ONNX PoC: Integrate Basic Inference Block for Signal Classification and Peak Detection

[RalphSteinhagen]

ONNX PoC:
Minimal Inference Block for Time-Series Classification (e.g. Peak Detection, Modulation Labelling)

This task explores the integration of ONNX Runtime into GNU Radio 4.0 by implementing a minimal inference block capable of applying pretrained models to real-time or buffered time-series input.

Scope

• Create a basic ONNX inference block that accepts 1D float input vectors (or std::complex<float> magnitudes).
• Load and evaluate a fixed ONNX model at runtime (e.g. MLP or CNN).
• Apply this to minimal signal processing MVPs such as:
  • Multi-peak detection in noisy FFT spectra
  • Modulation classification
  • Anomaly detection / signal presence
• Use ONNX Runtime with CPU backend only (no GPU offload or EP integration).
• Accept static shape models (initially), with input validation logic to match.
• Optional: visualise output via debug log or TagEmitter.

Design Notes

• Avoid complex tokenisation or postprocessing — focus on direct tensor-in, tensor-out logic.
• Reuse existing GR4 block model infrastructure.
• Support importing .onnx model path as a parameter.
• Add a onnx::Block namespace for future extension (e.g. quantised models, streaming RNNs).
• Potential model sources:
  • [Whisper-tiny] (as a stretch goal or toy speech demo)
  • [Coqui-TTS] (for future inverse/response model)
  • Small 1D CNN/MLP trained in scikit-learn, PyTorch, or Keras and exported to ONNX.

Rationale

This PoC is meant to establish a minimal, testable bridge between GNU Radio and ONNX Runtime, enabling future DSP+ML hybrid pipelines. Rather than integrating large and complex models up front, the focus is on:

• Simplicity of integration
• Practical MVPs in the RF/timeseries domain
• Documenting and resolving typical integration pitfalls (e.g. shape inference, memory reuse, ONNX session lifecycle)

Large models such as Whisper or TTS are not the primary goal here — they serve as long-term examples, not initial targets.