Shift Window Transformer for CSI feedback in Massive MIMO

In massive multiple-input multiple-output (MIMO) systems, the accuracy of the downlink channel state information (CSI) plays an important role in improving system performance. In frequency division duplexing (FDD) mode, CSI has to be sent back by user equipment (UE) due to lack of channel reciprocity, which is challengeable. Deep learning (DL) based methods have been proposed to improve the recovery and reconstruction performance of CSI feedback. In DL-based methods, although transformer-based models have achieved the state-of-the-art performance, these models have high computational complexity. In this paper, we propose a shift window transformer framework to improve CSI feedback performance, which reduces the amount of calculation by calculating self-attention in sub-windows. Experimental results show that the proposed model has good trade-off between feedback performance and computation complexity, and outperforms other methods at compression ratios of 1/16 and 1/64 under indoor scenarios.