time spent difference between rknn.inference and rknn.eval_perf

[复制链接] · 发表于 2019-3-15 16:48:00

I ran 100 times of rknn.inference() of mobilenet-v1-ssd and measure average time of each run using time.time(). It is

Time: 0.0248 sec

However, rknn.eval_perf() shows:

========================================================================
Total Time(us): 13779 (0.013779 sec)
FPS: 72.57
========================================================================

Both has the same input image, How can rknn_eval time be almost half of rknn.inference? Which one is more accurate?

Thanks,

只看该作者 · 发表于 2019-3-15 17:01:40

好像是常规现象。
我自己网络 rknn.eval_perf() 返回的时间大概是 rknn.inference() 的70%

只看该作者 · 发表于 2019-3-15 17:13:22

I have the same question

只看该作者 · 发表于 2019-3-18 16:04:02

rknn.inference()比rknn.eval_perf()耗时久是正常的。
inference()包含了数据通信拷贝等一系列额外的开销。
eval_perf()是用作性能评估，仅统计网络推理运行时间。