1

Hi I'm using the tensorflow ahead-of-time compilation(AOT) to compile tensorflow graph to self-contained binary. After switching from tf v1.14 to v2.2, the running time of the compiled binary increase by more than twice. Is this expected? If so, is there optimizations that I can do? If the AOT is by default using eager execution, is there a way to disable it, as this post mentions that disabling eager execution would make it faster.

Maggio Veneto
  • 11
  • 1

1 Answers1

0

Since Tensorflow 2.3 has addressed many of the performance issues, you should be able to find the better performance by upgrading to Tensorflow 2.3.

Disabling the eager execution is not the ideal way in TF2, instead you can use tf.function for performance optimized graphs.

You can follow this post for more details and benchmark on TF1 vs TF2 in this post.