Deep Learning Visualizer

An interactive neural network. Paint a digit in a 28×28 grid with the mouse, press space, and the model tells you which digit it thinks you drew. Press t to train it on MNIST and watch the cost fall, then press y to measure its accuracy.

The point is to make backpropagation and weight initialization something you can poke at rather than read about. Every prediction is computed live from the pixels under your cursor.

Controls

Input	Action
Left mouse button, drag	Paint the digit
space	Predict the drawn digit, printed to the terminal
r	Clear the grid
t	Train on MNIST
y	Report accuracy on the MNIST test set
s	Save the current weights to data/mnist/trained.mlp

Predictions and training progress are printed to the terminal that launched the program, not drawn in the window.

Requirements

The project targets a desktop OpenGL 3.3 core profile, so it needs a GPU and driver that can provide one. It builds on macOS and Linux, and CI runs it on Ubuntu.

Building requires a C++23 compiler, CMake, and a few system libraries.

macOS

Needs Xcode or the Command Line Tools for a C++23 compiler, plus CMake.

xcode-select --install
brew install cmake

macOS is detected automatically and the required system frameworks are linked for you. No OpenGL library needs to be installed, since the system provides it.

Linux (Ubuntu and Debian)

sudo apt-get install -y g++-14 cmake libwayland-dev libxkbcommon-dev xorg-dev

libwayland-dev, libxkbcommon-dev, and xorg-dev are what GLFW needs to create a window on Linux.

Building

Clone with the submodules, since GLFW and FreeType are pulled in as git submodules and both need to be compiled before the project can link against them.

git clone --recursive https://github.com/nycrat/deep-learning-visualizer
cd deep-learning-visualizer

Then build the dependencies. This installs GLFW and FreeType into external/, which is where the build looks for them.

./scripts/setup.sh

The MNIST files are already committed under data/mnist/, so there is nothing else to download.

Finally, build and run:

make
make run

make produces a debug build in bin/debug/. Set BUILD_TYPE=release for an optimized build in bin/release/:

make BUILD_TYPE=release

Release builds use -O3 -march=native, so they are fast but not portable between machines.

Running the tests

make test
make test BUILD_TYPE=release

See docs/testing.md for how the suite is laid out and how to write new tests.

Generating the documentation

make docs

Doxygen writes the site to docs/html, starting at docs/html/index.html. It is also published to GitHub Pages on every push; see .github/workflows/docs.yml.

Generating compile_commands.json

To get LSP working properly, use bear:

make clean
bear -- make