Waveform Extractor

SpikeInterface provides an efficient mechanism to extract waveform snippets.

The WaveformExtractor class:

• randomly samples a subset spikes with max_spikes_per_unit

• extracts all waveforms snippets for each unit

• saves waveforms in a local folder

• can load stored waveforms

• retrieves template (average or median waveform) for each unit

Here the how!

import matplotlib.pyplot as plt

from spikeinterface import download_dataset
from spikeinterface import WaveformExtractor, extract_waveforms
import spikeinterface.extractors as se

First let’s use the repo https://gin.g-node.org/NeuralEnsemble/ephy_testing_data to download a MEArec dataset. It is a simulated dataset that contains “ground truth” sorting information:

repo = 'https://gin.g-node.org/NeuralEnsemble/ephy_testing_data'
remote_path = 'mearec/mearec_test_10s.h5'
local_path = download_dataset(repo=repo, remote_path=remote_path, local_folder=None)

[INFO] Cloning dataset to Dataset(/home/docs/spikeinterface_datasets/ephy_testing_data)
[INFO] Attempting to clone from https://gin.g-node.org/NeuralEnsemble/ephy_testing_data to /home/docs/spikeinterface_datasets/ephy_testing_data
[INFO] Start enumerating objects
[INFO] Start counting objects
[INFO] Start compressing objects
[INFO] Start receiving objects
[INFO] Start resolving deltas
[INFO] Completed clone attempts for Dataset(/home/docs/spikeinterface_datasets/ephy_testing_data)
install(ok): /home/docs/spikeinterface_datasets/ephy_testing_data (dataset)
get(ok): mearec/mearec_test_10s.h5 (file) [from origin...]

Let’s now instantiate the recording and sorting objects:

recording = se.MEArecRecordingExtractor(local_path)
print(recording)
sorting = se.MEArecSortingExtractor(local_path)
print(recording)

MEArecRecordingExtractor: 32 channels - 1 segments - 32.0kHz - 10.000s
  file_path: /home/docs/spikeinterface_datasets/ephy_testing_data/mearec/mearec_test_10s.h5
MEArecRecordingExtractor: 32 channels - 1 segments - 32.0kHz - 10.000s
  file_path: /home/docs/spikeinterface_datasets/ephy_testing_data/mearec/mearec_test_10s.h5

The MEArec dataset already contains a probe object that you can retrieve an plot:

probe = recording.get_probe()
print(probe)
from probeinterface.plotting import plot_probe

plot_probe(probe)

Probe - 32ch - 1shanks

(<matplotlib.collections.PolyCollection object at 0x7f6ae6a00e20>, <matplotlib.collections.PolyCollection object at 0x7f6ae69d25e0>)

A WaveformExtractor object can be created with the extract_waveforms() function:

folder = 'waveform_folder'
we = extract_waveforms(
    recording,
    sorting,
    folder,
    ms_before=1.5,
    ms_after=2.,
    max_spikes_per_unit=500,
    overwrite=True
)
print(we)

extract waveforms memmap:   0%|          | 0/10 [00:00<?, ?it/s]
extract waveforms memmap:  90%|######### | 9/10 [00:00<00:00, 86.89it/s]
extract waveforms memmap: 100%|##########| 10/10 [00:00<00:00, 87.33it/s]
WaveformExtractor: 32 channels - 10 units - 1 segments
  before:48 after:64 n_per_units:500

Alternatively, the WaveformExtractor object can be instantiated directly. In this case, we need to set_params() to set the desired parameters:

folder = 'waveform_folder2'
we = WaveformExtractor.create(recording, sorting, folder, remove_if_exists=True)
we.set_params(ms_before=3., ms_after=4., max_spikes_per_unit=1000)
we.run_extract_waveforms(n_jobs=1, chunk_size=30000, progress_bar=True)
print(we)

extract waveforms memmap:   0%|          | 0/11 [00:00<?, ?it/s]
extract waveforms memmap:  91%|######### | 10/11 [00:00<00:00, 98.66it/s]
extract waveforms memmap: 100%|##########| 11/11 [00:00<00:00, 99.83it/s]
WaveformExtractor: 32 channels - 10 units - 1 segments
  before:96 after:128 n_per_units:1000

