I think its a great suggestion!

Same, I was kind of disappointed to see no mention of Windows in the download and install page.

Anyone have an idea of what the major incompatibilities/accommodations are? Is it it mostly issues with file paths, etc?

It's built with Bazel, which only supports linux/mac, but the good news is that windows support for bazel seems will be out by the end of this year.

:thumbsup:

You can already use TensorFlow on a Windows machine by using Docker.
Details are in this thread.
https://github.com/tensorflow/tensorflow/issues/42

I would like to use TensorFlow on Windows without Docker, in order to use GPU compute. Using Docker in this case is not using a Windows container, but a Linux virtual machine on Hyper-V or VirtualBox, and so the GPU will not be passed through.

:+1:

google, give us windows support , please!

+1/0.0

+1

Yeah, windows support would be super nice.

+1
+1
+1

yes please

Would love to use TensorFlow on Windows (native, not in a VM).

I'll take a look into whether Continuum can provide a conda package for tensorflow.

Judging by the use of bazel across the documentation, I assume its a matter of waiting for bazel to support windows. Is there anything specific to Tensorflow that would need to be addressed for windows to be supported, or is just bazel?

One question I have is when Bazel support is actually arriving on Windows. Looking at the Bazel repository it says they are planning to support Android in Windows, but I didn't see any reference to building (what I assume are) native packages.

Here is the bug we are using to track Bazel support for Windows: https://github.com/bazelbuild/bazel/issues/276. A month ago, @dslomov was able to get Bazel to bootstrap itself on Windows. The plan is to finish Windows support by end of this year.

@davidzchen I was actually working with that repo earlier this afternoon to try and give it a go without much luck, I might play around with it some more to see if I can get it to successfully bootstrap itself.

+1 native support on windows should be available

+1

or maybe we could port it to a more reasonable build system. perhaps cmake.

@ahmadia do you have any tips on how to install tensorflow using conda?

Right now we're in the same boat as everybody else because there is no port of TensorFlow to Windows. If somebody can put together a Windows port I'm happy to help with the binary build/deployment. Somebody has already put together a recipe for OS X/Linux available with conda install -c memex tensorflow.

+1 for this

+1

+1
also cmake support would be great

I put together an article with instructions and screenshots for getting TensorFlow to work on a Windows machine using Docker here using the tips from issue 42 if it's helpful for anyone.

Although I'm still not sure how I can take advantage of my GPU though if TensorFlow is running in a VM on my Windows machine.

+1 native windows support would be extremely helpful

+1

I want have a version of TensorFlow for Windows 7,if you have,Please send my email.
My email address: [email protected].
Thanks!

+1

Windows support is a bit of a dramatic effort. Maybe supporting Windows only for a truly standard compiler (gcc / tdm-gcc / mingwpy) and on a modern Python (3.4?) would be a much more easy target ?

Perhaps using Clang frontend with VC++ CodeGen would be a fastest pathway to get working Windows builds -> Clang with Microsoft CodeGen in VS 2015 Update 1. Project was designed to provide consistently working toolchain for cross platform builds of code primarily targeting windows, nonetheless scenario in which *nix code is compiled on windows could be equally interesting. MSFT C++ team states that code they write is contributed back to LLVM project so it should allow for fast and joint work on getting good cross platform support.

+1

+1

+1

+1

+1

+1

+1

+1 . This is the first Python package I recall not working on Windows. It seems odd to me, as my understanding is that NVIDIA's drivers have always been better on Windows than Linux, and I have a GTX 970 that's feeling bored these days.

I have ported most of the tensorflow c++ to a windows build use vc 2013. so far most code compile and linked use my own sources file. the biggest challenge for me to work around is that: it use static variable to register op and kernel. I build them as static library and app link to it will skip those static variable. I have to use a stub.cpp in app folder to include used ops and kernels.

@yuanhua8 any chance you push your changes to a github repo? BTW, I solved the static registration thing in caffe by using dumpin to generate a header that forces symbol liking. See https://github.com/BVLC/caffe/pull/2816 and https://github.com/willyd/caffe/tree/msvc for details.

Hope native Tensorflow for Windows will be ready soon, might have to wait until bazel for Windows is stable, thrilled to try Deep Learning course by Google and Udacity!

FYI Windows support for Bazel is currently one of our top priorities. Stay tuned.

Same here, i begun the deep learning course offered by Google and here i am stuck because i am running Windows.
When should we expect support for Windows ?

I noticed lot of people facing a problem running this under Windows so did a quick write up on how to set it under Windows using Vagrant and Docker:

https://medium.com/@Rapchik/running-google-s-deep-learning-course-material-under-windows-82d468b6d5be

@umarniz Thanks for the guide. However, I think, what most people here anticipate to see is TensorFlow running natively on Windows because of GPU support. That unfortunately doesn't work with Docker as far as I know nor any other solution using virtualization...

@SeveQ I completely agree and I personally prefer using Linux for all my ML experiments due to the ease of setting up multiple unique programming environments. This guide was meant to run the course material from Google as the training sets and examples they have are for beginners with smaller data sets which should be able to run in a CPU only environment too.

@umarniz Alright, didn't mean to denigrate your effort. Definitely not! You did a great job writing that up! Thanks again!

By the way, only to mention this... the Deep MNIST example, which is admittedly already not that simple anymore, takes several hours on my CPU whereas my GPU (GTX 980Ti) rushes through it in at most a few seconds. On a native Ubuntu that is. I've one set up on a USB stick. It works, however dual boot can be quite cumbersome and a huge demotivating factor. Even more so since a hibernated Windows on a UEFI system can be pretty picky when it comes to accessing data on NTFS partitions from Linux...

Sure, I totally agree that Linux (Ubuntu) is the best OS for ML tasks because it's the one most people use, and not without reason. But it still has its disadvantages for enthusiast people like me who not only do ML for a living but also as hobbyists, and who have other hobbies in parallel that rely on Windows... like gaming for example.

I guess that's some first world problems that I'd absolutely love to see solved nevertheless.

@SeveQ I didn't take it that way either :)

I agree with dual boot being a pain, I had to completely move to Ubuntu for a few months otherwise it becomes quite cumbersome to switch but I love the power that bash shell gives for working with large data sets.

I think i'll see what I can do to help add Windows support for TensorFlow myself.

+1

@umarniz Sounds good. I don't have the time right now to also dedicate myself to this, unfortunately. Other priorities like graduating... Windows support for TensorFlow would be a great help though in this matter. I'm graduating as MSc., systems engineer, specializing in, who would have thought, machine learning...

any update on this

I make a port of the c++ code into our internal branch. Sorry can not publish back since it take dependency of our code branch. Find some gap btw c++ and Python. But the core engine work for me in our windows/Vc environment.
H y

Sent from my iPhone

On Jan 30, 2016, at 2:46 AM, datashinobi [email protected] wrote:

any update on this

—
Reply to this email directly or view it on GitHub.

For all those who want to use Tensorflow and can surrender GPU support here is an Blogpost about installing Tensorflow with jupyter notebook support for Windows with Docker.

An update on Windows support for Bazel: the initial set of patches to get Bazel working on Windows has been merged (see bazelbuild/bazel#276), and @dslomov has been able to get Bazel to bootstrap itself on Windows.

We are planning to have experimental support for Windows in Bazel 0.3. If you would like to follow our progress, see the issues tagged "Windows" on the Bazel issue tracker.

C:\Users\Desktop>docker run -it b.gcr.io/tensorflow/tensorflow

Unable to find image 'b.gcr.io/tensorflow/tensorflow:latest' locally

docker: Error response from daemon: unable to ping registry endpoint https://b.g
cr.io/v0/
v2 ping attempt failed with error: Get https://b.gcr.io/v2/: dial tcp 64.233.188
.82:443: i/o timeout

Any news regarding being able to use TensorFlow in Windows, without Docker?
How hard would it be to offer a CMake build process that works both for Linux and Windows (creating a MVS project)?

I wrote instructions for installing tensorflow on a CentOS virtual machine. If you're having trouble working with Docker, this might be your solution.

Guys if you install bash on Windows using the new developer preview tensorflow does install and kinda work. So you can use that to run it on Windows without using docker. (And it would probably improve the bash for Windows because more people are using it.)

