I have been running networks on Azure percept devkit and i am interested in knowing how to measure performance such as FPS for the networks. I have not found any documentation of that anywhere, will be a big help if i can get some info on this.
I have been running networks on Azure percept devkit and i am interested in knowing how to measure performance such as FPS for the networks. I have not found any documentation of that anywhere, will be a big help if i can get some info on this.
Hello @bm21416-7310 Welcome to Microsoft Q&A Platform and thanks for your query. Community SME's on this topic or our team will review your scenario and circle back at the possible earliest time.
Hi @bm21416-7310,
You can access your model's telemetry via these instructions:
https://docs.microsoft.com/en-us/azure/azure-percept/how-to-view-telemetry
Let me know if this gives you the data and information you're looking for.
Thanks!
@bm21416-7310 - unfortunately, the Azure Percept DK does not currently have an API in the Myriad X SDK suitable for profiling neural networks.
However, you could try to calculate the FPS by taking the difference of two consecutive frames in the log output. For example, the below two frames are about 400 milliseconds apart. Note that seqno is the frame ID in consecutive order. So if subsequent frame IDs are not one integer apart, it means a frame was dropped.
nn: seqno=2, ts=1615811179194168848, []
nn: seqno=3, ts=1615811179235832767, []
Thanks,
-Mike
Thanks I will check this out. By the way what batch sizes are these networks set to? in the model URLs here: https://github.com/microsoft/azure-percept-advanced-development
The Timestamps are Unix timestamps in nanoseconds since January 1st 1970. So when you see ts=1615920044588919099, what that means is Tue, Mar 16, 2021 @ 18:40:44 GMT. Note that we use 64-bit unsigned integers for the timestamp, so in case the device is still functioning 17 years from now, we won't overflow.
I realized I made a mistake when calculating the example I previously gave you. The formula for this example is (later frame - earlier frame)/1000000 = the time difference in milliseconds. So my previous example should be (1615811179235832767 - 1615811179194168848)/1000000 = ~42ms.
Here is a handy Unix timestamp converter: https://www.unixtimestamp.com/
3 people are following this question.