cs231n-cyclone

This project estimates intensity and size of tropical cyclones from more than 70k frames collected worldwide.
Data is available at this webpage.

Requirements

Dependency can be installed using the following command:

pip install -r requirements.txt

Proposed model

Model Training

1. Overfit case

Training and validation data in the overfitting case has been prepared for you. Please go to the maxOverfit.ipynb file and preform interactive training there.

Notice don't run two sections in the Jupyter Notebook unless you want to re-generate the training and validation data: Read in info and get maxes, Calculate mean and std in training.

Another reminder is that the last two sections Playground and Calculate channel... in rolling manner is also for pre-processing and playing with data. They are relevenat in the training.

2. Normal case

Data has not been provided for this case, because they take up 60GB on disk. You need to generate the data by yourself.

Therefore, to perform training on a specified model, you have to do these commands (assume you have fulfilled all dependency requirements):

2.1 Create folders

You need to have two folders in ./normal/. That is ./normal/data and ./normal/results.

2.2 Generate training and validation data

python generate.py

Also, you can upload data files generated locally to cloud server

gcloud compute scp --recurse local/folder cloud/folder

2.3 Train

A use case is to train a Resnet34, for 100 epochs, with batch size 256, regularization strength 3e-3, initial learning rate 2e-3, step linear scheduler for learning rate decay and no center crop on images.

python train.py --modelSize 34 --epochTrain 100 --batchSize 256 --regStrength 3e-3 --lr 2e-3 --scheduler step

2.4 Results saving

The previous commandline arguments will create a folder ./results/resnet_34_100_0.003_0.002_step_False. 7 files exist in that folder. To download the folder to a local folder results, type the following in the Terminal:

gcloud compute scp --recurse your/cloud/server/directory/results/resnet_34_100_0.003_0.002_step_False your/local/directory/results

Citation

If you find this repository, e.g., the code and the datasets, useful in your research, please cite the following paper:

@inproceedings{chen2018rotation,
  title={Rotation-blended CNNs on a New Open Dataset for Tropical Cyclone Image-to-intensity Regression},
  author={Chen, Boyo and Chen, Buo-Fu and Lin, Hsuan-Tien},
  booktitle={Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery \& Data Mining},
  pages={90--99},
  year={2018},
  organization={ACM}
}