That's fantastic.
Using bash on Windows was my biggest hope to use tensorflow on Windows.

The biggest question is whether you can use the GPU drivers?

That's the main reason people asking for a native tensorflow installation in windows. Otherwise cpu support is just fine through docker or Linux virtual machine

I have no knowledge of that but we should ask the Windows team about that.
People have gotten GUI programs to work through windows but it uses CPU not
GPU so I'm thinking it does not currently have access
On Apr 19, 2016 5:37 PM, "hayder78" [email protected] wrote:

That's fantastic.
Using bash on Windows was my biggest hope to use tensorflow on Windows.

The biggest question is whether you can use the GPU drivers?

That's the main reason people asking for a native tensorflow installation
in windows. Otherwise cpu support is just fine through docker or Linux
virtual machine

—
You are receiving this because you commented.
Reply to this email directly or view it on GitHub
https://github.com/tensorflow/tensorflow/issues/17#issuecomment-212156142

Yeah. I read that GUI programs run through bash in windows will use a generic GPU driver. Which means does not use the nvidia drivers.

Let's pray that Google is working on a tensorflow on Windows under the hood.

I think bash for Windows needs time to be mature enough and bug free. In the meantime probably VM and docker will be a better choice.

I am now making a choice between tensorflow and CNTK.

+1 for Windows native support, including GPU.

The link to follow Bazel's Windows issues provided by @davidzchen isn't working, the correct one is https://github.com/bazelbuild/bazel/labels/category%3A%20windows

(I know everyone is hoping for Windows native support + GPU, I am waiting too, but we have to wait until Bazel for Windows gets stable, helping in testing and contributing to Bazel will probably speed up the process)

another +1 for windows with GPU support. Although TF seemed a better option, I had to switch to Theano as a result for lack of GPU support on windows. It will be nice if there is a mention on TF's roadmap as to whether support is going to be available in future releases, as it will help with loads of researchers decision making.

+1

+1

One interesting piece of news: @shanselman wrote a blog post about running TensorFlow on Bash for Windows.

We still intend to provide first-class Windows support, but adventurous users might find this a good way to get started in the mean time.

@mrry Hi Derek Murray,
I am glad to hear from you - as a Google software engineer - that you intend to provide a native windows support for tensorflow. Are you part of the tensorflow developer team at Google?

At least I have a hope now.

Any rough estimate when will be a beta release?

Another +1, very interested to see this happen. If and when Bazel actually runs on Windows, will TF actually compile on the platform? Or are there other portability issues as well e.g. network, file system calls?

+1

Folks, could you please use GitHub reactions rather than writing "+1" in a new message? Right now it produces a lot of spam for people who subscribed to updates in this thread. Thanks!

+1 for GitHub reactions!

Well bazel seems to work most of the time on windows now, aside from a few name errors (illegal characters or very long commands). There definitely are issues to getting a working windows tensorflow other than just having a working bazel. All the errors I have got so far are of the type:
ERROR: C:/tensorflow-orig/tensorflow/contrib/metrics/BUILD:16:1: in linkshared attribute of cc_binary rule //tensorflow/contrib/metrics:python/ops/_set_ops.so: 'linkshared' used in non-shared library. Since this rule was created by the macro 'tf_custom_op_library', the error might have been caused by the macro implementation in C:/tensorflow-orig/tensorflow/tensorflow.bzl:599:31.
So I am guessing build rules need to be updated for windows first.

The official roadmap of tensorflow :
https://github.com/tensorflow/tensorflow/blob/master/tensorflow/g3doc/resources/roadmap.md

Windows support is part of the near future roadmap that is targeted for the next few months!

@Sabrewarrior I tried to build the tensorflow on windows using bazel, but the configure file just give some errors. I wondering if you changed or added some parts on that file to be supported on windows

@Fhrozen https://github.com/Sabrewarrior/tensorflow/blob/test/tf_win_env.txt
These are my installed packages when I ran configure without any changes. I am running this on msys2 with python 2.7 on a Windows 10 machine.

@Sabrewarrior thanks for the answer. I checked it, was the spaces in the folder name my problem.
Now i got this error on the cuda version.

`Please specify which gcc nvcc should use as the host compiler. [Default is /mingw64/bin/gcc]:
Please specify the Cuda SDK version you want to use, e.g. 7.0. [Leave empty to use system default]: 7.5
Please specify the location where CUDA 7.5 toolkit is installed. Refer to README.md for more details. [Default is /usr/local/cuda]: C:/CUDA/v7.5
Please specify the Cudnn version you want to use. [Leave empty to use system default]: 4.0.7
Please specify the location where cuDNN 4.0.7 library is installed. Refer to README.md for more details. [Default is C:/CUDA/v7.5]: cuda/
Please specify a list of comma-separated Cuda compute capabilities you want to build with.
You can find the compute capability of your device at: https://developer.nvidia.com/cuda-gpus.
Please note that each additional compute capability significantly increases your build time and binary size.

Setting up Cuda include
Setting up Cuda
ln: fallo al crear el enlace simbólico '../../../third_party/gpus/cuda/extras/visual_studio_integration/MSBuildExtensions/CUDA': No such file or directory
ln: fallo al crear el enlace simbólico '../../../third_party/gpus/cuda/extras/visual_studio_integration/MSBuildExtensions/7.5.props': No such file or directory
ln: fallo al crear el enlace simbólico '../../../third_party/gpus/cuda/extras/visual_studio_integration/MSBuildExtensions/CUDA': No such file or directory
ln: fallo al crear el enlace simbólico '../../../third_party/gpus/cuda/extras/visual_studio_integration/MSBuildExtensions/7.5.targets': No such file or directory
ln: fallo al crear el enlace simbólico '../../../third_party/gpus/cuda/extras/visual_studio_integration/MSBuildExtensions/CUDA': No such file or directory
ln: fallo al crear el enlace simbólico '../../../third_party/gpus/cuda/extras/visual_studio_integration/MSBuildExtensions/7.5.xml': No such file or directory
xargs: bash: finish with status 255; aborting
`

@Fhrozen AFAIK you can't use gcc as a compiler for CUDA on Windows. Only Visual C++ Compiler is supported by CUDA on Windows (see here).

+1

AGAIN! PLEASE DON'T MAKE USELESS _+1_ COMMENTS! USE THE GITHUB REACTIONS INSTEAD!

+1

stop +1!!!

I repeat, stop +1 for $%{^¥ sake! Anyone who comments +1 after this is an idiot! :-P Use Github reactions, dammit!

Could someone give me some kind of moderator status? I'd like to clean this issue from useless +1 comments.

Hi team. I'm trying to understand what work needs to be done to be able to use TensorFlow on Windows (10, I guess) with GPU support via CUDA. From what I know so far, that rules out Docker because Docker cannot access the host's GPU, and means that we must use Visual Studio's cl.exe (which is what Bazel for Windows does use).

From that, is it just Bazel for Windows that needs to be progressed? Is that the only unsatisfied dependency here? Happy to be corrected on any point.

If I want to turn on the Bash on Windows, I need to check the Windows Subsystem for Linux (Beta) right? But I can't find this menu in the Windows Features dialog. I'm following the guide here and here.
Do I need to tick Hyper-V check box in the Windows Features dialog?
I've also turned on the Developer mode in the settings already.

It seems like I need to desert these Windows beta bash features and unwillingly use Docker?

@off99555 You're Windows 10 should have the latest preview release (Fast Ring). Check your build, if it is lower than ~14000, then you don't have this feature yet. (This update will be GA August 2nd)

Hello,

For Tensorflow Windows support, are we going to be able to build out Windows Binaries that are not msys2 binaries? This makes sense since it would be a possible to do actual deployment on most user's machines.

If this is not possible, can we have a way to just deploy a stripped down version of the prediction pass/forward pass of the library for this purpose? I can train on Linux just fine but for the actual prediction process, I will like to be able to deploy on a regular Windows box without msys2. Is this on the roadmap?

If you look at what CNTK is doing:
https://github.com/Microsoft/CNTK/wiki/Native-Evaluation-Interface

They have a much easier way to do evaluation/prediction using a stripped down DLL. Can we have something like this in Tensorflow as well for Windows ( and other platforms I can imagine will find it very useful).

Thanks!

