Deep learning inference in GNU Radio with ONNX
summary
gr-dnn is an open-source GNU Radio OOT block that runs ONNX models via ONNX Runtime (Python API) inside flowgraphs, requires GNU Radio 3.8+, and ships Docker images. Example: a 350k-parameter CNN for automatic modulation classification takes 128 complex IQ samples (flattened to 256 floats), outputs probabilities for 10 modulations, and achieves real-time inference only with GPU acceleration; CPU runs are slow and USB from the PlutoSDR becomes a bottleneck. Configure with input/output flatten sizes and batch size 200.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 2:24Open Neural Network eXchange (ONNX)
- 4:46ONNX support
- 7:31ONNX Runtime
- 12:11gr-dnn configuration
- 13:23Experimental setup
- 14:37Experimental model
- 15:46GNU Radio flowgraph
- 17:50GNU Radio output visualization
- 18:40External representation
- 19:16Performance
- 19:59Conclusions
- 21:52Resources
related talks
24mAn Open Channel Identifier using GNU Radio
14mBrain-on-Switch: Towards Advanced Intelligent Network Data Plane via NN-Driven Traffic...
27mEmbedded Deep Learning Super Resolution on GStreamer using ONNX Inference Runtime
26mDeep Learning in OpenCV
30mScalable ML acceleration with ONNX Runtime
32mMan or Machine?: Developing a "Turing Test" for Radio Intelligence
32mThe Role of GPUs in Modern SDR
08mLightning Talk - Deepwave Digital: Artificial Intelligence Radio Transceiver with ...