Recipe: Live Training Dashboard¶

What you'll build: a GUI window showing XorNetwork training in real time. The loss curve animates as ImPlotMetricsSink logs each step and TrainingDashboard draws the result.

This is a GUI demo, not a console recipe. It needs -DPULSATRIX_ENABLE_VIZ=ON and a display.

CMake target: training_dashboard_demo (examples/viz/training_dashboard_demo.cpp):

cmake -S . -B build -DPULSATRIX_ENABLE_VIZ=ON
cmake --build build --target training_dashboard_demo --config Release

Run it: ./build/training_dashboard_demo (Windows: build\Release\training_dashboard_demo.exe).

Code¶

ImPlotMetricsSink sink;
VizWindow window("pulsatrix -- Training Dashboard Demo", 1000, 700);
window.run([&]() {
    for (int i = 0; i < 4 && step < kMaxSteps; ++i, ++step) {
        // ... build input and target_tensor for XOR example (step % 4) ...
        net.train_step(input, target_tensor, optimizer, sink, step);  // logs loss via sink
    }
    ImGui::Begin("Training Dashboard");
    TrainingDashboard::Draw(sink);
    ImGui::End();
});

Full source: examples/viz/training_dashboard_demo.cpp.

What you'll see¶

A window titled "pulsatrix -- Training Dashboard Demo" opens and starts training XorNetwork (Linear(2,4) -> ReLU -> Linear(4,1), Adam) right away. It runs a few steps per frame, so the loss curve animates instead of finishing before the first frame renders. The dashboard shows a compact summary line per metric (last, min and max value, current step) above a live line chart of the loss.

What's happening¶

ImPlotMetricsSink is a MetricsSink, the same logging interface every training loop in this codebase accepts. Live plotting therefore needs no changes to XorNetwork::train_step().

log_scalar() buffers each step's loss into a time series per metric name. That part always builds and is unit-tested even without PULSATRIX_ENABLE_VIZ. Only TrainingDashboard::Draw() touches ImPlot. It runs once per frame inside VizWindow::run()'s callback and draws the buffered series.