0. Quick Start

$ sh run.sh

1. Preparing

1.1. Environments

1. System

   • OS

   $ uname -a
   Linux 78709cb2ea9c 4.4.0-59-generic #80-Ubuntu SMP Fri Jan 6 17:47:47 UTC 2017 x86_64 x86_64 x86_64 GNU/Linux
   

   커널: 4.4.0-59-generic 버전
   OS: x86 계열의 64bit 운영체제

   • CPU

   $ cat /proc/cpuinfo 
   Intel(R) Xeon(R) Gold 5120 CPU @ 2.20GHz
   

   • GPU
     • aistage 에서 제공받은 v100 GPU (NVIDIA 테슬라 Tesla V100 32G GPU)

2. Python

   • version: 3.7.11

1.2. Prepare Dataset

Get the trash dataset

$ wget https://aistages-prod-server-public.s3.amazonaws.com/app/Competitions/000076/data/data.tar.gz
$ tar xvfz data.tar.gz
$ mv ./data/ ./dataset
$ rm ./data.tar.gz
$ rm -rf ./dataset/*/.*.jpg

Check the dataset structure and format(COCO format)

./
└─dataset/
    ├─train/
    |   ├─0000.jpg
    |   ├─ ...
    |   └─48xx.jpg
    ├─test/
    |   ├─0000.jpg
    |   ├─ ...
    |   └─4870.jpg
    ├─train.json
    └─test.json

Trash Dataset Copy Rights: Naver Connect
Trash Dataset Copy Rights License: CC-BY-2.0

1.3. Prepare Libraries

1. MMDetection

   $ cd ./mmdetection
   $ pip install -v -e . 

   reference: MMDetection github

2. YOLOv5

   $ cd ./yolov5
   $ pip install -r requirements.txt

   reference: YOLOv5 github

3. convert2Yolo

   $ git clone https://github.com/ssaru/convert2Yolo.git
   $ cd ./convert2Yolo
   $ pip3 install -r requirements.txt

   reference: conver2Yolo github

4. WBF (Weighted Boxes Fusion)

   $ pip install ensemble-boxes

   reference: WBF github

1.4. Completed Structure

./
├─dataset/
|    ├─train/
|    |   ├─0000.jpg
|    |   ├─ ...
|    |   └─48xx.jpg
|    ├─test/
|    |   ├─0000.jpg
|    |   ├─ ...
|    |   └─4870.jpg
|    ├─train.json
|    └─test.json
|
├─mmdetection/
├─yolov5/
├─convert2Yolo
|
├─submissions_for_single_model/
|    ├─submission_yolov5s.csv
|    ├─submission_faster_rcnn.csv
|    ├─...
|    └─submission_swin.csv
|
├─run.py 
├─run.sh
|
└─submission.csv # Result of "python run.py" or "sh run.sh"

2. Training

2.1. Before training, convert dataset format (From COCO To YOLO)

• Prepare the *.names file as below.

   $ cat ./dataset/trash_coco.names
   General trash
   Paper
   Paper pack
   Metal
   Glass
   Plastic
   Styrofoam
   Plastic bag
   Battery
   Clothing

• Convert dataset format by using convert2Yolo

   $ cd ./convert2Yolo
   $ python3 example.py \
   	--datasets COCO \
   	--img_path ../dataset/ \
   	--label ../dataset/train.json \
   	--convert_output_path ../dataset/ \
   	--img_type ".jpg" \
   	--manifest_path ../dataset \
   	--cls_list_file ../dataset/trash_coco.names

• Then, we have the following structure.

   ./
   ├─dataset/
   |    ├─train/
   |    |   ├─0000.jpg
   |    |   ├─0000.txt # YOLO format
   |    |   ├─ ...
   |    |   ├─48xx.jpg
   |    |   └─48xx.txt # YOLO format
   |    ├─test/
   |    |   ├─0000.jpg
   |    |   ├─ ...
   |    |   └─4870.jpg
   |    ├─train.json # COCO format
   |    └─test.json  # COCO format
   |
   ├─mmdetection/
   ├─yolov5/
   ├─convert2Yolo
   |
   ├─submissions_for_single_model/
   |    ├─submission_yolov5s.csv
   |    ├─submission_faster_rcnn.csv
   |    ├─...
   |    └─submission_swin.csv
   |
   ├─run.py 
   ├─run.sh
   |
   └─submission.csv # Result of "python run.py" or "sh run.sh"
  

2.2. Train 4 models

1. detectors-cascade-rcnn-r50

   $ cd ./mmdetection
   $ python tools/train.py \
   	configs/trash/detectors_cascade_rcnn_r50.py

