neuralnet

Version 0.3

An Artificial Neural Network framework for Pure Data

[neuralnet] is an artificial neural network Pd external, written in pure C, without any dependencies. It is inspired by the book "Neural Networks from Scratch in Python" by Harrison Kinsley & Daniel Kukieła. It is an attempt to translate the Python code to C with the Pure Data API, to run neural networks within Pd.

[neuralnet] creates densely connected neural networks for classification, regression, and binary logistic regression. There are different activation functions and optimizers you can set, and various other settable parameters. The object's help patch and the examples found in the examples directory should cover all the necessary information.

Note about Make

This repository uses the pd-lib-builder Makefile system. You can get it from here. The directory of the Makefile should be in the same directory of the neuralnet directory. For example, run:

cd ../
git clone https://github.com/pure-data/pd-lib-builder.git
cd -

If you compile the object from sources, you should also change the following line in the Makefile:

PDLIBBUILDER_DIR=./pd-lib-builder/

to (one more dot)

PDLIBBUILDER_DIR=../pd-lib-builder/

If you fork this repository and build in GitHub, it should work as is.

Note about the examples

Example 03-mouse_input.pd uses [mousestate] from the Cyclone library, to get the coordinates of the mouse. Example 04-mnist_digits.pd uses the [command] external, plus some Python scripts (called via [command]). Example 05-audio_autoencoder.pd uses [command], though it is not necessary to use it, see patch for details.

All external objects used in the examples can be installed via the deken plugin (Help->Find externals).

If you mention this object in an academic paper/chapter/article, please include it in your bibliography with the following citation:

@article{Drymonitis2023neuralnet,
	author = {Drymonitis, Alexandros},
	journal = {AIMC 2023},
	year = {2023},
	month = {aug 29},
	note = {https://aimc2023.pubpub.org/pub/3j3fx7y1},
	publisher = {},
	title = {[neuralnet]: A {Pure} {Data} {External} for the {Creation} of {Neural} {Networks} {Written} in {Pure} {C}},
}

Log:

• Added four new activation functions
• Added the ability to access the inner structure of a network (e.g. the latent space)
• Added the ability to store weights and biases during training
• Added the ability to automatically save models during training
• Created an audio autoencoder example
• Added a signal-rate version of the object

Special thanks to Ben Wesch for providing macOS and Windows binaries, IOhannes m zmoelnig and Cristof Ressi, for helping out through Pd's mailing list. Dan Wilcox for spotting compiler warnings and typos in the help files.

Written by Alexandros Drymonitis