Skip to content

Latest commit

 

History

History
24 lines (24 loc) · 1.19 KB

README.md

File metadata and controls

24 lines (24 loc) · 1.19 KB
Raw

MCG+： An improved motor imagery classification based on Cosine similarity

This project is the for the paper: MCGNet+: An improved motor imagery classification based on Cosine similarity

Prerequistes

  • Python 3.6
  • Tensorflow 1.12.0
  • Keras 2.2.4
  • numpy 1.15.4
  • scipy 1.1.0
  • scikit-learn 0.21.3
  • the code is built upon the tensorflow framework

How to run

dataset

  • the dataset is come from the BNCI-competition2014, which can be load in:http://moabb.neurotechx.com/docs/generated/moabb.datasets.BNCI2014001.html
  • the dataset need to be preprocessed, we didn't use the raw data, instead we use the DE features, and the code for preprocessing the dataset and extracting the DE features can be found in the directory : process/process_data.py
  • we use the mutual information to generate the adjency matrix, this part can be find in: process/feature_generate.py

the configuration

the settings of the model can be found in: config/run.config

loss setting

In our paper, we use a loss refine method, and in the tensorflow framework the loss should be loss = abs(loss-0.8), which can be achieved by changing the: tf/Lib/site-packages/keras/engine/training.py.

run

run

python train.py