cross decode different areas using generalizing estimator

💬 Support & Discussions
1
  • MNE version: 1.2.1
  • operating system: macOS 12.6

I am using generalizing estimator where my model object is instantiated as:

tgm_fits[g][train_model_area] = GeneralizingEstimator(make_pipeline(StandardScaler(),PCA(n_components = 15), LinearDiscriminantAnalysis()), scoring='roc_auc', verbose=True)

train_scores_tgm[g][train_model_area] = tgm_fits[g][area].score(X = X_train_area, y = y_train)
test_scores_tgm[g][train_model_area] = tgm_fits[g][area].score(X = X_test_area, y = y_test)

I am trying to do something like this:

scores_cross_area_tgm[g][(train_model_area, test_area)] = tgm_fits[g][train_model_area].score(X = X_test_area_, y = y_test

This works fine when I train on a group of channels and test on the same group of channels.

But I want to cross-decode on a different group of channels, and the problem is that now we have a different number of channels, and the normalization and PCA step throw an error for test data for not having the same number of dimensions as the trained data.

Can I somehow bypass the preprocessing steps when I call .score() when the X is from a different group of sensors, and they are manually preprocessed to have PCA dimensions be 15.

Any advice would be very helpful!

2

as the feature space is expected to be the same between train and test you must have the same set of channels

I would suggest you subsample the channels so they match.

Alex

1 Like