alexstar55 / kaggle_airbus_ship_detection Goto Github PK

This project is for Kaggle competiton Airbus Ship Detection Challenge.

It can help you quickly get a baseline solution, which is not bad.

Kaggle新手银牌（21st）：Airbus Ship Detection 卫星图像分割检测

用Mask R-CNN训练自己的COCO数据集（Detectron）

辅助操作指南：Docker使用、镜像制作、Demo运行...

airbus                         
├─0_rle_to_coco                0、turn rle to coco
│  └─pycococreatortools
|
├─1_detectron_infer            1、files needed to be changed in detectron
|  ├─dataset_catalog.py            # ./detectron/datasets/dataset_catalog.py
│  ├─dummy_datasets.py             # ./detectron/datasets/dummy_datasets.py 
│  └─infer_airbus.py               # ./tools/infer_simple.py    
|
├─2_model                      2、model and trainning log
│  ├─log                           log and visualization script
│  └─model                         configure file and .pkl (.pkl not be uploaded)
|
└─3_submit                     3、generate your submission
   └─csv                           reference .csv file

Run codes in ./0_rle_to_coco. The guide has been written in markdwon file ./0_rle_to_coco/README.md

My codes are based on Detectron. So before using it, you need to install caffe2, which is quite troublesome. You can use my docker image, which is a little out of date, by the following command:

$ docker pull pascal1129/detectron:caffe2_cuda9_aliyun

In order to get the latest docker image, you can build the latest image with the official dockerfile: Detectron/docker/Dockerfile.

My codes are in the folder ./1_detectron_infer/, you can replace the origin files in detectron with my codes.

Confirm the .yaml file in ./2_model/model/ and start training. In addition, remember to use |tee command, so you can get the log file like ./2_model/log/20181103.log

Run ./2_model/analyse_log.py, then you can get the visualization picture.

Run ./3_submit/get_final_csv.py.