Deep learning library for C/C++

Bob Mottram 83260a0b44 Smaller number of inputs in deeplearn results in more confident convergence within a fixed number of training cycles 1 year ago
examples 496a449d21 Update to newer version of BSD license 1 year ago
img 0404d099a7 screenshot 3 years ago
man f7b3081293 Change domains from github 2 years ago
src 6c29d6ca12 Update year 1 year ago
unittests 83260a0b44 Smaller number of inputs in deeplearn results in more confident convergence within a fixed number of training cycles 1 year ago
.gitignore 99065c8805 Add images 2 years ago
LICENSE 496a449d21 Update to newer version of BSD license 1 year ago
Makefile 01ed28c007 Spectre mitigation causes compile problems on debian 1 year ago f7b3081293 Change domains from github 2 years ago a0655ed486 Documentation reproducibility 3 years ago

libdeep: C/C++ library for deep learning

This is a C library which can be used in deep learning applications. What differentiates libdeep from the numerous other deep learning systems out there is that it's small and that trained networks can be exported as completely standalone C or Python programs which can either be used via the commandline or within an Arduino IDE for creating robotics or IoT applications.

A Python API for libdeep can be found at


On Debian based systems:

sudo apt-get install build-essential doxygen xdot

If you want to be able to visualize call graphs for development or debugging purposes then you will need to install the egypt script.

On Arch based systems:

sudo pacman -S gcc doxygen egypt xdot

To build from source:

sudo make install

This creates the library and installs it into /usr/local

Unit Tests

You can run the unit tests to check that the system is working as expected:

cd unittests

Or to check for any memory leaks:

valgrind --leak-check=full ./tests

Source Documentation

To generate source code documentation make sure that you have doxygen installed and then run the script. A subdirectory called docs will be created within which html and latex formated documentation can be found. For general usage information you can also see the manpage.

man libdeep

Text or Numeric

Inputs within training data sets can include both text and/or numeric data. For example, within the XOR training data you can have something like:


or alternatively, using whatever text representations you prefer:


Showing the call graph

If you want to visualize the call graph for a particular source file for debugging or development purposes:

SOURCEFILE=deeplearn.c make graph

And you can change the SOURCEFILE value to whatever file you're interested in.


There are also some example programs within the examples directory. Reading the examples is the best way to learn how to use this library within your own code. Examples are:

  • Face recognition with a deep convolutional network
  • Automatic feature learning and image reconstruction from features
  • Determining whether a cancer is malignant or benign
  • Assessing wine quality from ingredients
  • Predicting concrete quality from ingredients

Using trained neural nets in your system

You can export trained neural nets either as a C program or a Python program. These programs are completely independent and can be used either as commands or integrated into a larger software application. This makes it easy to use the resulting neural net without needing to link to libdeep. See the source code in the examples directory for how to use the export function. If you include the word "sketch" or "arduino" within the filename to be exported to then it will appear as Arduino compatible C suitable for use within an Arduino IDE rather than as standard C.

In your training program:

deeplearn_export(&learner, "export.c");

Then when training is done:

gcc -o export_c export.c -lm
./export_c [first input] [second input]...


Although this software was primarily written to run on Linux-based systems it's pretty much just vanilla C99 standard code and so it should be easily portable to other platforms, such as Microsoft Windows and Mac systems. The independent random number generator should mean that results are consistent across different compilers and platforms.


To build packages for Debian (deb) see