Fixed interval windows processing#

This example shows how to process a dataset using the moabb.paradigms.FixedIntervalWindowsProcessing paradigm. This paradigm creates epochs at fixed intervals, ignoring the stim channel and events of the datasets. Therefore, it is compatible with all the datasets. Unfortunately, this paradigm is not compatible with the MOABB evaluation framework. However, it can be used to process datasets for unsupervised algorithms.

In this example, we will use the Zhou2016 dataset because it is relatively small and can be downloaded quickly.

# Authors: Pierre Guetschel <pierre.guetschel@gmail.com>
#
# License: BSD (3-clause)

import matplotlib.pyplot as plt
import mne
import numpy as np

from moabb import set_log_level
from moabb.datasets import Zhou2016
from moabb.paradigms import FixedIntervalWindowsProcessing, MotorImagery


set_log_level("info")

Process a dataset#

To process a dataset with moabb.paradigms.FixedIntervalWindowsProcessing , you can use the method as with every other paradigm. The only additional parameters are length, stride, start_offset, and stop_offset. They are all parametrised in seconds. length is the length of the epochs, stride is the time between the onset of two consecutive epochs, start_offset is the offset between each run start and their first epoch, and stop_offset is the offset between each run start and their last epoch. The default values are length=5, stride=10, start_offset=0, and stop_offset=None (i.e. end of the run).

An example usage of moabb.paradigms.FixedIntervalWindowsProcessing with the moabb.datasets.Zhou2016 dataset:

dataset = Zhou2016()
processing = FixedIntervalWindowsProcessing(
    # new parameters:
    length=100,
    stride=50,
    start_offset=300,
    stop_offset=900,  # we epoch 10 minutes per run, starting at 5 minutes (i.e. 300 seconds)
    # parameters common with other paradigms:
    resample=100,
    fmin=7,
    fmax=45,
    baseline=None,
    channels=None,
)
X, labels, metadata = processing.get_data(dataset=dataset, subjects=[1])

/home/runner/work/moabb/moabb/moabb/datasets/preprocessing.py:279: UserWarning: warnEpochs <Epochs | 10 events (all good), 0 – 100 s (baseline off), ~26.7 MB, data loaded,
 'Window': 10>
  warn(f"warnEpochs {epochs}")
/home/runner/work/moabb/moabb/moabb/datasets/preprocessing.py:279: UserWarning: warnEpochs <Epochs | 10 events (all good), 0 – 100 s (baseline off), ~26.7 MB, data loaded,
 'Window': 10>
  warn(f"warnEpochs {epochs}")
/home/runner/work/moabb/moabb/moabb/datasets/preprocessing.py:279: UserWarning: warnEpochs <Epochs | 10 events (all good), 0 – 100 s (baseline off), ~26.7 MB, data loaded,
 'Window': 10>
  warn(f"warnEpochs {epochs}")
/home/runner/work/moabb/moabb/moabb/datasets/preprocessing.py:279: UserWarning: warnEpochs <Epochs | 10 events (all good), 0 – 100 s (baseline off), ~26.7 MB, data loaded,
 'Window': 10>
  warn(f"warnEpochs {epochs}")
/home/runner/work/moabb/moabb/moabb/datasets/preprocessing.py:279: UserWarning: warnEpochs <Epochs | 10 events (all good), 0 – 100 s (baseline off), ~26.7 MB, data loaded,
 'Window': 10>
  warn(f"warnEpochs {epochs}")
/home/runner/work/moabb/moabb/moabb/datasets/preprocessing.py:279: UserWarning: warnEpochs <Epochs | 10 events (all good), 0 – 100 s (baseline off), ~26.7 MB, data loaded,
 'Window': 10>
  warn(f"warnEpochs {epochs}")

In this dataset, there are three sessions per subject and two runs per session:

for column in metadata.columns:
    print(f"{column}s: {metadata[column].unique()}")

subjects: [1]
sessions: ['0' '1' '2']
runs: ['0' '1']

We expect to obtained (stop_offset - start_offset - length) / stride; i.e. (900-300-100)/50=10 epochs per run. Here we have 3*2=6 runs. And indeed, we obtain a total of 6*10=60 epochs:

print(f"Number of epochs: {len(X)}")

Number of epochs: 60

Note

To apply a bank of bandpass filters, you can use the moabb.paradigms.FilterBankFixedIntervalWindowsProcessing paradigm instead.

Print the events#

We can print the position of the created epochs within the run next to the original events of the dataset. For this, we will first instantiate a moabb.paradigms.MotorImagery paradigm to recover the original events of the dataset:

paradigm = MotorImagery(
    resample=100,
    fmin=7,
    fmax=45,
    baseline=None,
    channels=None,
)

Then, we can recover the events of both paradigms using the _get_events_pipeline method:

events_pipeline_dataset = paradigm._get_events_pipeline(dataset)
events_pipeline_fixed = processing._get_events_pipeline(dataset)
raw = dataset.get_data(subjects=[1])[1]["0"]["0"]
events_dataset = events_pipeline_dataset.transform(raw)
events_fixed = events_pipeline_fixed.transform(raw)
events = np.concatenate([events_dataset, events_fixed])
event_id = dict(**paradigm.used_events(dataset), **processing.used_events(dataset))

Finally, we can plot the events. The artificial events created by moabb.paradigms.FixedIntervalWindowsProcessing are named "Windows":

fig = mne.viz.plot_events(
    events,
    sfreq=raw.info["sfreq"],
    event_id=event_id,
)
fig.subplots_adjust(right=0.7)
plt.show()

/home/runner/work/moabb/moabb/examples/plot_fixed_interval_windows.py:122: UserWarning: This figure was using a layout engine that is incompatible with subplots_adjust and/or tight_layout; not calling subplots_adjust.
  fig.subplots_adjust(right=0.7)

We can see that the epochs were effectively created at a fixed interval every 50 seconds between 300 and 900 seconds, and ignoring the original events of the dataset.

Total running time of the script: ( 0 minutes 12.123 seconds)

Estimated memory usage: 260 MB

Gallery generated by Sphinx-Gallery