Examples Gallery

Examples Gallery

Every example is a complete executable built from the repository's examples tree. Use this page to choose an experiment, open its source, and know what result to inspect.

Build and locate examples

Configure once, then build everything or one target:

cmake -S . -B build -DNUNN_ENABLE_OPENCL=OFF
cmake --build build --config Release
cmake --build build --config Release --target rnn_sine

Paths depend on the generator:

single-config:  build/examples/<name>/<name>
Visual Studio:  build\examples\<name>\Release\<name>.exe

The source-tree build defines every example below. The current install target contains a smaller runtime subset, so use the build tree when a newer demo is not present in an installed bin directory.

Start with these

Order	Example	Why it comes next
1	and_test	smallest trainable unit and linearly separable data
2	xor_test	hidden layer and backpropagation
3	linear_regression_demo	closed-form fit versus gradient descent
4	kmeans_demo	unsupervised assignment and centroids
5	rnn_sine	state, BPTT, and autoregressive evaluation
6	cnn_seq	end-to-end local filters and pooling
7	maze	state/action/reward loop
8	dqn_maze	replay buffer and target network
9	mnist_test	full dataset, test metrics, persistence, backend

Feedforward and regression

Target	Model	Run	Look for
and_test	Perceptron	and_test	all four truth-table rows classified correctly
xor_test	MlpNN	xor_test	nonlinear separation after training
counter_test	MlpNN	counter_test	encoded counter-state mapping
linear_regression_demo	LinearRegression	linear_regression_demo	OLS and gradient-descent coefficients, MSE, R²
titanic	supervised classifier	titanic	feature encoding on an embedded real-world table

AND and XOR form the most useful pair: the data changes only from linearly separable to nonlinearly separable, so the need for a hidden layer is isolated.

MNIST, OCR, and model tools

Target	Purpose	Run
mnist_test	train/test MlpNN or MlpMatrixNN on IDX data	mnist_test -p /path/to/mnist
ocr_test	Windows drawing UI, model loading, and MNIST training	launch ocr_test.exe
nunn_topo	render topology from JSON, legacy .net, or explicit sizes	nunn_topo --topology 2,3,1
net2json	convert legacy network streams to JSON	net2json <legacy.net> <model.json>

Useful topology commands:

nunn_topo --topology 2,3,1 --save xor.dot
nunn_topo --load model.json --save model.svg
nunn_topo --load model.net --save model.png

DOT needs no external renderer. SVG, PNG, and PDF require Graphviz. Large networks are compacted by default; use --full only when every node and edge is genuinely useful.

See MNIST and OCR for dataset layout and the full option matrix.

Recurrent models

Target	Models	Defaults and useful variants
rnn_sine	Vanilla RNN, GRU, LSTM	rnn_sine, rnn_sine --gru, rnn_sine --lstm
rnn_adding	all three side by side	rnn_adding [sequence_length] [hidden] [epochs] [lr]
rnn_char	Vanilla RNN, GRU, LSTM	rnn_char --gru 1200 128 120 0.8

rnn_sine reports an autoregressive rollout, not only a one-step fit. rnn_adding is the better long-memory comparison because exactly two marked values must survive irrelevant steps. rnn_char shows how temperature changes sampling from softmax output.

Read Recurrent Networks before interpreting architecture differences.

Convolution and attention

Target	Model	Run	Look for
cnn_seq	ConvNet	cnn_seq [epochs] [lr]	train/test accuracy on noisy one- versus two-cycle signals
transformer_char	MiniTransformer	transformer_char [epochs] [lr] [generated_length]	mean token loss and causal continuation

These examples keep inputs deliberately small so the full forward and backward implementations in nu_conv.cc and nu_transformer.cc remain practical to inspect.

Classical, representation, and generative models

Target	Model	Run	Primary metric or artifact
kmeans_demo	KMeans	kmeans_demo [k] [samples_per_cluster] [seed]	inertia, centroids, ASCII assignment map
pca_demo	Pca	pca_demo [components] [samples] [seed]	explained variance and reconstruction MSE
hopfield_test	HopfieldNN	hopfield_test	recall from a corrupted binary pattern
ae_demo	Autoencoder	ae_demo [epochs]	bottleneck codes and reconstruction
rbf_demo	Rbf	rbf_demo [centers] [epochs] [lr]	sine-regression train/test MSE
rbm_demo	Rbm	rbm_demo	reconstruction before/after CD-1
vae_demo	Vae	vae_demo	reconstructions, latent means, generated samples
som_demo	Som	som_demo	quantization error and organized prototype grid

Do not rank these models by one shared loss: each optimizes a different objective. Classical and Unsupervised Models explains the metrics and provides minimal API fragments.

Reinforcement learning and games

Target	Model	Run	Observe
maze	QLearn or Sarsa selected in source	maze	learned navigation policy
path_finder	graph Q-learning	path_finder	route recovered from state values
dqn_maze	Dqn	dqn_maze [episodes] [lr]	rolling successes, learn steps, greedy trace
tictactoe	MlpNN	tictactoe	console game-state evaluation
winttt	MlpNN	launch winttt.exe on Windows	GUI inference and packaged resources

For RL, record the reward definition and episode cap before comparing results. See Reinforcement Learning.

How to study one example

For any target:

1. Run it unchanged and save the output.
2. Read its main() from construction through evaluation.
3. Open the linked public header and identify every method called.
4. Follow one forward/update path in the implementation.
5. Open the matching test from tests.
6. Change one parameter, predict the effect, then rerun.

This turns the examples into controlled experiments rather than isolated demos.

Keep reading

Use Implementation Map for model-to-source-to-test cross-references and Training and Diagnostics for a repeatable experiment record.