1

I would like to know if it is possible to optimize the image recognition of Tensorflow (InceptionV.3). Indeed, I would like to introduce the latter on a VPS equipped only CPU. Now, I manage to get a result after 30-40 seconds, and I would like to know if there would not be a solution to otherwise reduce this time: / (Without GPU)

I get this warning when executing of this code:

2017-07-15 02: 40: 19.230276: W tensorflow / core / platform / cpu_feature_guard.cc: 45] The TensorFlow library was not compiled to use SSE4.1 instructions, but these are CPU computations.
2017-07-15 02: 40: 19.230359: W tensorflow / core / platform / cpu_feature_guard.cc: 45] The TensorFlow library was not compiled to use SSE4.2 instructions, but these are CPU computations.

Do you think the compilation will greatly reduce this time? I have not found any documents on their site that speak of this warning.  I'm new to Tensorflow, so I do not understand everything ... I'm on ubuntu 16.04 LTS, and i use python 2.7.12

Thank you in advance ! :)

Marshall Cocop
  • 151
  • 3
  • 16

1 Answers1

1

You get this warning because your CPU has features that tensorflow can use to work faster, but the "build" you downloaded does not support. The solution is to rebuild tensorflow from source, And yes it will work faster for you 99% of the time if that's the warning you get.

To receive instructions on how to build tensorflow from source - Check here

When running the bazel build command, be sure to use the --copt=-msse4.2 flag

Ofer Sadan
  • 11,391
  • 5
  • 38
  • 62