I don't understand why you are added support for Mac OS and not Windows. You cannot use normal CUDA GPU on most Macs. It is useless platform for deep learning training

Can the team tell us please about estimates for native windows support?

@mrry is actively working on this and can provide more updates on this if you have specific questions.

@aselle Thanks for the useful update, much appreciated!

@mrry Hello, how are the specifics of the Windows implementation going to be rolled out?

1) Will we have (ideally) a VS2015 solution that we can build with our apps/tools? A cmake method that generates the VS solution works too.

2) Is this going to be a statically linked or dynamically linked library?

3) Will there be a lightweight forward pass/evaluation module/library that we can use for deployment with our tools/apps?

The current resource I could find regarding deploying tensorflow with a C++ standalone app is here:
https://medium.com/jim-fleming/loading-a-tensorflow-graph-with-the-c-api-4caaff88463f#.7ejb8h7zk

My main concern is what is stated here:
"The build is huge, coming in at 103MB, even for this simple example. Much of this is for TensorFlow, CUDA support and numerous dependencies we never use. This is especially true since the C++ API doesn’t support much functionality right now, as a large portion of the TensorFlow API is Python-only. There is probably a better way of linking to TensorFlow (e.g. shared library) but I haven’t gotten it working yet."

If there is a better way to get around linking a large static library that will be great, especially for the purposes of a forward pass/evaluation which does not really need GPU support.

Thanks!

@mrry How can others get involved in this? How can we help to get TensorFlow running on Windows as soon as possible? At least we can help with testing and bug reporting.

+1

Any updates on windows? whats needed to get it run natively having GPU available as well?

We need windows support!!!!
Can the team tell us please about estimates for native windows support?

No windows? why? at least CPU only versions?

Here is a summary of what I think is happening in this Windows support issue:

Tensorflow requires Bazel to build from source, If I am not mistaken, Bazel is a build system like GNU Make, but not a compiler.

Based on what I saw as I briefly scanned through the code, source code of Tensorflow itself uses mostly standard C++ library for things like threading so it should be no problem to compile on Windows, essential third-party libraries it uses all have Windows support.

Therefore, I think the main problem lies on lack of build method for Windows rather than source code itself. There are works on using CMake to build instead of Bazel but not complete yet. If someone can translate Bazel build rules to CMake's, I think we will be able to build it on Windows.

• https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/makefile
• https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/cmake

Stable Bazel support for Windows still have a long way to go.

Please correct me if I made a mistake here.

@eiva If you want CPU-only support you can run Tensorflow in a virtual machine. I've run it on Ubuntu 14 on a Windows 10 host without issue.

@rongjiecomputer I honestly wonder why they used Bazel at all...

@eiva There's also a ready built Docker container that you can use on a Windows host. Works out of the box. It's just not that much fun to work with tensorflow without GPU support. For the cool stuff at least a single GPU is a requirement. You can of course run basic stuff like linear or logistic regression without a GPU. But that's just not tensorflow's long suit. It's made for deep learning tasks which basically require a GPU.

@marklit i use it this way already, but most of my app running on windows, and most of development is going on windows... Its not easy to switch os each time i doing something with different parts of system...

@Loo Rong
Your speculation is right!
https://www.quora.com/Why-did-Google-decide-to-use-Bazel-with-TensorFlow/answer/Derek-Murray-3

On Sep 12, 2016 12:17 PM, "Eugene Ivanchenko" [email protected]
wrote:

@marklit https://github.com/marklit i use it this way already, but most
of my app running on windows, and most of development is going on
windows... Its not easy to switch os each time i doing something with
different parts of system...

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
https://github.com/tensorflow/tensorflow/issues/17#issuecomment-246239719,
or mute the thread
https://github.com/notifications/unsubscribe-auth/AEBIBhhA-t1MrFnddm5k-wER0KAm_p30ks5qpNJTgaJpZM4GexU4
.

Thanks for all of the interest in TensorFlow on Windows! We're making progress on two main fronts:

1. Adapting TensorFlow's Bazel BUILD files to work on Windows. You may have seen some recent pull requests from @meteorcloudy in that direction (like #4449), and our eventual plan is to have full support for building TensorFlow, using Bazel, on Windows.
2. Modifying the TensorFlow runtime to build with the Visual C++ 2015 compiler. I've been working on this effort, and the diff between HEAD and my working branch is getting smaller every day. For now, I'm using CMake to build TensorFlow using Visual Studio/MSBuild, but I plan to switch to Bazel when the BUILD files work cross-platform.

We expect to make an announcement soon, with a binary PIP package and instructions for building TensorFlow on Windows. In the meantime, we will update this issue when we have news to share.

Responding to the specific questions from @kestrelm:

1. I am currently using CMake to generate a VS solution with multiple projects, and we'll merge this as part of item (2) above.
2. At the very least, we can generate a Python extension DLL containing the runtime and all kernels (the equivalent of tensorflow/python/_pywrap_tensorflow.so), and provide instructions for building a statically linked C++ binary (the equivalent of tensorflow/cc/tutorials/example_trainer.cc).
3. We can easily make a version of the runtime with those features (something like the Android subset of inference-related kernels in a DLL behind the C API), although it probably won't be part of the initial release. It will however be possible to modify the build files to make a target that has these properties, and I'll be happy to help anyone who's trying to do that.

Awesome work, @mrry!

We expect to make an announcement soon, with a binary PIP package and instructions for building TensorFlow on Windows

Can you give any rough estimates on when that announcement will be? I can't wait!

@mrry : Derek, solid update. Do you have a binary GPU release (BETA) before the rest of the makefile/compiler framework solidifies? At least folks could try out against the Python API ...

Dear Gurus:

I tried to build tensorflow from source code on Windows 10. It reports bazel is already supported on Windows although might not be stable. Nevertheless I got a working one.
However, when I tried to build tensorflow, there're two problems, first it indicates "bazel clean --expunge_async" failed with permission problem. I bet it's not that severe. The next problem is a little vital since it blocks the build. it reports:

...
INFO: Found 1 target...
INFO: Writing explanation of rebuilds to 'logfile'
ERROR: missing input file '@local_config_cuda//cuda:lib64/libcublas.so'.
ERROR: C:/tensorflow/tensorflow/tools/pip_package/BUILD:23:1: //tensorflow/tools/pip_package:build_pip_package: missing input file '@local_config_cuda//cuda:lib64/libcublas.so'.
Target //tensorflow/tools/pip_package:build_pip_package failed to build
ERROR: C:/tensorflow/tensorflow/tools/pip_package/BUILD:23:1 1 input file(s) do not exist.
INFO: Elapsed time: 34.644s, Critical Path: 1.52s

Actually during the configuration process, I deliberately disable GPU support, so why it tries to load dynamic library related to CUDA?

Since I disable GPU support, the stub libraries are correctly created with zero size and with the DLL file extension. However, as in the error message, I don't know where the requirement of loading CUDA related libraries originated. I checked several files but still have no idea. Meantime, I also think the file is poorly written, since it should look for DLL files instead SO files. However I still have no clue.

Anyone has any idea to this problem? Thanks a lot.

As you may have seen with PR #4778 merged, we now have preliminary support for building TensorFlow on Windows using CMake. It supports build a CPU-only version of TensorFlow for use in the C++ example trainer program, and a PIP package for Python use. For more details and instructions on how to build TensorFlow from source on Windows, see the CMake readme.

This is alpha-quality code, and we expect there to be bugs in this initial version. If you find one, please raise a new GitHub issue for the specific problem.

Our focus now shifts to getting GPU support for TensorFlow on Windows, and transitioning to a common Bazel-based build for Windows and Linux. If you have other feature requests, please raise a new GitHub issue.

I was using this prebuilt project https://github.com/shishaochen/TensorFlow-0.8-Win, but it does not support GPU.

@mrry Hi, Derek, appreciate your contribution to TensorFlow.
@cesardelgadof Hi, Cesar, thanks for your information.

However, I think the previously obstacle hindering build TensorFlow on Windows platform is the unavailability of bazel on Windows. Since now it's available, though may not be stable, so in my opinion we'd better try to solve the problem ahead when building TensorFlow using bazel.

I personally think guys at Google, if they do have the plan to support native build on Windows, they probably will choose bazel. And once the official solution available, I guess people will switch from cmake to bazel. So if there's such a day, why we work together to have the day come early?

