Exploratory data analysis with NWB

This example will focus on the basics of working with an NWBFile to do more than storing standardized data for use and exchange. For example, you may want to store results from intermediate analyses or one-off analyses with unknown utility. This functionality is primarily accomplished with linking and scratch space.


The scratch space is explicitly for non-standardized data that is not intended for reuse by others. Standard NWB:N types, and extension if required, should always be used for any data that you intend to share. As such, published data should not include scratch data and a user should be able to ignore any data stored in scratch to use a file.

Raw data

To demonstrate linking and scratch space, lets assume we are starting with some acquired data.

from pynwb import NWBFile, TimeSeries, NWBHDF5IO
from datetime import datetime
from dateutil.tz import tzlocal
import numpy as np

# set up the NWBFile
start_time = datetime(2019, 4, 3, 11, tzinfo=tzlocal())
create_date = datetime(2019, 4, 15, 12, tzinfo=tzlocal())

nwb = NWBFile(session_description='demonstrate NWBFile scratch',  # required
              identifier='NWB456',  # required
              session_start_time=start_time,  # required
              file_create_date=create_date)  # optional

# make some fake data
timestamps = np.linspace(0, 100, 1024)
data = np.sin(0.333 * timestamps) + np.cos(0.1 * timestamps) + np.random.randn(len(timestamps))
test_ts = TimeSeries(name='raw_timeseries', data=data, unit='m', timestamps=timestamps)

# add it to the NWBFile

with NWBHDF5IO('raw_data.nwb', mode='w') as io:

Copying an NWB file

To copy a file, we must first read the file.

raw_io = NWBHDF5IO('raw_data.nwb', 'r')
nwb_in = raw_io.read()

And then create a shallow copy the file with the copy method of NWBFile .

nwb_proc = nwb_in.copy()

Now that we have a copy, lets process some data, and add the results as a ProcessingModule to our copy of the file. [1]

import scipy.signal as sps

mod = nwb_proc.create_processing_module('filtering_module', "a module to store filtering results")

ts1 = nwb_in.acquisition['raw_timeseries']
filt_data = sps.correlate(ts1.data, np.ones(128), mode='same') / 128
ts2 = TimeSeries(name='filtered_timeseries', data=filt_data, unit='m', timestamps=ts1)


Now write the copy, which contains the processed data. [2]

with NWBHDF5IO('processed_data.nwb', mode='w', manager=raw_io.manager) as io:


Notice here that we are reusing the timestamps to the original TimeSeries.



The processed_data.nwb file (i.e., our copy) stores our processing module and contains external links to all data in our original file, i.e., the data from our raw file is being linked to, not copied. This allows us to isolate our processing data in a separate file while still allowing us to access the raw data from our processed_data.nwb file, without having to duplicate the data.

Adding scratch data

You may end up wanting to store results from some one-off analysis, and writing an extension to get your data into an NWBFile is too much over head. This is facilitated by the scratch space in NWB:N. [3]

First, lets read our processed data and then make a copy

proc_io = NWBHDF5IO('processed_data.nwb', 'r')
nwb_proc_in = proc_io.read()

Now make a copy to put our scratch data into [4]

nwb_scratch = nwb_proc_in.copy()

Now lets do an analysis for which we do not have a specification, but we would like to store the results for.

filt_ts = nwb_scratch.modules['filtering_module']['filtered_timeseries']

fft = np.fft.fft(filt_ts.data)

nwb_scratch.add_scratch(fft, name='dft_filtered', description='discrete Fourier transform from filtered data')

Finally, write the results.

with NWBHDF5IO('scratch_analysis.nwb', 'w', manager=proc_io.manager) as io:

To get your results back, you can index into scratch or use get_scratch:

scratch_io = NWBHDF5IO('scratch_analysis.nwb', 'r')
nwb_scratch_in = scratch_io.read()

fft_in = nwb_scratch_in.scratch['dft_filtered']

fft_in = nwb_scratch_in.get_scratch('dft_filtered')


This scratch space only exists if you add scratch data.



We recommend writing scratch data into copies of files only. This will make it easier to isolate and discard scratch data and avoids updating files that store precious data.

Gallery generated by Sphinx-Gallery