Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Activity
    Sean Morgan
    @seanpmorgan
    Thanks!
    Sean Morgan
    @seanpmorgan
    Now that windows build time is drastically reduced, would it be possible to publish tf-nightly for windows as well?
    Gunhan Gulsoy
    @gunan
    Technically, we never stopped
    but the build was never green, so autopushes did not trigger when the build is broken
    Sorry, I know that does not help you
    So, the actual helpful comment would be, we should look into getting the builds green to get it to autopush
    Sean Morgan
    @seanpmorgan
    Thanks for info. So it looks like there are windows whls on tf-nightly now (not sure how recently these have been going through).
    But yeah that solves my problem thx!
    Gunhan Gulsoy
    @gunan
    awesome! so someone did fix them already :)
    After new years, we will also look into python 3.8
    looks like grps is finally out with their new release
    Sean Morgan
    @seanpmorgan
    🎉
    Clayne Robison
    @claynerobison
    What time is the meeting on Jan 7? (Is there a meeting?)
    Sean Morgan
    @seanpmorgan
    Looks like py27 windows wheel is missing from 2.1 release? https://pypi.org/project/tensorflow/2.1.0/#files
    Is py27 not supported for windows (it's not in the nightlies either)
    Sean Morgan
    @seanpmorgan
    Hmmm wasn't part of 2.0 release either.. guess its just not supported.... Yeah "TensorFlow is not supported on Windows with Python 2.7"
    Gunhan Gulsoy
    @gunan
    on windows, we never had python 2.7 support.
    I think python 2.7 distributions on winodws are built with a too old version of visual studio, one that does not have C++11 support
    but I may be just making it up, we dealt with this 3 years ago, I am getting old :P
    Jonathan J. Helmus
    @jjhelmus
    Python 2.7 and extensions are typically built with VS 2008 on Windows.
    Abin Shahab
    @ashahab
    Hi, What's the plan regarding 1.x line of Tensorflow, is it going to be deprecated? Is there some documentation with then regarding 1.x vs 2.x?
    I saw that 1.15 is slated to be the last 1.x release https://github.com/tensorflow/tensorflow/releases/tag/v1.15.0
    Gunhan Gulsoy
    @gunan
    @jjhelmus yes! that was it!
    I remember if we did not rebuild python from sources, TF did not work.
    @ashahab yes, 1.15 is the last release, and no new features will be merged to 1.x branches
    all development is at head, and will be cut into 2.x releases.
    Jason Zaman
    @perfinion
    oh awesome there are now power systems on GCP, that'll make testing way easier
    jayfurmanek
    @jayfurmanek

    Hi Guys! Is there a plan of record for low-precision hardware support in Tensorflow? The obvious answer I guess is "Eigen" but eigen doesn't do anything below fp16.
    Low precision fucntion is better supported in TF Lite (via gemmlowp), but there is also a new library called ruy (https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/experimental/ruy) that might be the future of low pecision function in TF/TFLite ?

    We're looking at hardware that may include some native low precision modes so I'm exploring for the appropriate places to add it. Do you guys know? Or is lowp something only the TFLite team cares about?

    Austin Anderson
    @angerson
    Hmm... I'm not aware of any plans like that; it seems like something the TF Lite team would be more concerned with at first glance.
    Austin Anderson
    @angerson
    Let me ask around a little
    jayfurmanek
    @jayfurmanek
    Thanks Austin
    Clayne Robison
    @claynerobison
    This was a bit of a challenge for us with the MKL-DNN int8 support on Xeon, as TFLite didn't fit our architectural needs. We ended up having to create some custom ops and distribute some of our own quantization tools to really take advantage of it.
    Austin Anderson
    @angerson
    jayfurmanek: It seems like TF Lite is the big place for this currently, while "it's still tbd what the low precision kernel situation will be for main TF
    Raziel suggested you can also send hardware questions to thirup@google.com, one of the program managers (iirc) who can track outside interest in these hardware related issues.
    jayfurmanek
    @jayfurmanek
    Sounds good - thanks for the suggestions. Was there any word on if ruy vs gemmlowp was the preferred target in TFLite?
    Austin Anderson
    @angerson
    jayfurmanek: Raziel says ruy, as it's to be the replacement for gemmlowp.
    (also, ruy is getting a lot more development, whereas gemmlowp is mostly static right now)
    Fei Hu
    @feihugis
    Does anyone successfully build the master branch of TensorFlow on CentOS 7 with Bazel 1.2.1? Here is an issue:https://github.com/tensorflow/tensorflow/issues/35867#issuecomment-574879243
    William D. Irons
    @wdirons
    To get around errors like that, I build with BAZEL_LINKLIBS=-l%:libstdc++.a bazel build -c opt //tensorflow/tools/pip_package:build_pip_package Give that a shot and see if it works for you.
    Fei Hu
    @feihugis
    @wdirons Yeah, it works!! Thank you very much!
    Austin Anderson
    @angerson
    Just a reminder: our February meeting is tomorrow, Tuesday the 4th, at 2pm PST. Details here, feel free to suggest agenda items: http://bit.ly/tf-sig-build-notes
    Reuben Morais
    @reuben
    hi all, sorry if this is not the appropriate place to ask, but I'm trying to understand how TF builds its Python package, and I can't figure out why there's a bunch of logic around installing header files in tensorflow/tools/pip_package/setup.py. my understanding is that all of the native code and SWIG stuff is handled by Bazel, and the only way to install the tensorflow package is via binary wheels, so why all this header file shuffling?
    jayfurmanek
    @jayfurmanek
    The comment there kind of explains it, I think:
    """Override how headers are copied.
    The install_headers that comes with setuptools copies all files to
    the same directory. But we need the files to be in a specific directory
    hierarchy for -I <include_dir> to work correctly.
    """
    Reuben Morais
    @reuben
    well, I guess my question then is: for -I <include_dir> to work correctly from where? who is depending on this file structure?
    William D. Irons
    @wdirons
    The binary PIP package installs the header files and the library that you need to compile your op in locations that are system specific.
    Reuben Morais
    @reuben
    aha
    @wdirons awesome, thanks
    William D. Irons
    @wdirons
    The minimal bazel version for master needs to be moved up to 2.0 because of tensorflow/tensorflow@09fe958.
    Subin Modeel
    @sub-mod
    fyi manylinux2014 builders : pypa/manylinux#411