FeatureFormer Predicts Neural Network Performance with Node-wise Encoding

Deploying neural networks on resource-constrained edge devices necessitates accurate prediction of their latency and energy consumption for efficient architecture search. Existing predictors, often based on Graph Neural Networks (GNNs) or transformers, tend to overlook the specific computational costs at the node level within the network. This oversight limits their ability to precisely model the performance of critical operations. To address this, the researchers developed FeatureFormer, a novel neural performance predictor. FeatureFormer integrates explicit node-wise encodings of key metrics like Floating Point Operations (FLOPs), parameter counts, and memory proxies directly into a gated graph attention architecture. Alongside this, they present NNEQ, a new large-scale dataset for unified evaluation of latency and energy prediction. Extensive testing shows FeatureFormer consistently outperforms state-of-the-art methods across both metrics, even in challenging out-of-domain scenarios. The proposed encoding method is also broadly applicable, enhancing existing predictors with minimal overhead.