Skip to main content
Overview

Model Compression Competition

November 28, 2021
1 min read

Perspectives on compression

  • Reduce model size (= number of parameters)
  • Just make it faster
  • Reduce the number of operations

FLOPs

  • A factor representing the number of operations
  • An indirect factor determining computation speed

The ShuffleNetv2 paper proposed the following guidelines, considering factors beyond FLOPs that affect speed:

  • Memory access cost is lowest when input and output sizes are equal
  • Large group convolutions increase memory cost
  • Structures with multiple branching paths — that is, models configured in parallel — cause speed degradation
  • Element-wise operations have a non-negligible impact, so be careful with them
Loading comments...