2. cascade-rcnn-swin-base

   $ cd ./mmdetection
   $ python tools/train.py \
   	configs/trash/swin/cascade_rcnn_swin_base_fpn.py

3. cascade-rcnn-r50-fpn

   $ cd ./mmdetection
   $ python tools/train.py \
       configs/trash/cascade_rcnn_r50_fpn.py 

4. YOLOv5x6

   • Before training

   $ cd ./yolov5
   $ pip install -r requirements.txt

   • Train pre-trained model for 10-epochs

   $ python train.py \
       --img 1024 --batch 4 --epochs 10 --data custom.yaml \
       --weights yolov5x6.pt --cache --name 10epoch

   • Train pre-trained model for 20-epochs

   $ python train.py \
       --img 1024 --batch 4 --epochs 20 --data custom.yaml \
       --weights yolov5x6.pt --cache --name 20epochs

3. Inference and Make Submission

1. detectors-cascade-rcnn-r50

   $ cd ./mmdetection
   $ python tools/test.py \
   	confings/trash/detectors_cascade_rcnn_r50.py \
   	work_dirs/detectors_cascade_rcnn_r50/best*.pth \
   	--out work_dirs/detectors_cascade_rcnn_r50/best.pkl
   $ python tools/pkl_to_submission.py \
   	--pkl work_dirs/detectors_cascade_rcnn_r50/best.pkl \
   	--csv work_dirs/detectors_cascade_rcnn_r50.csv
   $ cp work_dirs/detectors_cascade_rcnn_r50.csv \
   	../submissions_for_single_model/submission_detectors_cascade_rcnn_r50.csv

2. cascade-rcnn-r50-fpn

   $ cd ./mmdetection
   $ python confings/trash/cascade_rcnn_r50_fpn.py \
   	work_dirs/cascade_rcnn_r50_fpn/best*.pth \
   	--out work_dirs/cascade_rcnn_r50_fpn/best.pkl
   $ python tools/pkl_to_submission.py \
   	--pkl work_dirs/cascade_rcnn_r50_fpn/best.pkl \
   	--csv work_dirs/cascade_rcnn_r50_fpn.csv
   $ cp work_dirs/cascade_rcnn_r50_fpn.csv \
   	../submissions_for_single_model/submission_cascade_rcnn_r50_fpn.csv

3. cascade-rcnn-swin-base

   $ cd ./mmdetection
   $ python confings/trash/swin/cascade_rcnn_swin_base_fpn.py \
   	work_dirs/cascade_rcnn_swin_base_fpn/best*.pth \
   	--out work_dirs/cascade_rcnn_swin_base_fpn/best.pkl
   $ python tools/pkl_to_submission.py \
   	--pkl work_dirs/cascade_rcnn_swin_base_fpn/best.pkl \
   	--csv work_dirs/cascade_rcnn_swin_base_fpn.csv
   $ cp work_dirs/cascade_rcnn_swin_base_fpn.csv \
   	../submissions_for_single_model/submission_cascade_rcnn_r50_fpn.csv

4. YOLOv5x6

   • Ensemble 2 models (Use YOLOv5 built-in ensemble)

      $ cd ./yolov5
      $ python detect.py \
      	--weights ./weights/runs/10epoch/train/best.pt ./weights/runs/20epoch/train/best.pt \
      	--source ../dataset/test/ \
      	--imgsz 1024 \
      	--max-det 100 \
      	--device 0 \
      	--classes 0 1 2 3 4 5 6 7 8 9 \
      	--save-txt --save-conf \
      	--nosave \
      	--augment

   • Convert labels to CSV file

      $ cd ./yolov5
      $ python convertcsv.py

4. Ensembling

• WBF (Weighted Boxes Fusion)
  • Config Format

     {
     	"csvs": ["./output_0.csv", 
     	         "./output_1.csv",
     	         "./output_2.csv",
     	         "./output_3.csv",
     	         "./output_4.csv",
     	         "./output_5.csv"
     	         ],
     	"save_path": "./result_outputs.csv",
     	"ensemble_mode": "wbf",
     	"weights": "None",
     	"iou_thr": 0.5,
     	"skip_box_thr": 0.0001,
     	"sigma": 0.1,
     	"img_size": 1024
     }

    csvs: 단일 모델이 예측한 결과 csv 파일들
    ensemble_mode: Choose the ensemble mode (wbf, nms, nmw, snms)

5. Get the submission.csv

$ python ensemble.py

6. Participants

이름	김서기	김승훈	손지아	이상은	조익수	배민한
역할