Skip to main content

Table 2 GPU memory/FLOPs/Params/RunTime comparison on kernel estimation layer (KEL), self-attention (SA), interlaced sparse self-attention layer (SSA) and the proposed approach

From: Sparse self-attention aggregation networks for neural sequence slice interpolation

  Memory(GB) FLOPs(G) Params(M) RunTime(ms)
KEL [9] - 138.4 - 6160
SA [27] 256.0 6599 0.0052 -
SSA [30] 9.775 81.24 0.0104 327
AAL(Ours) 2.331 10.42 0.0156 275
  1. All the numbers are tested on a single P40 GPU with CUDA10.2 and an input feature map of 1×64×512×512 during inference stage. Lower Memory, FLOPs and RunTime indicate better performance
\