the 3rd place solution code of Kaggle TReNDS Neuroimaging (https://www.kaggle.com/c/trends-assessment-prediction/overview)

Please read this post on Kaggle Discussion.

I mainly use local machine with following spec.

OS: Ubuntu 18.04 LTS
CPU: Intel(R) Core(TM) i7-9700K CPU @ 3.60GHz
GPU: GeForce RTX 2080 x2

You should install docker-ce. If you don't have docker installed, please refer to this page to install it.

Plese download data to ./input from https://www.kaggle.com/c/trends-assessment-prediction/data and unzip it.

$ docker build ./run_image/gpu -t kaggle/pytorch:trends

• compute statistic features of voxel and adjacency matrix of components from fnc data.

$ sh bin/preprocess.sh

I used Weight & Biases for management of experiments and Google Cloud Storage for saving result. If you want to use this function, please set yamls/store/*.yaml as below.

  wandb_project: [your wandb project name]
  gcs_project: [your gcp project name]
  bucket_name: [your gcs bucket name]

1. Train NN models

   $ sh bin/train_nn.sh
   

2. make_cnn_feature

   $ sh bin/make_cnn_feature.sh
   

3. train simple models

   $ sh bin/train_simple_models.sh
   

$ sh bin/stacking_lgbm.sh && sh bin/stacking_svm.sh

$ sh bin/blending_stacking.sh

The final output is generated in ./output/blending_lgbm_svm_stacking.csv.