Nevertheless I will try the cmake version, and hopefully someone else will also try the bazel build system. Probably we can figure out what's the nasty error comes from, and locate the poorly written configuration file.

Thanks for you all.

@mrry Thanks for the major contribution! I've tried to follow your instructions. All seem to work flawlessly (for about 45min) until it failed with:

The target "BeforeGenerateProjectPriFile" listed in a BeforeTargets attribute at "C:\Program Files (x86)\MSBuild\Microsoft\NuGet\Microsoft.NuGet.targets (186,61)" does not exist in the project, and will be ignored.
Done Building Project "H:\PycharmProjects\tensorflow\tensorflow\contrib\cmake\build\tf_python_build_pip_package.vcxproj" (default targets) -- FAILED.

"H:\PycharmProjects\tensorflow\tensorflow\contrib\cmake\build\tf_python_build_pip_package.vcxproj" (default target) (1) ->
"H:\PycharmProjects\tensorflow\tensorflow\contrib\cmake\build\pywrap_tensorflow.vcxproj" (default target) (3) ->
"H:\PycharmProjects\tensorflow\tensorflow\contrib\cmake\build\tf_core_cpu.vcxproj" (default target) (4) ->
"H:\PycharmProjects\tensorflow\tensorflow\contrib\cmake\build\tf_core_framework.vcxproj" (default target) (5) ->
(CustomBuild target) ->
C:\Program Files (x86)\MSBuild\Microsoft.Cpp\v4.0\V140\Microsoft.CppCommon.targets(171,5): error MSB6006: "cmd.exe" exited with code 1. [H:\PycharmProjects\t
ensorflow\tensorflow\contrib\cmake\build\tf_core_framework.vcxproj]

30 Warning(s)
1 Error(s)

@laudney Thanks for giving it a try, and I'm sorry to see that you're getting an error! Since this issue has a lot of subscribers, I've created issue #4798 to track the fix for this problem, so let's continue the discussion there.

@mingyr We hope to have TensorFlow building on Windows with Bazel as soon as possible. @meteorcloudy has been doing sterling work to get this to happen, and he just sent PR #4796, which enables Bazel to build //tensorflow/cc:tutorials_example_trainer on Windows. We'll aim to get that merged soon.

@mrry Thanks for #4798 , it now builds ok, also tested python whl install and no problems found so far.

Hope to see tensorflow in conda soon.

Use Tensorflow on virtualized environments as Docker/VirtualBox is a very bad solution. Tensorflow should run on Windows natively. (And since the most is coded in Python should not be a problem create a distribution for Windows.)

Hi everyone! As mentioned by @mrry, I am working on TensorFlow Windows build with Bazel. With PR #4796 and PR #4874 merged, the C++ example trainer now builds on Windows with Bazel 0.3.2!

Here is a brief instruction about how to build TF with Bazel on Windows:

• Install Bazel on Windows(installation instructions)
• Setup the environment Bazel needs for building C++ with MSVC
• Clone TF repository, run ./configure in MSYS
• Run bazel build -c opt --cpu=x64_windows_msvc --host_cpu=x64_windows_msvc //tensorflow/cc:tutorials_example_trainer --verbose_failures

For convenience, you can put the build options in ~/.bazelrc, mine is like:

build -c opt
build --cpu=x64_windows_msvc --host_cpu=x64_windows_msvc
build --copt="/w"                  # Suppress some warning messages
build --experimental_ui            # Enable a nice UI 

A known issue is that Bazel's MSVC wrapper script doesn't support python3 in Bazel 0.3.2, but it's already been fixed at https://github.com/bazelbuild/bazel/commit/ce5c33dd7e96aff0cf1eb993edd41a5fe8c423f2. You can solve it by setting BAZEL_PYTHON to a python 2.7 binary or building Bazel from HEAD.

Building the C++ example trainer with Bazel takes about 20 minutes on my machine(Windows 7, CPU 2.9GHz, RAM 64G), an incremental build after changing some source file (i.e. array_ops.cc) is around 1 minute.

I am also making some progress on building the python PIP package on Windows with Bazel, I will send a PR as soon as possible.

Please try to build TensorFlow with Bazel on Windows, and tell us what we could improve!

@meteorcloudy, Hi, Yun:
The C++ example trainer must built by the prebuilt bazel binary installed on Windows, or could I sync to the latest git version of bazel, and use the custom built bazel?

Actually I choose the second way, use the customer build bazel, however, it complains as follows:

ERROR: C:/tensorflow/tensorflow/core/BUILD:108:1: output 'tensorflow/core/exampl e/example.pb.h' was not created.
ERROR: C:/tensorflow/tensorflow/core/BUILD:108:1: output 'tensorflow/core/exampl e/example_parser_configuration.pb.h' was not created.
ERROR: C:/tensorflow/tensorflow/core/BUILD:108:1: output 'tensorflow/core/exampl e/example.pb.h' was not created.
ERROR: C:/tensorflow/tensorflow/core/BUILD:108:1: output 'tensorflow/core/exampl e/feature.pb.h' was not created.
......

Procedures for build I took:
$ cd c:/tensorflow
$ export JAVA_HOME="$(ls -d C:/Program\ Files/Java/jdk* | sort | tail -n 1)"
$ export BAZEL_SH=c:/tools/msys64/usr/bin/bash.exe
$ export BAZEL_VS="C:/Program Files (x86)/Microsoft Visual Studio 14.0"
$ export BAZEL_PYTHON=C:/Python27/python.exe
$ export PATH=$PATH:/c/Python27:/c/Python27/Scripts:/c/tools/swigwin-3.0.10:/c/bazel/output

./configure

bazel build -c opt --cpu=x64_windows_msvc --host_cpu=x64_windows_msvc //tensorflow/cc:tutorials_example_trainer --verbose_failures

Any suggestion?

Thanks.

@meteorcloudy Hi Yun,

I have followed similar procedure but I get the following error:
cd C:/tools/msys64/var/tmp/Bazel/V1uFCi$H/execroot/tensorflow
bazel-out/host/bin/external/protobuf/protoc.exe --cpp_out=bazel-out/vc_14_0_x64-opt/genfiles/ --plugin=protoc-gen-grpc=bazel-out/host/bin/external/grpc/grpc_cpp_plugin.exe --grpc_out=bazel-out/vc_14_0_x64-opt/genfiles/ -I. -Iexternal/protobuf/src -Ibazel-out/vc_14_0_x64-opt/genfiles/external/protobuf/src tensorflow/core/debug/debug_service.proto: com.google.devtools.build.lib.shell.BadExitStatusException: Process exited with status 1.
azel-out/vc_14_0_x64-opt/genfiles/external/protobuf/src: warning: directory does not exist.

I checked the folders and there is no src directory in protobuf. Also there is no grpc folder in bazel_out/host/bin/external/ (so no grpc_cpp_plugin.exe file).

this error is followed by a bunch of errors that files in tensorflow/core are not created

Any idea how to fix it?

Thanks

@mingyr @karthiek Thank you for trying to build TF with Bazel!
Unfortunately, according to http://ci.bazel.io/job/TensorFlow/ , TF build is even failing on Linux with Bazel at HEAD with similar error given by you. We are fixing it, meanwhile, can you try again with Bazel 0.3.2?

@meteorcloudy Hi Yun:

What do you mean by "can you try again with Bazel 0.3.2"?

You mean the official prebuilt Windows binary, or we can sync to the latest version of bazel, and use my build?

Actually I already sync to the latest version of bazel and use my build. If you indicate I should try the official prebuilt binary to see whether the problem is still pending, please kindly let me know.

B. R.

@mingyr Yes, I mean the "official prebuilt binary". I am having the same error after sync to HEAD.
Please see https://github.com/bazelbuild/bazel/issues/1929

@meteorcloudy Hi, Yun

Thanks for the suggestion. I tried using 0.3.2 pre-compiled binary and was able to compile and generate tutorilas_example_trainer.exe file.

@karthiek So glad you succeed! How long did it take? Is it faster than the CMake build?

@meteorcloudy It took around 15 minutes (Intel(R) Core(TM) i7-6700 CPU @ 3.40GHz). Its just yesterday I had the necessity to build tensor flow on windows. So, I directly tried the last method you suggested. I didn't try building it using CMake.

