fpgasystems / spooNN

FPGA-based neural network inference project with an end-to-end approach (from training to implementation to deployment)

spooNN

This is a repository for FPGA-based neural network inference, that delivered the highest FPS in the international contest for object detection as part of Design Automation Conference 2018 and 2019 (https://www.dac.com/content/2018-system-design-contest). The contents of spooNN enable an end-to-end capability to perform inference on FPGAs; starting from training scripts using Tensorflow to deployment on hardware. Target hardware platforms are PYNQ (http://www.pynq.io/) and ULTRA96 (https://www.96boards.org/product/ultra96/).

2018: The final rankings are published at http://www.cse.cuhk.edu.hk/~byu/2018-DAC-SDC/index.html

2019: The final rankings are published at http://www.cse.cuhk.edu.hk/~byu/2019-DAC-SDC/index.html

Repo organization

• hls-nn-lib: A neural network inference library implemented in C for Vivado High Level Synthesis (HLS).
• mnist-cnn: helloworld project, showing an end-to-end flow (training, implementation, FPGA deployment) for MNIST handwritted digit classification with a convolutional neural network.
• halfsqueezenet (targets PYNQ): The object detection network, that ranked second in DAC 2018 contest, delivering the highest FPS at lowest power consumption for object detection.
• recthalfsqznet (targets ULTRA96): The object detection network, that ranked second in DAC 2019 contest, delivering the highest FPS at lowest power consumption for object detection.