## General MNE-Python issues¶

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

Check out our installation instructions for Python and MNE-Python.

### I still can’t get it to work!¶

See Getting help.

### I can’t get Mayavi/3D plotting to work under Windows¶

If Mayavi plotting in Jupyter Notebooks doesn’t work well, using the IPython magic %gui qt after importing MNE/Mayavi/PySurfer should help.

from mayavi import mlab
%gui qt

### Python runs on macOS extremely slow even on simple commands!¶

Python uses some backends that interfere with the macOS energy saver when using an IDE such as Spyder or PyCharm. To test it, import time and run:

start = time.time(); time.sleep(0.0005); print(time.time() - start)

If it takes several seconds you can either:

• Install the module appnope and run in your script:

import appnope
appnope.nope()

• Change the configuration defaults by running in your terminal:

$defaults write org.python.python NSAppSleepDisabled -bool YES ### How do I cite MNE?¶ ### I’m not sure how to do X analysis step with my Y data…¶ Knowing “the right thing” to do with EEG and MEG data is challenging. We use the MNE Forum to discuss analysis strategies for different kinds of data. It’s worth searching the archives to see if there have been relevant discussions in the past, but don’t hesitate to ask a new question if the answer isn’t out there already. ### I think I found a bug, what do I do?¶ When you encounter an error message or unexpected results, it can be hard to tell whether it happened because of a bug in MNE-Python, a mistake in user code, a corrupted data file, or irregularities in the data itself. Your first step when asking for help should be the MNE Forum, not GitHub. This bears repeating: the GitHub issue tracker is not for usage help — it is for software bugs, feature requests, and improvements to documentation. If you open an issue that contains only a usage question, we will close the issue and direct you to the forum. If you’re pretty sure the problem you’ve encountered is a software bug (not bad data or user error): • Make sure you’re using the most current version. You can check it locally at a shell prompt with:$ mne sys_info

which will also give you version info about important MNE-Python dependencies.

• If you’re already on the most current version, if possible try using the latest development version, as the bug may have been fixed already since the latest release. If you can’t try the latest development version, search the GitHub issues page to see if the problem has already been reported and/or fixed.

• Try to replicate the problem with one of the MNE sample datasets. If you can’t replicate it with a built-in dataset, provide a link to a small, anonymized portion of your data that does yield the error.

If the problem persists, open a new issue and include the smallest possible code sample that replicates the error you’re seeing. Paste the code sample into the issue, with a line containing three backticks (`) above and below the lines of code. This minimal working example should be self-contained, which means that MNE-Python contributors should be able to copy and paste the provided snippet and replicate the bug on their own computers.

### 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.io.Raw.save(), mne.Epochs.save(), mne.write_evokeds(), mne.SourceEstimate.save(). If you have some data that you want to save but can’t figure out how, post to the MNE Forum or 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.

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:

events

## 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:

• mne.io.Raw.resample() 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

• mne.preprocessing.ICA.fit() 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 mne.io.Raw.filter() 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 mne.io.Raw.resample(), 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?¶

mne.io.Raw.resample() has a parameter npad=='auto'. This is the default, but if you’ve changed it you could try changing it back to 'auto', it might help.

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

## Forward and 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. It is thus suggested 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
>>> 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 GFP for all estimators as separate lines in the related panel.

### My watershed BEM meshes look incorrect¶

After using mne watershed_bem or mne.bem.make_watershed_bem() you might find that the BEM meshes for the brain, inner skull, outer skull, and/or scalp surfaces do not look correct in mne.viz.plot_alignment() and mne.viz.plot_bem().

MNE relies on FreeSurfer’s mri_watershed to compute the BEM meshes. Freesurfer’s watershed bem strategy is to:

1. Compute the outer skin (scalp) surface

2. Shrink outer skin inward make the “outer skull”

3. Compute brain surface

4. Expand brain surface outward to make the “inner skull”

A common problem is to see:

the surface inner skull is not completely inside surface outer skull

When looking at the meshes, the inner skull surface (expanded brain surface) will have defects, and these defects will protrude into the outer skull surface (shrunken scalp surface). In these cases, you can try (in rough ascending order of difficulty):

1. Changing the --preflood / -p parameter in mne watershed_bem.

2. Changing the --atlas and --gcaatlas options of mne watershed_bem.

3. Manually editing the meshes (see this tutorial).

4. Manually running mri_watershed with various FreeSurfer flags (e.g., -less to fix the output).

5. Going farther back in your Freesurfer pipeline to fix the problem. In particular, mri/brainmask.mgz could be incorrectly generated by the autorecon1 step and contain some dura and/or skull within the brain mask. You can check by using freeview or some other MRI-viewing tool.

It can be helpful to run recon_all -autorecon1 -xopts xopts.txt in a clean directory first to see if this fixes everything, and, if not, then resorting to manual control point setting and/or talairach adjustment. Once everything looks good at the end of -autorecon1, you can then run mne watershed_bem to see if the output is good. Once it is (and once brainmask.mgz is correct), you can then proceed with recon_all -autorecon2 and recon_all -autorecon3 to effectively complete all recon_all steps.

### References¶

1

Denis A. Engemann and Alexandre Gramfort. Automated model selection in covariance estimation and spatial whitening of MEG and EEG signals. NeuroImage, 108:328–342, 2015. doi:10.1016/j.neuroimage.2014.12.040.

2

Samu Taulu, Juha Simola, and Matti J. Kajola. Applications of the signal space separation method. IEEE Transactions on Signal Processing, 53(9):3359–3372, 2005. doi:10.1109/TSP.2005.853302.