Lightweight neural network based SNIP for CSI feedback in massive MIMO

CsiNet, a deep neural network framework, utilizes an autoencoder to efficiently transmit downlink channel state information (CSI) in the feedback link, which reduces the cost of feedback, and significantly improves the quality of the reconstruction. However, the model with massive parameters incurs a lot of storage space and high computational complexity, which is impractical for low-cost and low-power edge devices. In this work, a lightweight CsiNet based SNIP (Single-shot Network Pruning) is implemented, which prunes the model using the gradient information of the first training epoch, eliminating both pretraining and the complex pruning schedule. Numerical simulation results show that, under the same compression rates, the method can achieve a similar or even better reconstruction effect and more effectively reduce computational complexity, compared to traditional lightweight methods. We propose a lightweight CsiNet model based on SNIP. SNIP-based CsiNet model is more suitable for practical engineering applications, especially resource-constrained mobile edge devices.image