undr 1.0.1

Download and process Neuromorphic datasets

Unified Neuromorphic Datasets Repository

• Unified Neuromorphic Datasets Repository
  • Getting Started
    • Install the undr module
    • Generate a default configuration file
    • Download the datasets
    • Generate a BibTex file
  • Python module
    • Concepts
    • Python APIs
  • Dataset format specification
  • Dataset mirrors
    • Example configuration
      • Apache
      • Nginx
      • Upload a dataset
    • Download with existing CLI
  • Build documentation
• compilation

Getting Started

Install the undr module

pip3 install undr

Documentation: https://neuromorphicsystems.github.io/undr/

Generate a default configuration file

python3 -m undr init

The generated undr.toml file is written in TOML (https://github.com/toml-lang/toml). It lists the datasets that will be downloaded or streamed, hence it needs to be ajusted to your needs.

The line directory = 'datasets' specifies the directory where downloaded files are stored (relatively to the configuration file). All the files generated by undr (directory indexes, downloaded data, temporary files...) are stored in this directory.

Datasets are listed as [[datasets]] entries with three mandatory properties: name, url and mode. The optional server_type property is used internally to speed up the download process. To discard a dataset, you can either remove it from the configuration file or comment all its lines with # signs.

mode changes the download strategy on a per-dataset basis, with three possible values:

• 'remote' only downloads the dataset's file index. The undr Python package can be used to process the dataset files as if they were on your hard drive by streaming them from the server. This option is particularly useful for large datasets that do not fit on your disk but requires a fast internet connection since files are re-downloaded every time they are processed.
• 'local' downloads all the dataset files locally but does not decompress them (most datasets are stored as Brotli archives). The undr Python library transparently decompresses files in memory when you read them, making this option a good trade-off between disk usage and processing speed.
• 'local-decompressed' downloads all the dataset files locally and decompresses them. Decompressed files use a relatively inefficient plain binary file format so this option requires vast amounts of disk space (3 to 5 times as much as the Brotli archives). On the other hand, the plain binary format facilitates processing with other languages such as Matlab or C++.

undr also supports hybrid configurations where only part of a dataset is downloaded or decompressed. You may also use local directories without a server. See [NOT DOCUMENTED YET] for details.

Download the datasets

python3 -m undr install

This command downloads the datasets file indexes. If the mode is 'compressed' or 'decompress', it also downloads the dataset files (and possibly decompresses them).

This command can be interrupted at any time with CTRL + C. Re-running it will resume download where it left off.

Generate a BibTex file

python3 -m undr bibtex --output datasets.bib

The UNDR project does not claim authorship of the datasets. Please use this file to cite the origiinal articles.

Python module

pip3 install undr

Concepts

• priority
• -index.json
• files
• other_files

Python APIs

API name	Complexity	Configurability	Parallel processing	Progress display
loop	simple	high	no	no
map	simple	low	yes	yes
task	complex	high	yes	yes

All three approches support progress persistence. Progress persistence sightly increases the code complexity but makes it possible to resume processing after a network or power failure.

Dataset format specification

-index.json rationale

• prepend a special character to make sure the index file is ordered first in ASCII
• use an unreserved URL character to avoid escaping problems
• out of the unreserved URL characters (- / . / _ / ~, see https://www.rfc-editor.org/rfc/rfc3986#section-2):
  • . would result in a hidden file on UNIX systems
  • _ comes after alpha-numerical characters in ASCII
  • ~ is a shortcut for the user's home directory in many shells

Many command-line programs treat -index.json as a flag, hence a command such as cat -index.json returns an error. Prepending ./ to the filename avoids the problem: cat ./-index.json.

Dataset mirrors

Example configuration

Apache

<VirtualHost *:80>
    Alias / /path/to/local/directory/
    <Directory "/path/to/local/directory/">
        Require all granted
        Options +Indexes
    </Directory>
</VirtualHost>

To use another port, remember to edit /etc/apache2/ports.conf as well.

Nginx

server {
    listen 80;
    location / {
        alias /path/to/local/directory/;
        autoindex on;
        sendfile on;
        tcp_nopush on;
        sendfile_max_chunk 1m;
    }
}

Upload a dataset

• python3 -m undr check-conformity /path/to/dataset

• Caveat: An UNDR server can provide multiple compressed files (different formats) for each resource. The Python UNDR library always picks the best compression (smallest encoded size). check-conformity only checks the best compression and will not report errors for other compressions.

• for macOS users (.DS_Store)

Add to ~/.zshrc:

# rmdsstore removes .DS_Store files recursively
rmdsstore() {
    if [ $# -eq 0 ]; then
        printf 'usage: rmdsstore directory\n' >&2
        return
    fi
    find "$1" -name ".DS_Store" -delete -print
}

Run rmdsstore /path/to/dataset before running python3 -m undr check-conformity /path/to/dataset.

Download with existing CLI

wget --no-parent --recursive --level=inf http://localhost:5432/dvs09/
find . -iname '\*.br' | while read filename; do brotli -d -j "$filename"; done;

Build documentation

# setup
python3 -m venv .venv
source .venv/bin/activate
python documentation/build.py

pip install --upgrade pip setuptools pip install -r documentation/requirements.txt

compilation

cd python pip install . cd .. sphinx-build documentation documentation/_build