Last active
October 11, 2022 16:14
-
-
Save carlthome/6ae8a570e21069c60708017e3f96c9fd to your computer and use it in GitHub Desktop.
Example of how to use XLA AOT via tfcompile to build a Keras model into a shared library.
Model => 150 ms ± 199 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
XLA binary => 191 ms ± 604 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
Why XLA compiled binary is slower than the model itself?
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
same