Hi, I tried to follow @meteorcloudy approach but ended up with errors as below. Could anyone help?

bazel build -c opt --cpu=x64_windows_msvc --host_cpu=x64_windows_msvc //tensorflow/cc:tutorials_example_trainer --verbose_failures
INFO: Analysed target //tensorflow/cc:tutorials_example_trainer.
INFO: Found 1 target...
ERROR: C:/Programming/tensorflow/tensorflow/core/BUILD:1064:1: Executing genrule //tensorflow/core:version_info_gen failed: bash.exe failed: error executing command
cd C:/tools/msys64/var/tmp/Bazel/s5nfxnhX/execroot/tensorflow
SET PATH=C:\tools\msys64\usr\local\bin;C:\tools\msys64\usr\bin;C:\tools\msys64\usr\bin;C:\tools\msys64\opt\bin;C:\Windows\System32;C:\Windows;C:\Windows\System32\Wbem;C:\Windows\System32\WindowsPowerShell\v1.0\;C:\tools\msys64\usr\bin\site_perl;C:\tools\msys64\usr\bin\vendor_perl;C:\tools\msys64\usr\bin\core_perl;C:\Program Files\Google;C:\WinPython-64bit-2.7.10.2\python-2.7.10.amd64;C:\WinPython-64bit-2.7.10.2\python-2.7.10.amd64\Scripts;C:\tools\swigwin-3.0.10;C:\Program Files\java\jdk1.8.0_102\bin
C:/tools/msys64/usr/bin/bash.exe -c source external/bazel_tools/tools/genrule/genrule-setup.sh; tensorflow/tools/git/gen_git_source.py --generate tensorflow/tools/git/gen/spec.json tensorflow/tools/git/gen/head tensorflow/tools/git/gen/branch_ref "bazel-out/vc_14_0_x64-opt/genfiles/tensorflow/core/util/version_info.cc": com.google.devtools.build.lib.shell.BadExitStatusException: Process exited with status 1
Traceback (most recent call last):
File "tensorflow/tools/git/gen_git_source.py", line 258, in
generate(args.generate)
File "tensorflow/tools/git/gen_git_source.py", line 210, in generate
git_version = get_git_version(data["path"])
File "tensorflow/tools/git/gen_git_source.py", line 151, in get_git_version
"--long", "--dirty", "--tags"]).strip()
File "C:\WinPython-64bit-2.7.10.2\python-2.7.10.amd64\lib\subprocess.py", line 566, in check_output
process = Popen(stdout=PIPE, _popenargs, *_kwargs)
File "C:\WinPython-64bit-2.7.10.2\python-2.7.10.amd64\lib\subprocess.py", line 710, in init
errread, errwrite)
File "C:\WinPython-64bit-2.7.10.2\python-2.7.10.amd64\lib\subprocess.py", line 958, in _execute_child
startupinfo)
WindowsError: [Error 2] The system cannot find the file specified
ERROR: C:/tools/msys64/var/tmp/Bazel/s5nfxnhX/external/jpeg_archive/BUILD:74:1: declared output 'external/jpeg_archive/jconfig.h' was not created by genrule. This is probably because the genrule actually didn't create this output, or because the output was a directory and the genrule was run remotely (note that only the contents of declared file outputs are copied from genrules run remotely)
ERROR: C:/tools/msys64/var/tmp/Bazel/s5nfxnhX/external/gif_archive/BUILD:35:1: declared output 'external/gif_archive/windows/unistd.h' was not created by genrule. This is probably because the genrule actually didn't create this output, or because the output was a directory and the genrule was run remotely (note that only the contents of declared file outputs are copied from genrules run remotely)
Target //tensorflow/cc:tutorials_example_trainer failed to build
INFO: Elapsed time: 4.967s, Critical Path: 0.96s
FAILED: Build did NOT complete successfully

@bssrdf your symptoms seem similar to https://github.com/bazelbuild/bazel/issues/1463.
It is a known issue of bazel on Windows that we break if your workspace is not on C: drive.

If that's the case, the workaround is to create a junction - if your tensorflow checkout is in d:\src\tensorflow the following should work

C:\> mklink /j d_src d:\src
C:\> cd c:\d_src\tensorflow
C:\D_SRC\TENSORFLOW> bazel build ...

See https://github.com/bazelbuild/bazel/issues/1463#issuecomment-251967323
Sorry for the inconvenience, we are fixing this asap.

@bssrdf hmm on a second look maybe it is not it - could it be that you do not have git in PATH?

@meteorcloudy Hi, Yun:

I also confirm that there's no problem compiling the C++ example trainer using pre-compiled bazel binary version 0.3.2.

My notebook is a old one and meantime it runs other tasks, so the compiling time is nonsense from bench marking perspective.

Thanks for your work and look farword to compiling the python PIP package on Windows

B. R.

@mingyr That's great!
But just a notification for everyone, the example trainer is not buildable at TF HEAD, because my fix was reverted at be3bc472a52571a83f048479d6a4fa528b5a495e for some reason.
The good news is, building python PIP package with Bazel is close! Please see PR #4942

@dslomov thanks for the reply. I did have to solve the bazel non-c: drive workspace problem as you suggested. I installed git for windows and git can be seen in the PATH by MSYS2.

I just relaunched the bazel build command and it's working now, like a miracle. I don't know what happened yesterday. Anyway, this is just another confirmation of success on Windows 10 home, VS 2015 community edition, prebuilt bazel 0.3.2 and Winpython-2.7.10.2

Thanks a lot for Yun's work.

@meteorcloudy, is it possible to build the GPU version following the procedure (with changes in bazel build of course)?

@mrry Hi Derek, thanks for your contribution, but I have this problem with tensorboard, it does not working:

C:\Users\Cesar\Documents\Notebooks\DeepLearning>tensorboard --logdir='./my_graph/'
Traceback (most recent call last):
  File "c:\program files\anaconda3\lib\runpy.py", line 184, in _run_module_as_main
    "__main__", mod_spec)
  File "c:\program files\anaconda3\lib\runpy.py", line 85, in _run_code
    exec(code, run_globals)
  File "C:\Program Files\Anaconda3\Scripts\tensorboard.exe\__main__.py", line 5, in <module>
ImportError: No module named 'tensorflow.tensorboard'

@bssrdf Probably not, because the code doesn't support GPU on Windows yet. @mrry did all the work of making TF code compile with MSVC, and he has been focusing on fixing GPU support.

Do we have some sort of educated guess on ETA for production use?

Since it seems relevant to people here using Bazel on Windows; i created a chocolatey package to make installation easier. https://chocolatey.org/packages/bazel. Please try it out and if there are any problems (it _is_ new), tweet or raise an issue over at bazelbuild/bazel.

GPU support is out thanks to @mrry and @guschmue.
Instructions will be updated on readme.

Has anyone been able to compile on Windows and create a PIP package? Thank you. :)

master (66f979714c9a8582059f383ec1505d13c9e8c523) should build ok for windows cpu and gpu. Instructions are in tensorflow/contrib/cmake/README.md (use cmake 3.6 for now).

Tensorboard is not included?, because it not works:

C:\Users\Cesar\Documents\Notebooks\DeepLearning>tensorboard --logdir='./my_graph/'
Traceback (most recent call last):
  File "c:\program files\anaconda3\lib\runpy.py", line 184, in _run_module_as_main
    "__main__", mod_spec)
  File "c:\program files\anaconda3\lib\runpy.py", line 85, in _run_code
    exec(code, run_globals)
  File "C:\Program Files\Anaconda3\Scripts\tensorboard.exe\__main__.py", line 5, in <module>
ImportError: No module named 'tensorflow.tensorboard'

Some python modules are missing from the pip: examples, some of contrib and tensorboard, doesn't get copied into the wheel when building with cmake.
Let me check what it is just a matter of adding it with an option to tf_python.cmake or if there are complications.

I compiled tensorflow for python as per the instructions, but I get a crash originating from cudnn when using tf.nn.conv2d on my GPU.

