propose a method that focus on GPU latency, reach compatative results to MobileNet.

• tension among: # of params, acc, latency

  • of params should be resonably large

  • faster on GPU, while remains similar acc.
  • def a reward: acc * (latency / tar) ^ omega
  • latency relationship shown by random network suggest no obvious linear correspondence ==> design network with architecture in mind!
  • suffer from underfitting
• convert reward maxmization to LP
• fast eval of acc and latency
  • eval: FairNAS(supernet?)
  • latency: lookup table
• GPU: fast conv, less (% of) time