Frequently Asked Questions

General MNE-Python issues

Help! I can’t get Python and MNE-Python working!

Check out our section on how to get Anaconda up and running over at the getting started page.

I still can’t get it to work!

For analysis talk, join the MNE mailing list. File specific feature requests or bug reports on GitHub. You can also chat with developers on Gitter.

How do I cite MNE?

If you use the implementations provided by the MNE software in your research, you should cite:

If you use the Python code you should cite as well:

    1. Gramfort, M. Luessi, E. Larson, D. Engemann, D. Strohmeier, C. Brodbeck, R. Goj, M. Jas, T. Brooks, L. Parkkonen, M. Hämäläinen, MEG and EEG data analysis with MNE-Python, Frontiers in Neuroscience, Volume 7, 2013, ISSN 1662-453X, [DOI]

To cite specific versions of the software, you can use the DOIs provided by Zenodo.

You should as well cite the related method papers, some of which are listed in Related publications.

I’m not sure how to do X analysis step with my Y data...

Knowing “the right thing” to do with EEG and MEG data is challenging. We use the MNE mailing list to discuss how to deal with different bits of data. It’s worth searching the archives to see if there have been relevant discussions before.

I think I found a bug, what do I do?

Please report any problems you find while using MNE-Python to the GitHub issues page. Try using the latest master version to see if the problem persists before reporting the bug, as it may have been fixed since the latest release.

It is helpful to include system information with bug reports, so it can be useful to include the output of the mne.sys_info() command when reporting a bug, which should look something like this:

>>> import mne
>>> mne.sys_info()  
Platform:      Linux-4.2.0-27-generic-x86_64-with-debian-jessie-sid
Python:        2.7.11 |Continuum Analytics, Inc.| (default, Dec  6 2015, 18:08:32)  [GCC 4.4.7 20120313 (Red Hat 4.4.7-1)]
Executable:    /home/larsoner/miniconda/bin/python

mne:           0.12.dev0
numpy:         1.10.2 {lapack=mkl_lapack95_lp64, blas=mkl_intel_lp64}
scipy:         0.16.1
matplotlib:    1.5.1

sklearn:       Not found
nibabel:       Not found
nitime:        Not found
mayavi:        Not found
nose:          1.3.7
pandas:        Not found
pycuda:        Not found
skcuda:        Not found

Why is it dangerous to “pickle” my MNE-Python objects and data for later use?

Pickling data and MNE-Python objects for later use can be tempting due to its simplicity and generality, but it is usually not the best option. Pickling is not designed for stable persistence, and it is likely that you will not be able to read your data in the not-too-distant future. For details, see:

MNE-Python is designed to provide its own file saving formats (often based on the FIF standard) for its objects usually via a save method or write_* method, e.g.,, mne.write_evokeds(), If you have some data that you want to save but can’t figure out how, shoot an email to the MNE mailing list or post it to the GitHub issues page.

If you want to write your own data to disk (e.g., subject behavioral scores), we strongly recommend using h5io, which is based on the HDF5 format and h5py, to save data in a fast, future-compatible, standard format.

I downloaded a dataset once, but MNE-Python is asking to download it again. Why?

The default location for the MNE-sample data is ~/mne_data. If you downloaded data and an example asks you whether to download it again, make sure the data reside in the examples directory and that you run the script from its current directory:

$ cd examples/preprocessing

Then in Python you can do:

In [1]: %run

See Datasets for a list of all available datasets and some advanced configuration options, e.g. to specify a custom location for storing the datasets.

A function uses multiple CPU cores even though I didn’t tell it to. Why?

Ordinarily in MNE-python the parallel module is used to deploy multiple cores via the n_jobs variable. However, functions like mne.preprocessing.maxwell_filter() that use scipy.linalg do not have an n_jobs flag but may still use multiple cores. This is because scipy.linalg is built with linear algebra libraries that natively support multithreading:

To control how many cores are used for linear-algebra-heavy functions like mne.preprocessing.maxwell_filter(), you can set the OMP_NUM_THREADS or OPENBLAS_NUM_THREADS environment variable to the desired number of cores for MKL or OpenBLAS, respectively. This can be done before running Python, or inside Python you can achieve the same effect by, e.g.:

>>> import os
>>> num_cpu = '4' # Set as a string
>>> os.environ['OMP_NUM_THREADS'] = num_cpu

This must be done before running linear algebra functions; subsequent changes in the same Python session will have no effect.

Resampling and decimating data

What are all these options for resampling, decimating, and binning data?

There are many functions in MNE-Python for changing the effective sampling rate of data. We’ll discuss some major ones here, with some of their implications:

  • is used to resample (typically downsample) raw data. Resampling is the two-step process of applying a low-pass FIR filter and subselecting samples from the data.

    Using this function to resample data before forming mne.Epochs for final analysis is generally discouraged because doing so effectively loses precision of (and jitters) the event timings, see this gist as a demonstration. However, resampling raw data can be useful for (at least):

    • Computing projectors in low- or band-passed data
    • Exploring data
  • decimates data without low-passing, but is only used for fitting a statistical model to the data.

