{
  "nbformat_minor": 0, 
  "nbformat": 4, 
  "cells": [
    {
      "execution_count": null, 
      "cell_type": "code", 
      "source": [
        "%matplotlib inline"
      ], 
      "outputs": [], 
      "metadata": {
        "collapsed": false
      }
    }, 
    {
      "source": [
        "\n# Linear classifier on sensor data with plot patterns and filters\n\n\nDecoding, a.k.a MVPA or supervised machine learning applied to MEG and EEG\ndata in sensor space. Fit a linear classifier with the LinearModel object\nproviding topographical patterns which are more neurophysiologically\ninterpretable [1]_ than the classifier filters (weight vectors).\nThe patterns explain how the MEG and EEG data were generated from the\ndiscriminant neural sources which are extracted by the filters.\nNote patterns/filters in MEG data are more similar than EEG data\nbecause the noise is less spatially correlated in MEG than EEG.\n\nReferences\n----------\n\n.. [1] Haufe, S., Meinecke, F., G\u00f6rgen, K., D\u00e4hne, S., Haynes, J.-D.,\n       Blankertz, B., & Bie\u00dfmann, F. (2014). On the interpretation of\n       weight vectors of linear models in multivariate neuroimaging.\n       NeuroImage, 87, 96\u2013110. doi:10.1016/j.neuroimage.2013.10.067\n\n"
      ], 
      "cell_type": "markdown", 
      "metadata": {}
    }, 
    {
      "execution_count": null, 
      "cell_type": "code", 
      "source": [
        "# Authors: Alexandre Gramfort <alexandre.gramfort@telecom-paristech.fr>\n#          Romain Trachel <trachelr@gmail.com>\n#          Jean-Remi King <jeanremi.king@gmail.com>\n#\n# License: BSD (3-clause)\n\nimport mne\nfrom mne import io, EvokedArray\nfrom mne.datasets import sample\nfrom mne.decoding import Vectorizer, get_coef\n\nfrom sklearn.preprocessing import StandardScaler\nfrom sklearn.linear_model import LogisticRegression\nfrom sklearn.pipeline import make_pipeline\n\n# import a linear classifier from mne.decoding\nfrom mne.decoding import LinearModel\n\nprint(__doc__)\n\ndata_path = sample.data_path()"
      ], 
      "outputs": [], 
      "metadata": {
        "collapsed": false
      }
    }, 
    {
      "source": [
        "Set parameters\n\n"
      ], 
      "cell_type": "markdown", 
      "metadata": {}
    }, 
    {
      "execution_count": null, 
      "cell_type": "code", 
      "source": [
        "raw_fname = data_path + '/MEG/sample/sample_audvis_filt-0-40_raw.fif'\nevent_fname = data_path + '/MEG/sample/sample_audvis_filt-0-40_raw-eve.fif'\ntmin, tmax = -0.1, 0.4\nevent_id = dict(aud_l=1, vis_l=3)\n\n# Setup for reading the raw data\nraw = io.read_raw_fif(raw_fname, preload=True)\nraw.filter(.5, 25)\nevents = mne.read_events(event_fname)\n\n# Read epochs\nepochs = mne.Epochs(raw, events, event_id, tmin, tmax, proj=True,\n                    decim=4, baseline=None, preload=True)\n\nlabels = epochs.events[:, -1]\n\n# get MEG and EEG data\nmeg_epochs = epochs.copy().pick_types(meg=True, eeg=False)\nmeg_data = meg_epochs.get_data().reshape(len(labels), -1)"
      ], 
      "outputs": [], 
      "metadata": {
        "collapsed": false
      }
    }, 
    {
      "source": [
        "Decoding in sensor space using a LogisticRegression classifier\n\n"
      ], 
      "cell_type": "markdown", 
      "metadata": {}
    }, 
    {
      "execution_count": null, 
      "cell_type": "code", 
      "source": [
        "clf = LogisticRegression()\nscaler = StandardScaler()\n\n# create a linear model with LogisticRegression\nmodel = LinearModel(clf)\n\n# fit the classifier on MEG data\nX = scaler.fit_transform(meg_data)\nmodel.fit(X, labels)\n\n# Extract and plot spatial filters and spatial patterns\nfor name, coef in (('patterns', model.patterns_), ('filters', model.filters_)):\n    # We fitted the linear model onto Z-scored data. To make the filters\n    # interpretable, we must reverse this normalization step\n    coef = scaler.inverse_transform([coef])[0]\n\n    # The data was vectorized to fit a single model across all time points and\n    # all channels. We thus reshape it:\n    coef = coef.reshape(len(meg_epochs.ch_names), -1)\n\n    # Plot\n    evoked = EvokedArray(coef, meg_epochs.info, tmin=epochs.tmin)\n    evoked.plot_topomap(title='MEG %s' % name)"
      ], 
      "outputs": [], 
      "metadata": {
        "collapsed": false
      }
    }, 
    {
      "source": [
        "Let's do the same on EEG data using a scikit-learn pipeline\n\n"
      ], 
      "cell_type": "markdown", 
      "metadata": {}
    }, 
    {
      "execution_count": null, 
      "cell_type": "code", 
      "source": [
        "X = epochs.pick_types(meg=False, eeg=True)\ny = epochs.events[:, 2]\n\n# Define a unique pipeline to sequentially:\nclf = make_pipeline(\n    Vectorizer(),                       # 1) vectorize across time and channels\n    StandardScaler(),                   # 2) normalize features across trials\n    LinearModel(LogisticRegression()))  # 3) fits a logistic regression\nclf.fit(X, y)\n\n# Extract and plot patterns and filters\nfor name in ('patterns_', 'filters_'):\n    # The `inverse_transform` parameter will call this method on any estimator\n    # contained in the pipeline, in reverse order.\n    coef = get_coef(clf, name, inverse_transform=True)\n    evoked = EvokedArray(coef, epochs.info, tmin=epochs.tmin)\n    evoked.plot_topomap(title='EEG %s' % name[:-1])"
      ], 
      "outputs": [], 
      "metadata": {
        "collapsed": false
      }
    }
  ], 
  "metadata": {
    "kernelspec": {
      "display_name": "Python 2", 
      "name": "python2", 
      "language": "python"
    }, 
    "language_info": {
      "mimetype": "text/x-python", 
      "nbconvert_exporter": "python", 
      "name": "python", 
      "file_extension": ".py", 
      "version": "2.7.13", 
      "pygments_lexer": "ipython2", 
      "codemirror_mode": {
        "version": 2, 
        "name": "ipython"
      }
    }
  }
}