To speed up computation, waveforms can also be extracted using parallel processing (recommended!). We can define some 'job_kwargs' to pass to the function as extra arguments:

job_kwargs = dict(n_jobs=2, chunk_duration="1s", progress_bar=True)

folder = 'waveform_folder_parallel'
we = extract_waveforms(
    recording,
    sorting,
    folder,
    ms_before=3.,
    ms_after=4.,
    max_spikes_per_unit=500,
    overwrite=True,
    **job_kwargs
)
print(we)

extract waveforms memmap:   0%|          | 0/10 [00:00<?, ?it/s]
extract waveforms memmap:  10%|#         | 1/10 [00:00<00:03,  2.40it/s]
extract waveforms memmap:  50%|#####     | 5/10 [00:00<00:00, 11.93it/s]
extract waveforms memmap: 100%|##########| 10/10 [00:00<00:00, 16.87it/s]
WaveformExtractor: 32 channels - 10 units - 1 segments
  before:96 after:128 n_per_units:500

The 'waveform_folder' folder contains:

• the dumped recording (json)

• the dumped sorting (json)

• the parameters (json)

• a subfolder with “waveforms_XXX.npy” and “sampled_index_XXX.npy”

import os

print(os.listdir(folder))
print(os.listdir(folder + '/waveforms'))

['waveforms', 'recording.json', 'sorting.json', 'templates_average.npy', 'params.json', 'recording_info']
['sampled_index_#7.npy', 'waveforms_#0.npy', 'waveforms_#3.npy', 'waveforms_#9.npy', 'sampled_index_#2.npy', 'waveforms_#8.npy', 'sampled_index_#5.npy', 'sampled_index_#1.npy', 'waveforms_#7.npy', 'sampled_index_#9.npy', 'sampled_index_#4.npy', 'waveforms_#4.npy', 'waveforms_#1.npy', 'sampled_index_#0.npy', 'sampled_index_#8.npy', 'sampled_index_#3.npy', 'sampled_index_#6.npy', 'waveforms_#5.npy', 'waveforms_#6.npy', 'waveforms_#2.npy']

Now we can retrieve waveforms per unit on-the-fly. The waveforms shape is (num_spikes, num_sample, num_channel):

unit_ids = sorting.unit_ids

for unit_id in unit_ids:
    wfs = we.get_waveforms(unit_id)
    print(unit_id, ':', wfs.shape)

#0 : (53, 224, 32)
#1 : (50, 224, 32)
#2 : (43, 224, 32)
#3 : (30, 224, 32)
#4 : (48, 224, 32)
#5 : (37, 224, 32)
#6 : (51, 224, 32)
#7 : (110, 224, 32)
#8 : (194, 224, 32)
#9 : (129, 224, 32)

We can also get the template for each units either using the median or the average:

for unit_id in unit_ids[:3]:
    fig, ax = plt.subplots()
    template = we.get_template(unit_id=unit_id, mode='median')
    print(template.shape)
    ax.plot(template)
    ax.set_title(f'{unit_id}')

• 
• 
• 

(224, 32)
(224, 32)
(224, 32)

Or retrieve templates for all units at once:

all_templates = we.get_all_templates()
print(all_templates.shape)


'''
Sparse Waveform Extractor
-------------------------

'''

(10, 224, 32)

'\nSparse Waveform Extractor\n-------------------------\n\n'

For high-density probes, such as Neuropixels, we may want to work with sparse waveforms, i.e., waveforms computed on a subset of channels. To do so, we two options.

Option 1) Save a dense waveform extractor to sparse:

In this case, from an existing waveform extractor, we can first estimate a sparsity (which channels each unit is defined on) and then save to a new folder in sparse mode:

from spikeinterface import compute_sparsity

# define sparsity within a radius of 40um
sparsity = compute_sparsity(we, method="radius", radius_um=40)
print(sparsity)

# save sparse waveforms
folder = 'waveform_folder_sparse'
we_sparse = we.save(folder=folder, sparsity=sparsity, overwrite=True)

