Hello PYNQ and RFSoC community!
I wanted to share a project I’ve been working on as part of the StrathSDR group at the University of Strathclyde: a real-time modulation classification demo running on the AUP RFSoC platforms with a streaming CNN accelerator.
This design connects the RF-ADCs directly to a CNN model operating at the line-rate, performing real-time classification of modulation schemes from a live signal sent in loopback. It’s all implemented on the RFSoC with PYNQ, showcasing how deep learning can be tightly integrated with RF signal processing.
This work also demonstrates the effects of training a CNN model using Quantisation-Aware Training (QAT) and Post-Training Quantisation (PTQ), and integrates the resulting models in a real-time system.
For both QAT and PTQ, four different fixed-point weight precisions are compared in real time (16-, 8-, 4-, and 2-bits).
The repository provides an installable demo to try out for yourself. It also includes the training scripts, CNN IP source models/code, and the DeepRFSoC dataset.
Looking forward to your thoughts and discussions!