Question on performance

[previ]

Device: RaspberryPi3+:

CPU architecture: armv7l

OS: Raspbian stretch

Conditions
I installed tensor flow lite for python 3.5, multi threaded (this wheel: tensorflow-1.14.0-cp35-cp35m-linux_armv7l.whl).
I run the test as documented in the readme.

Question
I cannot get better timing than 0.27 seconds with num_threads=4 while the readme reports a much faster 0.16.
Any hints?
Thanks!

[PINTO0309]

Please try tensorflow-1.13.1-cp35-cp35m-linux_armv7l.whl or tensorflow-1.11.0-cp35-cp35m-linux_armv7l.whl. Tensorflow is bloated as the version goes up. I have not yet clearly compared the performance.

[previ]

I did some more tests, I could not get any better timing until, guess what? I switched the power input from my PC USB output to a proper 2.1A power supply.
Now I get timings consistent with your test, actually as short as:
0.127 seconds with ts 1.14
0.371 with ts 1.13.1
0.317 with ts 1.11

So 1.14 looks much faster than predecessors, I did not try 2.0 beta.

Thanks

[PINTO0309]

@previ
It is a very interesting result. By the way, I am using a 5V 3.0A high power AC adapter. It is only speculation, but the following may have an impact.

1. Power supply capacity of AC adapter
2. Difference in weight loading time between the first and second inference processes

[JCStarry]

@PINTO0309 I test on raspberry pi4 B, python 3.6. it very fast!

[PINTO0309]

@JCStarry
I encourage you to try the following. It may be a little faster.

Tensorflow Lite stand alone Installer
https://github.com/PINTO0309/TensorflowLite-bin.git

MobileNetV2
https://storage.googleapis.com/download.tensorflow.org/models/tflite_11_05_08/mobilenet_v2_1.0_224_quant.tgz

tensorflow/tensorflow#31229 (comment)

[JCStarry]

@JCStarry
I encourage you to try the following. It may be a little faster.

Tensorflow Lite stand alone Installer
https://github.com/PINTO0309/TensorflowLite-bin.git

MobileNetV2
https://storage.googleapis.com/download.tensorflow.org/models/tflite_11_05_08/mobilenet_v2_1.0_224_quant.tgz

tensorflow/tensorflow#31229 (comment)

This is the result of testing MobileNetV2 ，ありがとう！！for your work
@PINTO0309

[PINTO0309]

As there has been no progress, I'm closing it.