kaggle-understanding-clouds

Code for 3rd place solution in Kaggle Understanding Clouds from Satellite Images Challenge.

To read the brief description of the solution, please, refer to the Kaggle post. If you run into any trouble with the setup/code or have any questions please contact me at davidcao1991@gmail.com

Solution Overview

Archive Contents

.
./input                          : images and .csv files for mask labels
./output                         : trained models, logs and predictions
./kaggle-cloud-organization      : code

Software:

• Python 3.7.3
• CUDA 10.1
• cuddn 7602
• nvidia drivers v418.67

python packages are detailed separately in requirements.txt

Data setup:

Assumes the Kaggle API is installed, below are the shell commands used in each step, as run from the top level directory.

$ mkdir -p input
$ cd input
$ kaggle competitions download -c understanding_cloud_organization
$ unzip train_images_zip -d ./images
$ unzip test_images_zip -d ./images

Please make sure the unzipped train and test images are in the same folder named "images".

Data proprecessing

This will convert the raw images to 384x576 and build .csv file for 5-fold info. The 5-fold info .csv is already included in ./kaggle-cloud-organization/files/

$ cd kaggle-cloud-organization
$ python preprocessing_images.py
$ python make_folds.py

Model build

There are 4 steps to reproduce the model from scratch. Shell command to run each step is below.

1. Train seg1 model and predict.

$ bash ./kaggle-cloud-organization/run_seg1.sh

2. Train seg2 models and predict.

$ bash ./kaggle-cloud-organization/run_seg2.sh

3. Train cls model and predict.

$ bash ./kaggle-cloud-organization/run_cls.sh

4. Ensemble the final predictions

$ cd output
$ mkdir -p b5-Unet-inception-FPN-b7-Unet-b7-FPN-b7-FPNPL
$ mkdir -p ensemble
$ cd ..
$ python ./kaggle-cloud-organization/mask-ensemble-5fold.py
$ python ./kaggle-cloud-organization/2-stage-ensemble-5fold.py

The final prediction file can be found as ./output/ensemble/test_5fold_tta3_cls.csv