mlearnx/EEG-Finger-Predictor

First project of the EPFL Spring 2018 Deep Learning Class

EEG-Finger-Predictor

First project of the EPFL Spring 2018 Deep Learning Class

Dependencies

• Python 3.5.4
• PyTorch 3.0.1
• Numpy 1.13.3
• Matplotlib 2.0.2

Folder structure

• src/: Path to all the source code
  • data_handler.py: Contains functionality to load, preprocess and augment the dataset
  • dlc_bci.py: Course provided loader of the BCI dataset
  • models.py: Contains the PyTorch LSTM and CNN model architectures
  • plots.py: Contains functionality to plot training and cross validation loss and accuracy curves
  • test.py: Script to train and test best CNN architecture
  • train_best_lstm.py: Script to train and test best LSTM architecture
  • train_cnn.py: Gridsearch running cross validation over parameter configurations to find best CNN architecture
  • train_lstm.py: Gridsearch running cross validation over parameter configurations to find best LSTM architecture
  • train_mlp.py: Multilayer perceptron baseline
  • train_utils.py: Contains training and K-fold cross validation helper functions
• data_bci/: Path where data will automatically be downloaded to on the first run
• README.md

How to run test files

To get the same numerical results as in the report, run the models on a GPU.

To run file that returns the train and test accuracies / losses of the best CNN model run:

$ cd src && python test.py

To run file that returns the train and test accuracies / losses of the best LSTM model run (Takes very long to train if not on a GPU!):

$ cd src && python train_best_lstm.py