Closed adjhawar closed 2 years ago
Hi @adjhawar
You could use the -e option in ExecuteNetwork to generate profiling information, ideally redirecting it to a file with >. Then comparing the two files should reveal the answer.
Hope this helps.
Closing as I think this has been resolved. If you need more help please open another issue.
I am trying to run onnx models on mobile device using ArmNN ( neon ) as the execution provider. When I run a model on the device, I have observed a time difference of 1-1.5 seconds between the first run and the subsequent runs. Is this behaviour expected?
What I have observed that the initial inference run does certain computations which are missing in the next runs. Below are the log snippets:
First Run: Time between A and B is around 60-90 ms
Second Run: Time Between A and B is around 5-7 ms
In the second run, the log with ArmNN 2D convolution is not there. I even tried looking for the conv.cc that is being referred in the logs but could not find it anywhere. I am stuck here, trying to find out the cause of the difference in inference time. Any leads on this will be highly appreciated.