koodg123 / resnet_fpga

UCSD CSE 237D Spring '20 Course Project

ResNet on FPGA

This is a project at UCSD as part of both the course CSE 237D (SP'20) and the research on IoT neural networks in Ryan Kastner's group.

Overview

Automatic modulation classification of radio signals is useful for spectrum sensing but still require lots of improvement. Neural network (NN) solutions like ResNets are state-of-the-art for this but need an efficient hardware implementation. We propose that an FPGA implementation would be able to reduce memory footprint and improve latency.

Goals

• Implementation with baseline accuracy (PyTorch)
• Quantization of NNs (Brevitas, TWN Generator)
• Generation of RTL design (Vivado HLS)
• Target test results (RFSoC ZCU111)

Please glance at the commit messages to see progress.

Team

• Kartik Kulgod (SW)
• Zesen Zhang (SW)
• Nitish Kulshrestha (HW)

References

• Real-time Automatic Modulation Classification using RFSoC
• OTA Deep Learning based MC: https://arxiv.org/pdf/1712.04578.pdf
• Unrolling ternary NN: https://arxiv.org/pdf/1909.04509.pdf