Harmful Brain Activity Classification (HBAC)

This repo contains my solution for the Kaggle HMS - Harmful Brain Activity Classification competition.

This first commit contains my solution and the steps to reproduce, but future versions will be modified as I extend this project a little bit over the next couple of months to explore some ideas on interpretability, generalizability, and to reproduce other competitors solutions.

Reproducability

To reproduce all of the steps that I took for my final solution, first clone this library with:

$ git clone https://github.com/ryanirl/hbac.git
$ cd hbac

Then, download the Kaggle dataset and then unpack it into ./data/ (such that for example ./data/train_eeg/ is a valid directory).

Once this has been completed, you'll need to preprocess the dataset by running the following commands.

Note

This will take ~25 minutes to complete depending on your hardware.

$ preprocess.py -i ./data/ --modality eeg --num_workers 4 
$ preprocess.py -i ./data/ --modality eeg_spectrogram --num_workers 4
$ preprocess.py -i ./data/ --modality spectrogram --num_workers 4

Finally, once the data is downloaded and preprocessed, you can start training the models. Training a single fold takes about 4 hours to train from scratch. Since the dataset had high variance, and performance greatly improved with ensembling, I chose to use a 10-fold GroupKFold ensemble for my final solution. Training was originally performed on a single GTX 1070 8GB and takes ~40 hours to complete (4 hours per fold times 10 folds). To see the results of each model read my writeup.

Note

In retrospect, very similar performance can be achieved by using just the raw 1D EEG model without the mid and ekg information. Additionally, maybe including the EEG spectrograms was a bit overkill.

To train the each model, run the following commands:

$ sh scripts/train_eeg.sh
$ sh scripts/train_eeg_spec.sh
$ sh scripts/train_spectrogram.sh
$ sh scripts/train_multimodal.sh

Note: The above scripts will train a single fold, not 10 folds. If you want to train your will need to perform the following changes:

# Remove these two lines
train_stage_1 0 10
train_stage_2 0 10

# Replace them with this, which will train on all folds.
for i in `seq 0 9`
do
    train_stage_1 "$i" 10
    train_stage_2 "$i" 10
done

---

Unless stated otherwise in specific file, the license is Apache 2.0.