addressed.
• Easy: easy to use, easy to add new features(backbone/head/neck), easy to deploy.
• Large-Scale Dataset Training Detail
• Equalized Focal Loss for Dense Long-Tailed Object Detection EFL
• Improve-YOLOX YOLOX-RET
• Quantization Aware Training(QAT) interface based on MQBench.

The master branch works with PyTorch 1.8.1. Due to the pytorch version, it can not well support the 30 series graphics card hardware.

pip install -r requirments

Some example scripts are supported in scripts/.

Export eod into ROOT and PYTHONPATH

ROOT=../../
export ROOT=$ROOT
export PYTHONPATH=$ROOT:$PYTHONPATH

Step1: edit meta_file and image_dir of image_reader:

dataset:
  type: coco # dataset type
    kwargs:
      source: train
      meta_file: coco/annotations/instances_train2017.json 
      image_reader:
        type: fs_opencv
        kwargs:
          image_dir: coco/train2017
          color_mode: BGR

Step2: train

python -m eod train --config configs/det/yolox/yolox_tiny.yaml --nm 1 --ng 8 --launch pytorch 2>&1 | tee log.train

• --config: yamls in configs/
• --nm: machine number
• --ng: gpu number for each machine
• --launch: slurm or pytorch

Step3: fp16, add fp16 setting into runtime config

runtime:
    fp16: True

Step1: edit config of evaluating dataset

Step2: test

python -m eod train -e --config configs/det/yolox/yolox_tiny.yaml --nm 1 --ng 1 --launch pytorch 2>&1 | tee log.test

Step1: add visualizer config in yaml

inference:
  visualizer:
    type: plt
    kwargs:
      class_names: ['__background__', 'person'] # class names
      thresh: 0.5

Step2: inference

python -m eod inference --config configs/det/yolox/yolox_tiny.yaml --ckpt ckpt_tiny.pth -i imgs -v vis_dir

• --ckpt: model for inferencing
• -i: images directory or single image
• -v: directory saving visualization results

EOD supports mpirun mode to launch task, MPI needs to be installed firstly

# download mpich
wget https://www.mpich.org/static/downloads/3.2.1/mpich-3.2.1.tar.gz # other versions: https://www.mpich.org/static/downloads/

tar -zxvf mpich-3.2.1.tar.gz
cd mpich-3.2.1
./configure  --prefix=/usr/local/mpich-3.2.1
make && make install

Launch task

mpirun -np 8 python -m eod train --config configs/det/yolox/yolox_tiny.yaml --launch mpi 2>&1 | tee log.train

• Add mpirun -np x; x indicates number of processes
• Mpirun is convenient to debug with pdb
• --launch: mpi

• custom dataset
• rank_dataset

• YOLOX
• YOLOX-Ret
• EFL
• YOLOV5
• RetinaNet
• QAT

• Training on custom data
• Register modules
• EOD developing mode

• Learn about configs
• Datasets
• Augmentations and preprocesses
• Loss functions

• Hooks
• Visualization

• YOLOX
• YOLOv5

Thanks to all past contributors, especially opcoder,