Object-Detection-level2-9

P-Stage level-2 object detection competition. (이미지 내 쓰레기 객체 감지 및 분류)

9조 하나둘셋Net() Solution

Archive contents

baseline/
├── mmdetection/
│ ├── cascade_rcnn_swin/
| | ├── cascade_rcnn_swin_large.py
| | ├── cascade_rcnn_swin.py
| | ├── cascade_rcnn.py
| | ├── dataset.py
| | ├── runtime.py
| | └── scheduler.py
│ ├── faster_rcnn/
| | ├── faster_rcnn_r152_fpn_dconv_c3-c5_1x_coco.py
| | ├── faster_rcnn_r152_fpn_1x_coco.py
| | ├── faster_rcnn_r50_fpn_1x_coco.py
| | ├── faster_rcnn_r50_fpn.py
| | ├── dataset.py
| | ├── default_runtime.py
| | └── schedule_1x.py
| └── inference.ipynb
├── ensemble/
│ └── Ensemble.ipynb
├── YOLOv5/
│ ├── data/
| | └── coco_trash.yaml
│ ├── runs/
| | ├── hyp.yaml
| | └── opt.yaml
└── inference.ipynb
eda/
├── eda.ipynb
└── eda-2.ipynb
utils/
|  └── inference_checker/
|    └── main.py
|  └── trainset_check/
└──  └── main.py

• baseline/mmdetection/cascade_rcnn: config files for Cascade R-CNN model
• baseline/mmdetection/faser_rcnn: config files for Faster R-CNN model
• baseline/mmdetection/inference.ipyb: inference code for mmdetection library
• baseline/ensemble/Ensemble.ipynb: ensemble code
• baseline/YOLOv5/data/coco_trash.yaml: converted Trash dataset to YOLO data form
• baseline/YOLOv5/runs: config file of our model
• baseline/YOLOv5/inferece.ipynb: inference code for YOLOv5 library
• eda/eda.ipynb: result of EDA
• eda/eda-2.ipynb: another result of EDA
• utils/inference_chcker/main.py: python program for visualization of result of inference
• utils/trainset_check/main.py: python program for visualization of training set
  

Requirements

• Linux version 4.4.0-59-generic
• Python >= 3.8.5
• PyTorch >= 1.7.1
• conda >= 4.9.2
• tensorboard >= 2.4.1

Hardware

• CPU: Intel(R) Xeon(R) Gold 5220 CPU @ 2.20GHz
• GPU: Tesla V100-SXM2-32GB

Reference

git clone https://github.com/open-mmlab/mmdetection.git : install mmdetection library
git clone https://github.com/ultralytics/yolov5.git : install YOLOv5 library.

Training

mmdetection

• mmdetection 라이브러리 clone 후 configs file 실행
  example
  
• cd mmdetection
  
• python tools/train.py baseline/mmdetection/cascade_rcnn/cascade_rcnn_swin_large.py
  

YOLOv5
example

• cd YOLOv5
  
• python train.py --img 1024 --batch 4 --epochs 50 --data baseline/data/coco_trash.yaml --weights yolov5x6.pt

Ensemble

• baseline/ensemble/Ensemble.ipynb 파일 실행
• target folder 생성 후 ensemble 하고자 하는 파일 target1.csv, target2.csv 형식으로 저장
• ipynb 파일 run (Weighted Boxes Fusion)

Tools

inference_checker

cd utils/inference_checker
python main.py
open submission file (command or control + O)

trainset_check

cd utils/treainset_check
python main.py