Python 3.7 on Windows.

[steffanjensen]

Please make sure that this is a feature request. As per our GitHub Policy, we only address code/doc bugs, performance issues, feature requests and build/installation issues on GitHub. tag:feature_template

System information

• TensorFlow version (you are using):
  Tried many.

• Are you willing to contribute it (Yes/No):
  Yes

Describe the feature and the current behavior/state.
Make Tensorflow work for python 3.7 on Windows.

Will this change the current api? How?
I don't know.

Who will benefit with this feature?
Users and developers.

Any Other info.
It's a huge issue, dlib has same problem.

[pandrey-fr]

Hi,
The latest builds are normally compatible with python 3.7, whatever the (non-exotic) OS. Could you please be a little more specific as to which version of TF you are targetting, which installation method you have tried and what problems (including error stacks) you have encountered?

[steffanjensen]

I am just targeting
pip install tensorflow

I have a repo with over 100 users, so many people are using windows.
https://github.com/instabotai/

When you enter
pip install tensorflow

In python 3.7 it is not recognized by pip.. So "nothing" happens"..

Also if i look in here:
https://www.tensorflow.org/install/pip#package-location

I can see tensorflow pip support linux python 3.7 but not windows python 3.7.

[mihaimaruseac]

There should be python3.7 for Windows too, we are building the pips.

[steffanjensen]

There should be python3.7 for Windows too, we are building the pips.

Thanks for your answer, then i will stick with tensorflow..

Are we talking a week before pip release, month or more ?

Because if it's more than i week, i will make a Docker option for my users with issues win/py3.7..

Thanks for answering me :)

[mihaimaruseac]

I'll be looking to see why they don't get released for 1.14.1, this should probably land next week or the one after that.

Keep also an eye on 2.0, there will be release candidates and if pip is missing please ping on here so we can look at why.

[steffanjensen]

Anything new?

[mihaimaruseac]

They will be provided with the new releases. There should be soon a release candidate 0 (RC0) for the next version and that one should be have the needed pips.

[steffanjensen]

something new ?

[mihaimaruseac]

There hasn't been a new release yet, but hopefully we'll have 2.0 RC0 early next week

[steffanjensen]

Great, people have been looking for this for years ;)

[mihaimaruseac]

TF 2.0 RC0 is out, and I see a Python 3.7 link for each os at https://pypi.org/project/tensorflow/2.0.0rc0/#files

Please test it and let us know if there are still issues with python 3.7 on Windows

[ebaj]

pip install tensorflow==2.0.0rc0
is still not working for 3.7 on Windows. Has anybody been able to get this to work without downgrading yet?

[steffanjensen]

it works on some windows system, it's all about the cpu :/

[mihaimaruseac]

@ebaj What version of pip you use? What CPU? What compiler? Which version of Windows? Using WSL?

[mihaimaruseac]

@ebaj, also please post output of pip debug --verbose

[kitesurf]

I'm having the same issues. Tried tensorflow tensorflow-gpu and all pip upgrades...
Windows Server 2019 with newest c++ on a Google Virtual Engine with Intel Xeon 2.2

here my pip debug --verbose:

pip version: pip 19.2.3 from c:\users\kitesurf\appdata\local\programs\python\python37-32\lib\site-packages\pip (python 3.7)
sys.version: 3.7.3 (v3.7.3:ef4ec6ed12, Mar 25 2019, 21:26:53) [MSC v.1916 32 bit (Intel)]
sys.executable: c:\users\kitesurf\appdata\local\programs\python\python37-32\python.exe
sys.getdefaultencoding: utf-8
sys.getfilesystemencoding: utf-8
locale.getpreferredencoding: cp1252
sys.platform: win32
sys.implementation:
  name: cpython
Config variable 'Py_DEBUG' is unset, Python ABI tag may be incorrect
Config variable 'WITH_PYMALLOC' is unset, Python ABI tag may be incorrect
Compatible tags: 14
  cp37-cp37m-win32
  cp37-none-win32
  py3-none-win32
  cp37-none-any
  cp3-none-any
  py37-none-any
  py3-none-any
  py36-none-any
  py35-none-any
  py34-none-any
  py33-none-any
  py32-none-any
  py31-none-any
  py30-none-any

[steffanjensen]

Yeah, I have seen many report unable to use TensorFlow on Windows Servers.

[mihaimaruseac]

@kitesurf Our windows tag is cp37-cp37m-win_amd64, but you have cp37-cp37m-win32. That is, we only release for 64 bits but your CPU is 32 bits, hence we don't have a pip for you.

You can try building from source

[steffanjensen]

@kitesurf Our windows tag is cp37-cp37m-win_amd64, but you have cp37-cp37m-win32. That is, we only release for 64 bits but your CPU is 32 bits, hence we don't have a pip for you.

