According to the paper, The accuracy and runtime by the algorithm in the paper are 74.96 and 79.48 ms,
while the accuracy and runtime by random select are 73.78 ± 0.85 and 77.31 ± 0.9 ms.
Although the paper claim that random search cannot beat the NAS result by accuracy (73.78 + 0.85 = 74.63 < 74.96), how could you assure that result is not caused by larger runtime? For 77.31 + 0.9 = 78.21 ms < 79.48 ms.
If so, we don't need to calculate CE loss. We only need to randomly select a subset of networks according to the desired latency, and choose the best-performed network.
According to the paper, The accuracy and runtime by the algorithm in the paper are 74.96 and 79.48 ms, while the accuracy and runtime by random select are 73.78 ± 0.85 and 77.31 ± 0.9 ms.
Although the paper claim that random search cannot beat the NAS result by accuracy (73.78 + 0.85 = 74.63 < 74.96), how could you assure that result is not caused by larger runtime? For 77.31 + 0.9 = 78.21 ms < 79.48 ms.
If so, we don't need to calculate CE loss. We only need to randomly select a subset of networks according to the desired latency, and choose the best-performed network.