Accelerating Neural ODE Based CNN Models on Low-Cost FPGAs

ODENet is a deep neural network architecture in which a stacking structure of ResNet is implemented with an ordinary differential equation (ODE) solver. ANODE is an extended approach of ODENet for stably high accuracy by introducing more precise training algorithm using more memory in training phase. It is also possible to improve the accuracy while keeping the same number of parameters on resource-limited edge devices. In this paper, using Euler method as an ODE solver, a part of ODENet and ANODE are implemented as a dedicated logic on a low-cost FPGA (Field-Programmable Gate Array) board, such as PYNQ-Z2 board. As ODENet variants, reduced ODENets (rODENets) each of which heavily uses a part of ODENet layers and reduces/eliminates some layers differently are proposed and analyzed for low-cost FPGA implementation. In addition, rANODE is proposed as an ANODE variant in a way similar to the rODENet. They are evaluated in terms of parameter size, accuracy, execution time, and resource utilization on the FPGA. The results show that an overall execution time of rODENet and rANODE variants is improved by up to 2.66 times compared to a pure software execution while keeping a comparable accuracy to the original ODENet and ANODE.