Here we provide two new spiking datasets for the evaluation of spiking neural networks. The Spiking Heidelberg Digits (SHD) dataset and the Spiking Speech Command (SSC) dataset are both audio-based classification datasets for which input spikes and output labels are provided. The datasets are released under the Creative Commons Attribution 4.0 International License.
When using these data or the code for your work, please cite:
Cramer, B., Stradmann, Y., Schemmel, J., and Zenke, F. (2022). The Heidelberg Spiking Data Sets for the Systematic Evaluation of Spiking Neural Networks. IEEE Transactions on Neural Networks and Learning Systems 33, 2744–2757. https://doi.org/10.1109/TNNLS.2020.3044364.
https://compneuro.net/datasets/
Mirrors: https://zenkelab.org/datasets/ and https://ieee-dataport.org/open-access/heidelberg-spiking-datasets
We provide two distinct classification datasets for spiking neural networks.
Name | Classes | Samples (train/valid/test) | Parent dataset | URL |
---|---|---|---|---|
SHD | 20 | 8156/-/2264 | Heidelberg Digits (HD) | https://compneuro.net/datasets/hd_audio.tar.gz |
SSC | 35 | 75466/9981/20382 | Speech Commands v0.2 | https://arxiv.org/abs/1804.03209 |
Both datasets are based on respective audio datasets. Spikes in 700 input channels were generated using an artificial cochlea model. The conversion model is freely available from https://github.com/electronicvisions/lauscher . The SHD consists of approximately 10000 high-quality aligned studio recordings of spoken digits from 0 to 9 in both German and English language. Recordings exist of 12 distinct speakers two of which are only present in the test set. The SSC is based on the Speech Commands release by Google which consists of utterances recorded from a larger number of speakers under less controlled conditions. It contains 35 word categories from a larger number of speakers.
For the leaderboard see https://zenkelab.org/resources/spiking-heidelberg-datasets-shd/
For maximum compatibility, the SHD datasets are provided in HDF5 format which can be read by most major programming languages.
root
|-spikes
|-times[]
|-units[]
|-labels[]
|-extra
|-speaker[]
|-keys[]
|-meta_info
|-gender[]
|-age[]
|-body_height[]
Each datum consists of two lists that contain the firing times and the unit id of which neuron has fired at the corresponding firing time.
For a tutorial how to train a spiking neural network on this dataset checkout: https://github.com/fzenke/spytorch/blob/master/notebooks/SpyTorchTutorial4.ipynb
The following code illustrates howto download and access the dataset in Python. The example code uses the PyTables package (https://www.pytables.org) to load HDF5 files.
import os
import urllib.request
import gzip, shutil
from tensorflow.keras.utils import get_file
cache_dir=os.path.expanduser("~/data")
cache_subdir="hdspikes"
print("Using cache dir: %s"%cache_dir)
# The remote directory with the data files
base_url = "https://compneuro.net/datasets"
# Retrieve MD5 hashes from remote
response = urllib.request.urlopen("%s/md5sums.txt"%base_url)
data = response.read()
lines = data.decode('utf-8').split("\n")
file_hashes = { line.split()[1]:line.split()[0] for line in lines if len(line.split())==2 }
def get_and_gunzip(origin, filename, md5hash=None):
gz_file_path = get_file(filename, origin, md5_hash=md5hash, cache_dir=cache_dir, cache_subdir=cache_subdir)
hdf5_file_path=gz_file_path[:-3]
if not os.path.isfile(hdf5_file_path) or os.path.getctime(gz_file_path) > os.path.getctime(hdf5_file_path):
print("Decompressing %s"%gz_file_path)
with gzip.open(gz_file_path, 'r') as f_in, open(hdf5_file_path, 'wb') as f_out:
shutil.copyfileobj(f_in, f_out)
return hdf5_file_path
# Download the Spiking Heidelberg Digits (SHD) dataset
files = [ "shd_train.h5.gz",
"shd_test.h5.gz",
]
for fn in files:
origin = "%s/%s"%(base_url,fn)
hdf5_file_path = get_and_gunzip(origin, fn, md5hash=file_hashes[fn])
print(hdf5_file_path)
# Similarly, to download the SSC dataset
files = [ "ssc_train.h5.gz",
"ssc_valid.h5.gz",
"ssc_test.h5.gz",
]
for fn in files:
origin = "%s/%s"%(base_url,fn)
hdf5_file_path = get_and_gunzip(origin,fn,md5hash=file_hashes[fn])
print(hdf5_file_path)
# At this point we can visualize some of the data
import tables
import numpy as np
fileh = tables.open_file(hdf5_file_path, mode='r')
units = fileh.root.spikes.units
times = fileh.root.spikes.times
labels = fileh.root.labels
# This is how we access spikes and labels
index = 0
print("Times (ms):", times[index])
print("Unit IDs:", units[index])
print("Label:", labels[index])
# A quick raster plot for one of the samples
import matplotlib.pyplot as plt
fig = plt.figure(figsize=(16,4))
idx = np.random.randint(len(times),size=3)
for i,k in enumerate(idx):
ax = plt.subplot(1,3,i+1)
ax.scatter(times[k],700-units[k], color="k", alpha=0.33, s=2)
ax.set_title("Label %i"%labels[k])
ax.axis("off")
plt.show()
The artificial cochlear model used for converting the datasets is freely available at https://github.com/electronicvisions/lauscher
Copyright 2019-2020 Benjamin Cramer & Friedemann Zenke
This work is licensed under a Creative Commons Attribution 4.0 International License.
compneuro.net is here to hosts a (random) selection of useful resources for computational neuroscience. It is currently maintained by Friedemann Zenke https://fzenke.net If you have something to contribute, please get in touch.
Thanks for reading!