Sample code that causes the crash is

import tensorflow as tf
import numpy as np
sess = tf.InteractiveSession()
image = tf.placeholder(tf.float32, shape=[None, 10, 10, 1], name='image')
kernel = tf.truncated_normal([5, 5, 1, 1], stddev=0.1)
conv = tf.nn.conv2d(image, kernel, strides=[1, 1, 1, 1], padding='SAME')
sess.run(tf.initialize_all_variables())
sess.run(conv, feed_dict={image: [np.zeros((10, 10, 1))]})

This causes the python process to crash and gives this output in the jupyter console

WARNING:root:kernel 800a0f42-d62b-4128-a61a-ba64ee725ca8 restarted
I d:\tensorflow\tensorflow\tensorflow\stream_executor\dso_loader.cc:125] successfully opened CUDA library cublas64_80.dll locally
I d:\tensorflow\tensorflow\tensorflow\stream_executor\dso_loader.cc:125] successfully opened CUDA library cudnn64_5.dll locally
I d:\tensorflow\tensorflow\tensorflow\stream_executor\dso_loader.cc:125] successfully opened CUDA library cufft64_80.dll locally
I d:\tensorflow\tensorflow\tensorflow\stream_executor\dso_loader.cc:125] successfully opened CUDA library nvcuda.dll locally
I d:\tensorflow\tensorflow\tensorflow\stream_executor\dso_loader.cc:125] successfully opened CUDA library curand64_80.dll locally
I d:\tensorflow\tensorflow\tensorflow\core\common_runtime\gpu\gpu_device.cc:944] Found device 0 with properties:
name: GeForce GTX 970
major: 5 minor: 2 memoryClockRate (GHz) 1.3165
pciBusID 0000:01:00.0
Total memory: 4.00GiB
Free memory: 3.31GiB
I d:\tensorflow\tensorflow\tensorflow\core\common_runtime\gpu\gpu_device.cc:965] DMA: 0
I d:\tensorflow\tensorflow\tensorflow\core\common_runtime\gpu\gpu_device.cc:975] 0:   Y
I d:\tensorflow\tensorflow\tensorflow\core\common_runtime\gpu\gpu_device.cc:1034] Creating TensorFlow device (/gpu:0) -> (device: 0, name: GeForce GTX 970, pci bus id: 0000:01:00.0)
E d:\tensorflow\tensorflow\tensorflow\core\common_runtime\gpu\gpu_device.cc:662] Could not identify NUMA node of /job:localhost/replica:0/task:0/gpu:0, defaulting to 0.  Your kernel may not have been built with NUMA support.
E d:\tensorflow\tensorflow\tensorflow\stream_executor\cuda\cuda_dnn.cc:385] could not create cudnn handle: CUDNN_STATUS_NOT_INITIALIZED
E d:\tensorflow\tensorflow\tensorflow\stream_executor\cuda\cuda_dnn.cc:392] error retrieving driver version: Permission denied: could not open driver version path for reading: /proc/driver/nvidia/version
E d:\tensorflow\tensorflow\tensorflow\stream_executor\cuda\cuda_dnn.cc:352] could not destroy cudnn handle: CUDNN_STATUS_BAD_PARAM
F d:\tensorflow\tensorflow\tensorflow\core\kernels\conv_ops.cc:532] Check failed: stream->parent()->GetConvolveAlgorithms(&algorithms)

While messing around with the trying different versions of cudnn64_5.dll I was able to get the code to work once and confirmed that the GPU was being used, but even after trying all the versions of cudnn that I downloaded I have had no luck making it work a second time.

tried your script and it works for me. We tested only with cudnn5.1, everything older I don't know.
I'd get a clean install of cudnn5.1, build with that and make sure the cudnn64_5.dll in PATH is 5.1.
Another thing to check: your free memory is a little lower than I'd expect - is there maybe another python running that has the device open?

I tried a clean rebuild with a fresh cudnn downloaded and on the PATH (cudnn-8.0-windows10-x64-v5.1) but no cigar. cudnn64_5.dll from this download is 'File version': 6.14.11.8000 and has 'File description': NVIDIA CUDA 80.0.29 CUDNN Library .

cublas64_80.dll from the CUDA 8.0 path has the same 'File version' but 'File description': NVIDIA CUDA 80.0.45 BLAS Library

cufft64_80.dll and curand64_80.dll also have matching 'File versions' but are 8.0.44 in 'File description'

The GPU drivers I have installed are version 369.30 from latest CUDA drivers and nvcuda.dll in System32 has 'File version': 6.14.13.6930 and has 'File description': NVIDIA CUDA 80.0.44 driver .

I believe the low device memory is just because I have a 970 which only has 3.5GB of useful memory and CUDA doesn't like using the last 0.5GB.

I have the same cudnn version and its works fine for me. There are a bunch of conv python tests in the tf tree and they are all passing. I mostly use a 970 for testing. Your log was specific that it had issues initializing cudnn. Let me take a look at the init code.

OK the issue appears to be a driver one. I can successfully use cudnn and convolutions once each time I restart the video driver (using device manager disable then enable). Once I have used it in one process however, any future uses fail to initialise.

Furthermore I think other processes are having driver-related crashes once cudnn has been initialised.

Update: I've uninstalled and reinstalled my graphics driver and now I can have one process (jupyter kernel) with cudnn work correctly, but until I close that kernel every attempt to use cudnn in other kernels will encounter the same crash. If I restart the working kernel I can run the code in a different kernel without having to reload the driver.

It compiled for me, but I had to add shell=True in the line 150 of the file gen_git_source.py. Otherwise, I would get the same error as showed in:
http://stackoverflow.com/questions/24306205/file-not-found-error-when-launching-a-subprocess

@synap5e if you run this in a normal tf.Session() a second instance will not work because tf manages all memory on the gpu. You can change the config for Session() like this:
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.4
with tf.Session(config=config) as sess:
...
with that I can run 2 scripts in a loop that hit cudnn. This is the same on linux.
But it crashes is different on windows - on linux it gets CUDA_ERROR_OUT_OF_MEMORY, on windows it get CUDNN_STATUS_NOT_INITIALIZED. Going to take a look at this.

Ah so the problem was that my system was using too much video memory even before initialising tensorflow. I guess the driver stuff was just bringing the used device memory down so that tensorflow could initialise properly. Thanks!

Guys, is it possible to build on Windows without MSVS? Windows 10 SDK is supposed to contain the C++ compiler, wouldn't that suffice for the task?

@zandaqo You only need standalone MSVC compiler with Windows 10 SDK and CMake, so there is no need to install the whole Visual Studio. Windows 10 SDK contains dll, static libraries, header files and source codes for the libraries and some useful tools, but it does not contain compiler.

Hello!

I'm new to both Python and Tensorflow, but I have done some MNIST stuff in the past. I followed the instructions on tensorflow/contrib/cmake/README.md, and it worked fine, GPU seems to be working fine as well. Tensorflow on Windows seems to be breaking in some places, as expected, and I just want to check whether my understanding of the breakage is right.

I tried to follow https://www.tensorflow.org/versions/r0.11/tutorials/mnist/beginners/index.html and it doesn't work, since this import path (tensorflow.examples.tutorials.mnist) doesnt exist. This is ultimately due to tf.contrib not being included in the pip package, right?

Just out of curiosity, I did something apparently pretty awful: I copied the contents of the tensorflow repo over the lib location at site-packages\tensorflow, which I expected to cause a lot of stuff to break. This is how it breaks:

  File "C:\Program Files\Anaconda3\lib\site-packages\tensorflow\contrib\framework\python\ops\variables.py", line 23, in <module>
    from tensorflow.contrib.framework.python.ops import gen_variable_ops
ImportError: cannot import name 'gen_variable_ops'

I couldn't find the definition of gen_variable_ops anywhere, but the documentation says that tf.load_op_library() is not currently implemented. So, gen_variable_ops is implemented somewhere else, which needs to be loaded using tf.load_op_library(). So, the lack of tf.load_op_library is what is causing the problem, right?

