Fig. 13: Attention map of the self-attention module (400 × 400) from the last block of 4DVarFormer.
The self-attention in each head exhibits a nearly full-rank attention map, indicating that 4DVarFormer can learn a wide range of features.
Fig. 13: Attention map of the self-attention module (400 × 400) from the last block of 4DVarFormer. | npj Climate and Atmospheric Science