You can try building from source

Why not isn't it easy to implement or are you guys jumping on the 64bit should be a standard thing for 2.0?

Remember many users are still using tensorflow for servers, most servers are 32bit.

[mihaimaruseac]

Our support matrix is already extremely large, supporting 32 bits on x86 CPUs will almost double it and this is very untractable.

[steffanjensen]

I know what you mean, but people want to use a package, all users easy can install. There is no point in making your app support TensorFlow, and then telling all your users which use 32 how to compile from source.

[ebaj]

@mihaimaruseac Is it really the CPU, or actually the type of Python installed?
My physical CPU is a 64-bit Intel but I was running 32-bit Python, as many users do, so platform.architecture was showing up as 32 bit, which is where I think the mismatch arose.
After testing with 64-bit Python 3.7.4, tensorflow installed successfully with pip.

I believe the "latest version" download from Python defaults to the 32-bit installer so I imagine many people will be in this situation, if this is indeed the source of the issue.
In addition, this thread on stackflow highlights why many people even on 64-bit machines still choose to run 32-bit Python.

With TensorFlow's well-earned popularity, I hope the resources will exist to support this userbase.

[steffanjensen]

Yes it works with python version 64 bit. You are correct on all your statements.
Most user by default gets confused, because more either download 32 bit because it's the default python version, or they use a VM or like my server a 32 bit CPU. I own a repo with hundreds of users and I get daily issues from people who need support for this issue. This issue is a must, it destroys it for people who want to use it and not spending hours daily because of TensorFlow issues users has.

I don't think the devs want this supported, they are devs not product creators. This would most likely not happen, they have a "dev" roadmap. They will wait 2 years until 64 bit becomes standard for python, they build for the "future".

I know many devs who does not understand how to compile tensorflow from source, so the normal users don't have a chance.

I am thinking about creating a package for all package mangers for my app, just to fix this problem.

Huge Issues

• Pip installer does not work with 32 bit python.
• Pip installer does also not work with 32 bit because of above issue.
• Python downloads 32 bit by default BIG PROBLEM

[pandrey-fr]

Most user by default gets confused, because more either download 32 bit because it's the default python version, or they use a VM or like my server a 32 bit CPU.

You know 64 bits processors have been the norm for ten years, right? And I don't know where you find your python downloads, but 64 bits is the default.

I don't think the devs want this supported, they are devs not product creators.

Skip the newspeak pal, and have a little respect for tensorflow devs. You can pretty easily compile a 32 bits version if you need; they already make an impressive job supporting a large number of 64 bits system, and you cannot blame them for putting their efforts into fixing bugs, adding functionalities and enhancing existing ones rather than avoiding you the not-so-tremendous effort to have to compile the package once in a while.

I am thinking about creating a package for all package mangers for my app, just to fix this problem.

Makes sense ; you have a use-case problem, make a use-case solution. That being said, instead of adding a package manager layer, I would intuitively recommend just packing (pre-compiled) dependencies with your app (e.g. with pex or docker). You know, like a dev deploying their tool might do.

[steffanjensen]

You are talking like I have the problem, yet it is my hundreds of users who has the problem. That you are trying to say "do this, do that". You don't understand the larger issue on scale.

When python is 32 bit by default, and you are talking about norms. Maybe you should follow the norm for python, so python users can install this package, instead of talking about what others do.

50% of all python users has issues installing TensorFlow, but that is just the norm for years. Maybe you should fix that?

But if you guys don't want to fix it, I will most likely just fix it myself, my users does not care about the norm, but just it works, maybe you should try that to.

I even have a docker, so because you don't want to support 32 bit your solution is, for users to go into the bios, enable Hyper-V go back install the docker and then install the package. haha

Also docker does not work on these version of windows:
Windows 10 Home; Windows 10 Mobile; Windows 10 Mobile Enterprise, because they does not support Hyper-V.

Which is 50% of the windows market

I will look into pex, sorry for rant i have just seen many post about this issue for years and i get over 10 messages from people daily, who asks me how to install TensorFlow, because of 32bit.

It's really frustrating to create an app and then TensorFlow fills 50% of the user support and 80% runs away because of Tensorflow installation issues.

[mihaimaruseac]

I'm disputing the 50% statistics, the 80% numbers, referencing a 10 years old Stack Overflow issue and many other anectdata in this thread.

However, this is getting offtopic and the main issue seems to be solved: python3.7 pips for Windows are now released.

For 32 bits architecture support, I created a feature request: #32315. Closing this one

[tensorflow-bot]

Are you satisfied with the resolution of your issue?
Yes
No

[steffanjensen]

#32627