Seizure Prediction
- Competition page at Kaggle
- This is a proof-of-concept for applying deep learning techniques to EEG data converted into spectrograms.
- This code builds a separate model for each subject (there are three subjects).
These steps take about 4 hours on a system with 4 processors, a single GPU and a spinning hard-disk. Tested only on Ubuntu.
-
Download and install neon 1.6.0
git clone https://github.com/NervanaSystems/neon.git cd neon git checkout v1.6.0 make source .venv/bin/activate -
Verify neon installation
Make sure that this command does not result in any errors:
./examples/cifar10_msra.py -e1 -
Install prerequisites
pip install scipy sklearn scikits.audiolab -
Download the data files from Kaggle:
Save all files to a directory (referred to as /path/to/data below) and unzip the .zip files.
-
Clone this repository
git clone https://github.com/anlthms/sp-2016.git cd sp-2016 -
Train models and generate predictions
./run.sh /path/to/data /path/to/output 2>&1 | tee run.logwhere /path/to/data must contain the data subdirectories (train_1, train_2 etc.) as well as sample_submission.csv and /path/to/output is a new directory that will be created to store intermediate output files.
-
Evaluate predictions
Submit subm.csv to Kaggle
- The model requires 8GB of device memory.
- If using AWS, see slide 10 on [this deck] (https://github.com/anlthms/meetup2/blob/master/audio-pattern-recognition.pdf) for instructions on how to configure an EC2 instance.
- The first run takes longer due to conversion of .mat files into .wav files.
- Conversion of data to spectrograms is performed on the fly by neon.