4 ms·

Tensorflow 1.6.0 Released

125 points by MichaelBurge 9y ago

minimaxir 9y ago
1.6.0 does not appear to be installable via pip yet, though. (you can install tensorflow==1.6.0rc1 if you really want it now) The inclusion of AVX instructions should give model training a notable speed boost at lower numbers of CPUs.
- puzzle 9y ago
  I thought the plan was to have common code for the generic, non-critical parts like graph handling, while operators would be shipped in multiple variants, keyed and loaded at runtime by CPUID. You can even delegate that to glibc if they are packaged as shared libraries and it's a recent enough version. Maybe in 1.7?
  - netheril96 9y ago
    Do you have any evidence that the TF developers are going to do that?
    - puzzle 9y ago
      I heard it from someone at Google at a public event, but there was no date attached to it. I should have asked.
- ajay-d 9y ago
  Looks like it's on pypi now https://pypi.python.org/pypi/tensorflow-gpu https://pypi.python.org/pypi/tensorflow-gpu
tormeh 9y ago
Any idea why AMDs fork[0] hasn't been merged yet? 0: https://github.com/ROCmSoftwarePlatform/hiptensorflow https://github.com/ROCmSoftwarePlatform/hiptensorflow
- ma2rten 9y ago
  There is some discussion on it here: https://github.com/tensorflow/tensorflow/issues/22 https://github.com/tensorflow/tensorflow/issues/22
  - mastazi 9y ago
    While that thread mentions the HIP port at some point, it's really about the OpenCL port.
xvilka 9y ago
Still no OpenCL support...
- make3 9y ago
  you're complaining about a free, open source project. your contributions are welcome
  - loranbriggs 9y ago
    +1
  - ancarda 9y ago
    I get the sentiment, but how many people can just casually add OpenCL to Tensorflow? For a lot of people, even if talented programmers, they may not be able to do anything but ask and beg. My webcam doesn’t work with Linux. Should I write a driver despite not knowing the kernel API, C, or how webcams even work? I would be better off sponsoring development somehow — this is something I’ve wondered if it’s possible to do. Someone with the skills could be paid by me, and other affected users.
    - ndr 9y ago
      Asking nicely would help. Offering to pay someone to put in the work would be even better.
- lallysingh 9y ago
  https://github.com/tensorflow/tensorflow/issues/22 https://github.com/tensorflow/tensorflow/issues/22 from above

125 points by MichaelBurge 9y ago

1.6.0 does not appear to be installable via pip yet, though. (you can install tensorflow==1.6.0rc1 if you really want it now) The inclusion of AVX instructions should give model training a notable speed boost at lower numbers of CPUs.

puzzle 9y ago
I thought the plan was to have common code for the generic, non-critical parts like graph handling, while operators would be shipped in multiple variants, keyed and loaded at runtime by CPUID. You can even delegate that to glibc if they are packaged as shared libraries and it's a recent enough version. Maybe in 1.7?
- netheril96 9y ago
  Do you have any evidence that the TF developers are going to do that?
  - puzzle 9y ago
    I heard it from someone at Google at a public event, but there was no date attached to it. I should have asked.
ajay-d 9y ago
Looks like it's on pypi now https://pypi.python.org/pypi/tensorflow-gpu https://pypi.python.org/pypi/tensorflow-gpu

tormeh 9y ago

Any idea why AMDs fork[0] hasn't been merged yet? 0: https://github.com/ROCmSoftwarePlatform/hiptensorflow https://github.com/ROCmSoftwarePlatform/hiptensorflow

ma2rten 9y ago
There is some discussion on it here: https://github.com/tensorflow/tensorflow/issues/22 https://github.com/tensorflow/tensorflow/issues/22
- mastazi 9y ago
  While that thread mentions the HIP port at some point, it's really about the OpenCL port.

xvilka 9y ago

Still no OpenCL support...

make3 9y ago
you're complaining about a free, open source project. your contributions are welcome
- loranbriggs 9y ago
  +1
- ancarda 9y ago
  I get the sentiment, but how many people can just casually add OpenCL to Tensorflow? For a lot of people, even if talented programmers, they may not be able to do anything but ask and beg. My webcam doesn’t work with Linux. Should I write a driver despite not knowing the kernel API, C, or how webcams even work? I would be better off sponsoring development somehow — this is something I’ve wondered if it’s possible to do. Someone with the skills could be paid by me, and other affected users.
  - ndr 9y ago
    Asking nicely would help. Offering to pay someone to put in the work would be even better.
lallysingh 9y ago
https://github.com/tensorflow/tensorflow/issues/22 https://github.com/tensorflow/tensorflow/issues/22 from above