Pruning remote photoplethysmography networks using weight-gradient joint criterion

With the rapid advancement of remote photoplethysmography (rPPG), there is an urgent need to deploy rPPG algorithms on edge devices for efficient and accurate inference. However, due to limited computational resources, many rPPG neural networks require tailoring before they can be applied to these devices. Most existing network pruning algorithms rely on a single indicator to measure the importance of a connection, often resulting in the premature removal of crucial connections during the early stages of training. In this paper, we propose a novel pruning scheme that jointly considers the weight and gradient of a connection as the importance metric, while also taking into account the dynamics of the connection during the training process. Specifically, connections with large weights and small gradients are identified as stable and important, and should be retained. Secondly, connections with small weights and large gradients, although potentially significant for development, are likely to be removed but should be allowed to regenerate. Additionally, connections with small weights and small gradients, which are stable and necessary, are also considered. An importance indicator is designed for each of these three types of connections and is utilized in the drop, regenerate, and trim steps, respectively. The proposed pruning scheme is evaluated on two existing networks (DeeprPPG and PhysNet) using the PURE dataset. The results demonstrate that our approach possesses smaller network sparsity, fewer parameters, and fewer floating-point operations (FLOPs) to achieve a given level of accuracy compared to existing pruning methods. This study validates the feasibility of fine-grained pruning for small networks and highlights the effectiveness of considering the dynamics of connections during the training process.