# we_sparse is a sparse WaveformExtractor
print(we_sparse)

wf_full = we.get_waveforms(we.sorting.unit_ids[0])
print(f"Dense waveforms shape for unit {we.sorting.unit_ids[0]}: {wf_full.shape}")
wf_sparse = we_sparse.get_waveforms(we.sorting.unit_ids[0])
print(f"Sparse waveforms shape for unit {we.sorting.unit_ids[0]}: {wf_sparse.shape}")

ChannelSparsity - units: 10 - channels: 32 - ratio: 0.18
WaveformExtractor: 32 channels - 10 units - 1 segments
  before:96 after:128 n_per_units:500 - sparse
Dense waveforms shape for unit #0: (53, 224, 32)
Sparse waveforms shape for unit #0: (53, 224, 6)

Option 2) Directly extract sparse waveforms:

We can also directly extract sparse waveforms. To do so, dense waveforms are extracted first using a small number of spikes ('num_spikes_for_sparsity')

folder = 'waveform_folder_sparse_direct'
we_sparse_direct = extract_waveforms(
    recording,
    sorting,
    folder,
    ms_before=3.,
    ms_after=4.,
    max_spikes_per_unit=500,
    overwrite=True,
    sparse=True,
    num_spikes_for_sparsity=100,
    method="radius",
    radius_um=40,
    **job_kwargs
)
print(we_sparse_direct)

template_full = we.get_template(we.sorting.unit_ids[0])
print(f"Dense template shape for unit {we.sorting.unit_ids[0]}: {template_full.shape}")
template_sparse = we_sparse_direct.get_template(we.sorting.unit_ids[0])
print(f"Sparse template shape for unit {we.sorting.unit_ids[0]}: {template_sparse.shape}")

extract waveforms shared_memory:   0%|          | 0/10 [00:00<?, ?it/s]
extract waveforms shared_memory:  10%|#         | 1/10 [00:00<00:03,  2.52it/s]
extract waveforms shared_memory: 100%|##########| 10/10 [00:00<00:00, 20.73it/s]

extract waveforms memmap:   0%|          | 0/10 [00:00<?, ?it/s]
extract waveforms memmap:  10%|#         | 1/10 [00:00<00:03,  2.42it/s]
extract waveforms memmap:  70%|#######   | 7/10 [00:00<00:00, 16.18it/s]
extract waveforms memmap: 100%|##########| 10/10 [00:00<00:00, 17.56it/s]
WaveformExtractor: 32 channels - 10 units - 1 segments
  before:96 after:128 n_per_units:500 - sparse
Dense template shape for unit #0: (224, 32)
Sparse template shape for unit #0: (224, 6)

As shown above, when retrieving waveforms/template for a unit from a sparse 'WaveformExtractor', the waveforms are returned on a subset of channels. To retrieve which channels each unit is associated with, we can use the sparsity object:

# retrive channel ids for first unit:
unit_ids = we_sparse.unit_ids
channel_ids_0 = we_sparse.sparsity.unit_id_to_channel_ids[unit_ids[0]]
print(f"Channel ids associated to {unit_ids[0]}: {channel_ids_0}")

Channel ids associated to #0: ['5' '6' '7' '16' '17' '28']

However, when retrieving all templates, a dense shape is returned. This is because different channels might have a different number of sparse channels! In this case, values on channels not belonging to a unit are filled with 0s.

all_sparse_templates = we_sparse.get_all_templates()

# this is a boolean mask with sparse channels for the 1st unit
mask0 = we_sparse.sparsity.mask[0]
# Let's plot values for the first 5 samples inside and outside sparsity mask
print("Values inside sparsity:\n", all_sparse_templates[0, :5, mask0])
print("Values outside sparsity:\n", all_sparse_templates[0, :5, ~mask0])

plt.show()

Values inside sparsity:
 [[ 6.7338767   3.1060634   2.8129902   1.3475728   2.3475006 ]
 [ 1.953573    4.932686    4.9273763   6.148149    4.4570413 ]
 [ 1.1766979   0.23954824 -0.29712072 -0.0430811  -1.0933229 ]
 [ 2.9538999   3.3712146   3.991932    5.281219    2.9890084 ]
 [ 1.5635666   0.10017583  0.05006344  0.44048104  2.034473  ]
 [ 2.9059367   1.1767691  -0.92374414  0.08211926  0.9413183 ]]
Values outside sparsity:
 [[0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0.]]