Build tf `r1.8` fails on jetson.
Closed this issue · 6 comments
This post is not quite an issue. 😄 According to tensorflow/tensorflow#18643, there are some problems to build tensorflow branch r1.8 on aarch64 (Jetson TX1 & 2).
The error msg looks like:
undefined reference to 'png_init_filter_functions_neon'
I'm posting here just to let more know the workaround (pr: tensorflow/tensorflow/pull/19075):
Change file third_party/png.BUILD in tensorflow repo, near line 38:
add line:
copts = ["-DPNG_ARM_NEON_OPT=0"],
bellow line: linkopts = ["-lm"],
After the amendment, rerun the BuildTensorFlow.sh script or the bazel build cmd.
Hello , Could you sharing your experience with installing TensorFlow on TX2 ?
I have flashed TX2 with Jetpack 3.2.
"$ sudo bash BuildTensorFlow.sh" failed.
Or could you leave a way to contact?
@DefineFC
I see google has included a "raspberry pi" fix in the master branch. I'm hoping this automagically migrates and works with the TX2. I am testing an update to my build script now. Hopefully all is smooth. Thank you for pointing this out. I'm undecided about including a patch in my helperscript or just leaving it out and putting a warning about r1.8 being broken.
@timoonboru
Just
Change file
third_party/png.BUILDin tensorflow repo, near line 38:
add line:
copts = ["-DPNG_ARM_NEON_OPT=0"],
bellow line:linkopts = ["-lm"],
And then, do what @JasonAtNvidia tells us to do. If you're still having trouble accomplishing, try sudo bash BuildTensorFlow.sh -b r1.7 as described in this repo.
Read the doc very carefully and do exactly as told to. There is no magic.
@JasonAtNvidia Hoping tf will fix that issue painlessly, too.
Jason, did you really build the wheel for Compute Capability 5.2, and 6.3 as described in README.md? IMHO, Compute Capability 5.3, and 6.2 seems more common for Jetson.
@DefineFC It is a typo, it is 5.3 for the TX1 and 6.2 for TX2. It would fail miserably if I tried building for CC6.3 because the Jetson is not capable of it.
Today I uploaded a git patch that takes care of the 2 files that need to be modified for a successful build. I also posted updated links to TF 1.8.0 wheel binaries to use directly without the need to build TF yourself.