  • mne.Epochs.decimate(), which does the same thing as the decim parameter in the mne.Epochs constructor, sub-selects every \(N^{th}\) sample before and after each event. This should only be used when the raw data have been sufficiently low-passed e.g. by to avoid aliasing artifacts.

  • mne.Epochs.resample(), mne.Evoked.resample(), and mne.SourceEstimate.resample() all resample data. This process avoids potential aliasing artifacts because the resampling process applies a low-pass filter. However, this filtering introduces edge artifacts. Edge artifacts also exist when using, but there the edge artifacts are constrained to two times: the start and end of the recording. With these three methods, edge artifacts are introduced to the start and end of every epoch of data (or the start and end of the mne.Evoked or mne.SourceEstimate data), which often has a more pronounced effect on the data.

  • mne.SourceEstimate.bin() can be used to decimate, with or without “binning” (averaging across data points). This is equivalent to applying a moving-average (boxcar) filter to the data and decimating. A boxcar in time is a sinc in frequency, so this acts as a simplistic, non-ideal low-pass filter; this will reduce but not eliminate aliasing if data were not sufficiently low-passed. In the case where the “filter” or bin-width is a single sample (i.e., an impulse) this operation simplifies to decimation without filtering.

Resampling raw data is taking forever! What do I do? was significantly sped up for version 0.12 by using the parameter npad=='auto'. Try it, it might help!

If you have an NVIDIA GPU you could also try using CUDA, which can sometimes speed up filtering and resampling operations by an order of magnitude.

Inverse Solution

How should I regularize the covariance matrix?

The estimated covariance can be numerically unstable and tends to induce correlations between estimated source amplitudes and the number of samples available. The MNE manual therefore suggests to regularize the noise covariance matrix (see Regularization of the noise-covariance matrix), especially if only few samples are available. Unfortunately it is not easy to tell the effective number of samples, hence, to choose the appropriate regularization. In MNE-Python, regularization is done using advanced regularization methods described in [1]. For this the ‘auto’ option can be used. With this option cross-validation will be used to learn the optimal regularization:

>>> import mne
>>> epochs = mne.read_epochs(epochs_path) 
>>> cov = mne.compute_covariance(epochs, tmax=0., method='auto') 

This procedure evaluates the noise covariance quantitatively by how well it whitens the data using the negative log-likelihood of unseen data. The final result can also be visually inspected. Under the assumption that the baseline does not contain a systematic signal (time-locked to the event of interest), the whitened baseline signal should be follow a multivariate Gaussian distribution, i.e., whitened baseline signals should be between -1.96 and 1.96 at a given time sample. Based on the same reasoning, the expected value for the global field power (GFP) is 1 (calculation of the GFP should take into account the true degrees of freedom, e.g. ddof=3 with 2 active SSP vectors):

>>> evoked = epochs.average() 
>>> evoked.plot_white(cov) 

This plot displays both, the whitened evoked signals for each channels and the whitened GFP. The numbers in the GFP panel represent the estimated rank of the data, which amounts to the effective degrees of freedom by which the squared sum across sensors is divided when computing the whitened GFP. The whitened GFP also helps detecting spurious late evoked components which can be the consequence of over- or under-regularization.

Note that if data have been processed using signal space separation (SSS) [2], gradiometers and magnetometers will be displayed jointly because both are reconstructed from the same SSS basis vectors with the same numerical rank. This also implies that both sensor types are not any longer linearly independent.

These methods for evaluation can be used to assess model violations. Additional introductory materials can be found here.

For expert use cases or debugging the alternative estimators can also be compared:

>>> covs = mne.compute_covariance(epochs, tmax=0., method='auto', return_estimators=True) 
>>> evoked = epochs.average() 
>>> evoked.plot_white(covs) 

This will plot the whitened evoked for the optimal estimator and display the GFPs for all estimators as separate lines in the related panel.

Morphing data

Should I morph my source estimates using morph or morph_precomputed?

The two functions mne.SourceEstimate.morph() and mne.SourceEstimate.morph_precomputed() perform the same operation: taking surface-based source space data from one subject and morphing it to another using a smoothing procedure. However, they can take different amounts of time to perform the computation.

To use mne.SourceEstimate.morph_precomputed(), you must first precompute a morphing matrix with mne.compute_morph_matrix() which can take some time, but then the actual morphing operation carried out by mne.SourceEstimate.morph_precomputed() is very fast, even for mne.SourceEstimate objects with many time points. The method mne.SourceEstimate.morph(), by contrast, smooths the data by operating directly on the data, which can be very slow with many time points. If there are thousands of time points, then mne.SourceEstimate.morph_precomputed() will be much faster; if there are a few time points, then mne.SourceEstimate.morph() will be faster. For data sizes in between, we advise testing to determine which is best, although some developers choose to always use mne.SourceEstimate.morph_precomputed() since it will rarely take a long time.


[1]Engemann D. and Gramfort A. (2015) Automated model selection in covariance estimation and spatial whitening of MEG and EEG signals, vol. 108, 328-342, NeuroImage.
[2]Taulu, S., Simola, J., Kajola, M., 2005. Applications of the signal space separation method. IEEE Trans. Signal Proc. 53, 3359–3372.