Performance Analysis of Deep Learning Inference on the Banana Pi BPI-F3 Board Using the Image Classification Problem as an Example

The paper analyzes the inference performance of the well-known neural networks ResNet-50 and MobileNetV2, which provide a solution for the problem of image classification, on the Banana Pi BPI-F3 board, which is built on the RISC-V architecture. The inference is launched by available frameworks: PyTorch, TensorFlow Lite, Apache TVM and ExecuTorch. The models are converted to the format of each target framework. The correctness of the problem solving is checked using the obtained neural networks. It is demonstrated that the accuracy indicators of image classification using these models correlate well with the published ones. Then, the optimal parameters for launching the inference for each framework and model are selected. A comparative analysis of the inference performance shows that ExecuTorch demonstrates the best results for both models. For the ResNet-50 model, the number of frames processed per second (FPS) varies from 2.649 to 3.339 fps with optimal parameters depending on the batch size of images processed in one forward pass through the network, for MobileNetV2 – from 11.26 to 29.96 fps. TensorFlow Lite is inferior to ExecuTorch by an average of ~ 2.1 times. PyTorch and Apache TVM demonstrate lower performance indicators. Probably, this is due to the fact that they are not fully optimized for the RISC-V architecture.