A small experimental project that implements a simple feedforward neural network from scratch in C++ and visualizes the training process for a regression task.
The goal of the project is educational: to demonstrate how a neural network learns to approximate a function (for example, a sine wave) and to visually observe the training dynamics.
The visualization is implemented using Dear ImGui, and the neural network itself is written without external ML libraries.
- Implementation of a fully connected neural network from scratch
- Support for multiple layers
- tanh activation
- Mean Squared Error (MSE) loss
- Gradient descent training
- Tracking and storing training history
- Interactive training visualization
- Zoom and pan of the coordinate system
- Playback of training iterations
Input
↓
Dense
↓
Tanh
↓
Dense
↓
Output
Dense(1 → 32)
Tanh
Dense(32 → 16)
Tanh
Dense(16 → 1)
The network approximates a continuous function from a generated dataset.
The dataset is generated using a Python script.
Example function:
y = sin(x) + noise
To generate a dataset:
python scripts/generate_dataset.py
This creates:
dataset.csv
Each file contains:
x,y
The project uses Dear ImGui to visualize:
- training samples
- test samples
- current model prediction
- training progress over iterations
Controls:
Control Description
Play / Pause Start or stop training playback Reset Reset and drag to default Epochs slider Move through training steps Speed Control playback speed Mouse drag Move canvas Mouse wheel Zoom
During training the following data is recorded:
- Loss per epoch
- Predict values per epoch
This allows replaying the entire training process visually. Stopping criterion used in the trainer: training stops when loss stops decreasing
The project uses the following libraries:
- C++20
- OpenGL
- GLFW
- GLAD
- Dear ImGui
- OpenMP
Example using CMake:
mkdir build
cmake -S . -B build
cmake --build build
Run:
./build/nonlinear-perceptron