I've managed to build successfully the pip package following the cmake path, and installed it with pip install tensorflow-0.11.0rc1_cmake_experimental-py3-none-any.whl which reported a successful installation. However, when I try to import tensorflow, I get the following errors:

>>> import tensorflow as tf
Traceback (most recent call last):
  File "C:\Users\Tester\AppData\Local\Programs\Python\Python35\lib\site-packages\tensorflow\python\pywrap_tensorflow.py", line 18, in swig_import_helper
    return importlib.import_module(mname)
  File "C:\Users\Tester\AppData\Local\Programs\Python\Python35\lib\importlib\__init__.py", line 126, in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
  File "<frozen importlib._bootstrap>", line 986, in _gcd_import
  File "<frozen importlib._bootstrap>", line 969, in _find_and_load
  File "<frozen importlib._bootstrap>", line 958, in _find_and_load_unlocked
  File "<frozen importlib._bootstrap>", line 666, in _load_unlocked
  File "<frozen importlib._bootstrap>", line 577, in module_from_spec
  File "<frozen importlib._bootstrap_external>", line 906, in create_module
  File "<frozen importlib._bootstrap>", line 222, in _call_with_frames_removed
ImportError: DLL load failed: The specified module could not be found.

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "C:\Users\Tester\AppData\Local\Programs\Python\Python35\lib\site-packages\tensorflow\python\__init__.py", line 54, in <module>
    from tensorflow.python import pywrap_tensorflow
  File "C:\Users\Tester\AppData\Local\Programs\Python\Python35\lib\site-packages\tensorflow\python\pywrap_tensorflow.py", line 21, in <module>
    _pywrap_tensorflow = swig_import_helper()
  File "C:\Users\Tester\AppData\Local\Programs\Python\Python35\lib\site-packages\tensorflow\python\pywrap_tensorflow.py", line 20, in swig_import_helper
    return importlib.import_module('_pywrap_tensorflow')
  File "C:\Users\Tester\AppData\Local\Programs\Python\Python35\lib\importlib\__init__.py", line 126, in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
ImportError: No module named '_pywrap_tensorflow'

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "C:\Users\Tester\AppData\Local\Programs\Python\Python35\lib\site-packages\tensorflow\__init__.py", line 23, in <module>
    from tensorflow.python import *
  File "C:\Users\Tester\AppData\Local\Programs\Python\Python35\lib\site-packages\tensorflow\python\__init__.py", line 60, in <module>
    raise ImportError(msg)
ImportError: Traceback (most recent call last):
  File "C:\Users\Tester\AppData\Local\Programs\Python\Python35\lib\site-packages\tensorflow\python\pywrap_tensorflow.py", line 18, in swig_import_helper
    return importlib.import_module(mname)
  File "C:\Users\Tester\AppData\Local\Programs\Python\Python35\lib\importlib\__init__.py", line 126, in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
  File "<frozen importlib._bootstrap>", line 986, in _gcd_import
  File "<frozen importlib._bootstrap>", line 969, in _find_and_load
  File "<frozen importlib._bootstrap>", line 958, in _find_and_load_unlocked
  File "<frozen importlib._bootstrap>", line 666, in _load_unlocked
  File "<frozen importlib._bootstrap>", line 577, in module_from_spec
  File "<frozen importlib._bootstrap_external>", line 906, in create_module
  File "<frozen importlib._bootstrap>", line 222, in _call_with_frames_removed
ImportError: DLL load failed: The specified module could not be found.

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "C:\Users\Tester\AppData\Local\Programs\Python\Python35\lib\site-packages\tensorflow\python\__init__.py", line 54, in <module>
    from tensorflow.python import pywrap_tensorflow
  File "C:\Users\Tester\AppData\Local\Programs\Python\Python35\lib\site-packages\tensorflow\python\pywrap_tensorflow.py", line 21, in <module>
    _pywrap_tensorflow = swig_import_helper()
  File "C:\Users\Tester\AppData\Local\Programs\Python\Python35\lib\site-packages\tensorflow\python\pywrap_tensorflow.py", line 20, in swig_import_helper
    return importlib.import_module('_pywrap_tensorflow')
  File "C:\Users\Tester\AppData\Local\Programs\Python\Python35\lib\importlib\__init__.py", line 126, in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
ImportError: No module named '_pywrap_tensorflow'


Error importing tensorflow.  Unless you are using bazel,
you should not try to import tensorflow from its source directory;
please exit the tensorflow source tree, and relaunch your python interpreter
from there.

Did I miss some dependency here or something was wrong with the build/installation?

@zandaqo, I got this problem when I first run tensorflow on Windows. You just need to compile the examples, and add the folder with the compiled examples to your path. When you compile the sample codes, it will also compile some DLLs needed for Tensorflow.

@ErivaldoJunior By compiling the examples do you mean this: MSBuild /p:Configuration=Release tf_tutorials_example_trainer.vcxproj? I did that before compiling the pip package. It compiles to .\Release, is that the folder to add to path? I have just tried adding it, but no luck so far.

I do see files named pywrap_tensor mentioned in the errors in that .\Release folder, maybe manually copying them somewhere might help Python to see them. Not sure where though.

@zandaqo, that is right. I just add that Release folder to my Windows path and it worked. However, I tested by launching python my_tensorflow_script.py from a CMD window. I don't know if it work if you use some kind of IDE or iPython.

@ErivaldoJunior Thanks, you were right about missing DLLs, though in my case it turned out to be zlib.dll that was missing from both the package and .\Release folder. I've manually copied zlib.dll into the folder, as it was suggested here and the import now works.

I successed to build the tensorflow! But cmake 3.3.1 causes build error, 3.6.3 works fine.

Good to know! We'd welcome a PR with an appropriate minimum version
statement in the cmakelist file.

On Saturday, November 5, 2016, Takahiro Kubo [email protected]
wrote:

I successed to build the tensorflow! But cmake 3.3.1 causes build error,
3.6.3 works fine.

—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
https://github.com/tensorflow/tensorflow/issues/17#issuecomment-258649999,
or mute the thread
https://github.com/notifications/unsubscribe-auth/AAjO_Ys4MI03pcc3tCR_BzsqIXFrQTTcks5q7RVRgaJpZM4GexU4
.

@martinwicke as in here it seems the minimum known working Cmake version is 3.5. Should I change minimum requirement to 3.5 and update the readme for minimum 3.5 up to 3.6?

Nice to see that TensorFlow can now work on Windows with GPU now. Are there plans for a version that need not be installed from sources? And if so, any ETA on that? It would be great if the setup process were as straightforward as on Linux, I'm using TensorFlow on Linux without problems and would like to do so on Windows, but the requisites and setup process documented in the readme currently looks a bit cumbersome for those of us that don't have much time...

Hey guys! With the release of Bazel 0.4.0, TensorFlow PIP package can now build with Bazel on Windows.
Thanks to @petemounce's contribution, you can use choco install bazel --version 0.4.0 to install Bazel easily.

@gunan has already set up a ci job for it: http://ci.tensorflow.org/job/tensorflow-pr-win-bazel/
Basically, we are using this script. So with proper modification to the environment variables (if your installations of Visual Studio, Python 3 or msys is not at the standard places), you'll be able to build the PIP package with Bazel as well!

Update: I would recommend to run the commands in the script manually if you have trouble running it directly.

@meteorcloudy Thanks for update!
Build with bazel did not work for me. I get ERROR: CreateFile(C:\tmp\Bazel\44EoGNou\install)
Any way to debug that? I have localized windows (russian), maybe that's the cause

Full log. Running in PowerShell as admin:

