1

I have to measure the inference time of my ML Model in Swift. To do so, I wanted to perform the inference several times and then average all inference times to have a more precise value. I noted that the inference time grew with the iterations. I must have done something wrong, I'm not familiar with Swift:

Duration :  4756 ms
Duration :  4879 ms
Duration :  5325 ms
Duration :  5712 ms
Duration :  5952 ms
Duration :  6059 ms
Duration :  6223 ms
Duration :  6244 ms
Duration :  6088 ms
Duration :  6286 ms

Here is a snippet of my code:

for _ in 1...nb_it {

    // Timer
    let tic = CFAbsoluteTimeGetCurrent()

    // Inference    
    _ = try model.prediction(input: input, options: options)

    // Timer
    let toc = CFAbsoluteTimeGetCurrent()
    let duration = Int32(1000 * (toc - tic))

    // Report
    print(String(format: "Duration : %5d ms", duration))
}

Is it normal? Any idea if it's not? Thanks in advance!

Vincent Garcia
  • 635
  • 10
  • 19

1 Answers1

2

Yes, this is possible if the phone gets hot. Your model seems to be pretty large (4 seconds per prediction) so it is doing a lot of work, heating up the CPU and/or GPU. The phone will slow down to avoid overheating.

Matthijs Hollemans
  • 7,706
  • 2
  • 16
  • 23