Performance Evaluation of Apache TVM’s MetaSchedule for Image Classification Models on Edge Device

Edge deep learning is rapidly being deployed in consumer devices such as home appliances, car navigation systems, and IoT devices. Given these devices’ limited computational power and memory, optimizing deep learning models, expressed as tensor programs, is essential. This paper evaluates the inference performance of Apache TVM’s MetaSchedule for optimizing image classification models on edge devices. MetaSchedule, a part of the Apache TVM open-source machine learning compiler, automates tensor program optimization for various hardware platforms. Our evaluation of 52 image classification models shows that MetaSchedule outperforms ONNX Runtime in 45 models with up to 4.77x speedup (1.68x on average). However, MetaSchedule lags in models like EfficientNet-v2, VGG, and ResNet50 with larger batch sizes. Future work will focus on expanding model support for MetaSchedule and enhancing optimization performance, contributing to low-latency, privacy-preserving, and bandwidth-efficient inference on edge devices.