C:\libs\tensorflow>c:\tools\msys64\usr\bin\bash -l C:\libs\tensorflow/tensorflow/tools/ci_build/windows/cpu/pip/build_tf_windows.sh
+ set -e
++ dirname 'C:\libs\tensorflow/tensorflow/tools/ci_build/windows/cpu/pip/build_tf_windows.sh'
+ script_dir='C:\libs\tensorflow/tensorflow/tools/ci_build/windows/cpu/pip'
+ cd 'C:\libs\tensorflow/'
+ export TMPDIR=C:/tmp
+ TMPDIR=C:/tmp
+ export BAZEL_SH=C:/tools/msys64/usr/bin/bash
+ BAZEL_SH=C:/tools/msys64/usr/bin/bash
+ export 'PYTHON_BIN_PATH=C:\tools\Anaconda3/python'
+ PYTHON_BIN_PATH='C:\tools\Anaconda3/python'
+ export 'BAZEL_PYTHON=C:\tools\Anaconda3/python'
+ BAZEL_PYTHON='C:\tools\Anaconda3/python'
+ export 'BAZEL_VS=C:/Program Files (x86)/Microsoft Visual Studio 14.0'
+ BAZEL_VS='C:/Program Files (x86)/Microsoft Visual Studio 14.0'
+ export 'PATH=/c/tools/bazel:/c/Program Files/Anaconda3:/usr/local/bin:/usr/bin:/bin:/opt/bin:/c/Windows/System32:/c/Windows:/c/Windows/System32/Wbem:/c/Windows/System32/WindowsPowerShell/v1.0/'
+ PATH='/c/tools/bazel:/c/Program Files/Anaconda3:/usr/local/bin:/usr/bin:/bin:/opt/bin:/c/Windows/System32:/c/Windows:/c/Windows/System32/Wbem:/c/Windows/System32/WindowsPowerShell/v1.0/'
+ bazel clean
Extracting Bazel installation...
ERROR: CreateFile(C:\tmp\Bazel\44EoGNou\install): ▒▒ ▒▒▒▒▒▒▒ ▒▒▒▒▒ ▒▒▒▒▒▒▒▒▒ ▒▒▒▒.
 (2)
.....................................................................................................................................................................................................................................................................................................................................
INFO: Starting clean (this may take a while). Consider using --expunge_async if the clean takes more than several minutes.
++ bazel info output_base
+ output_base=C:/tmp/Bazel/44EoGNou
+ bazel shutdown
+ rm -rf C:/tmp/Bazel/44EoGNou
rm: невозможно удалить 'C:/tmp/Bazel/44EoGNou/server/jvm.out': Device or resource busy

@KhabarlakKonstantin That's (ERROR: CreateFile(C:\tmp\Bazel\44EoGNou\install)) bazelbuild/bazel#1744 I think. It only happens the first time after bazel is installed. There don't seem to be any repercussions.

@KhabarlakKonstantin As @petemounce pointed out, you can safely ignore that error. And you can also skip rm -rf C:/tmp/Bazel/44EoGNou if the Device or resource busy error is blocking you. I guess it's because it tried to remove jvm.out before jvm fully shutdown after invoking bazel shutdown.

@petemounce @meteorcloudy Thanks for help
The problem is that it doesn't start building. I've commented out bazel shutdown and it goes now a little bit furhter, but still crashes

INFO: Starting clean (this may take a while). Consider using --expunge_async if the clean takes more than several minutes.
++ bazel info output_base
+ output_base=C:/tmp/Bazel/44EoGNou
+ echo ''
+ ./configure
/c/libs/tensorflow /c/libs/tensorflow
Found possible Python library paths:
  C:\Program Files\Anaconda3\lib\site-packages
  C:\Program Files\Anaconda3
Please input the desired Python library path to use.  Default is [C:\Program Files\Anaconda3\lib\site-packages]
Using python library path: C:\Program Files\Anaconda3\lib\site-packages
Junction created for util\python\python_include <<===>> C:\Program Files\Anaconda3\include
Junction created for util\python\python_lib <<===>> C:\Program Files\Anaconda3\lib\site-packages
Junction created for third_party\py\numpy\numpy_include <<===>> C:\Program Files\Anaconda3\lib\site-packages\numpy\core\include

No more output here

@KhabarlakKonstantin It seems OpenCL configuration is just added into ./configure but not turned off by default on Windows, so please manually run that script or add export TF_NEED_OPENCL=0 before echo "" | ./configure

pcloudy@PCLOUDY1-W MSYS ~/workspace/tensorflow
$ ./configure
~/workspace/tensorflow ~/workspace/tensorflow
Please specify the location of python. [Default is /c/Program Files/Anaconda3/python]:
Found possible Python library paths:
  C:\Program Files\Anaconda3\lib\site-packages
  C:\Program Files\Anaconda3
Please input the desired Python library path to use.  Default is [C:\Program Files\Anaconda3\lib\site-packages]

Using python library path: C:\Program Files\Anaconda3\lib\site-packages
Junction created for util\python\python_include <<===>> C:\Program Files\Anaconda3\include
Junction created for util\python\python_lib <<===>> C:\Program Files\Anaconda3\lib\site-packages
Junction created for third_party\py\numpy\numpy_include <<===>> C:\Program Files\Anaconda3\lib\site-packages\numpy\core\include
Do you wish to build TensorFlow with OpenCL support? [y/N] n
No OpenCL support will be enabled for TensorFlow
Configuration finished

Is anybody else having these errors with Pragma (ignore pywrap one)?

Building with GPU and all specs in accord to readme.

@Carmezim, it worked fine for me. I did not get any of these errors.

I have a pr to fix this here: https://github.com/tensorflow/tensorflow/pull/5421

@guschmue Nice! Thanks for the quick response

After fixing _Pragma, this is what I get in the same file:

C:/tensorflow/tensorflow/core/kernels/parameterized_truncated_normal_op_gpu.cu.cc(168): error : more than
one operator "/" matches these operands: [C:\tensorflow\tensorflow\contrib\cmake\build\tf_core_gpu_kernels.v
cxproj]

5421 will fix this as well.

Can I build with GPU support using bazel?

@edwin100394 no GPU support on windows using bazel yet.
We are still working on it.

On Mon, Nov 7, 2016 at 9:33 PM, edwin100394 [email protected]
wrote:

Can I build with GPU support using bazel?

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
https://github.com/tensorflow/tensorflow/issues/17#issuecomment-259049695,
or mute the thread
https://github.com/notifications/unsubscribe-auth/AHlCOXNRNoMFmnSPfnkTMLsRHcIcl9TXks5q8AmigaJpZM4GexU4
.

@meteorcloudy Adding export TF_NEED_OPENCL=0 helped me to go a little further

Now I'm getting another error:

ERROR: C:/libs/tensorflow/tensorflow/core/BUILD:1115:1: Executing genrule //tensorflow/core:version_info_gen failed: bash failed: error executing command
  cd C:/tmp/Bazel/44EoGNou/execroot/tensorflow
  SET PATH=C:\tools;C:\tools\Anaconda;C:\tools\msys64\usr\local\bin;C:\tools\msys64\usr\bin;C:\tools\msys64\usr\bin;C:\tools\msys64\opt\bin;C:\Windows\System32;C:\Windows;C:\Windows\System32\Wbem;C:\Windows\System32\WindowsPowerShell\v1.0\;C:\tools\msys64\usr\bin\site_perl;C:\tools\msys64\usr\bin\vendor_perl;C:\tools\msys64\usr\bin\core_perl
  C:/tools/msys64/usr/bin/bash -c source external/bazel_tools/tools/genrule/genrule-setup.sh; tensorflow/tools/git/gen_git_source.py --generate tensorflow/tools/git/gen/spec.json tensorflow/tools/git/gen/head tensorflow/tools/git/gen/branch_ref "bazel-out/vc_14_0_x64-py3-opt/genfiles/tensorflow/core/util/version_info.cc": com.google.devtools.build.lib.shell.BadExitStatusException: Process exited with status 1
Traceback (most recent call last):
  File "tensorflow/tools/git/gen_git_source.py", line 260, in <module>
    generate(args.generate)
  File "tensorflow/tools/git/gen_git_source.py", line 212, in generate
    git_version = get_git_version(data["path"])
  File "tensorflow/tools/git/gen_git_source.py", line 152, in get_git_version
    str("--work-tree=" + git_base_path), "describe", "--long", "--dirty", "--tags"
  File "C:\tools\Anaconda\lib\subprocess.py", line 626, in check_output
    **kwargs).stdout
  File "C:\tools\Anaconda\lib\subprocess.py", line 693, in run
    with Popen(*popenargs, **kwargs) as process:
  File "C:\tools\Anaconda\lib\subprocess.py", line 947, in __init__
    restore_signals, start_new_session)
  File "C:\tools\Anaconda\lib\subprocess.py", line 1224, in _execute_child
    startupinfo)
FileNotFoundError: [WinError 2] The system cannot find the file specified
Target //tensorflow/tools/pip_package:build_pip_package failed to build
INFO: Elapsed time: 10.997s, Critical Path: 3.44s
FAILED: Build did NOT complete successfully

I've started from scratch and I've build bazel from sources. Now it building almost till the end

It fails at link time in the end:

ERROR: C:/libs/tensorflow/tensorflow/python/BUILD:1907:1: Linking of rule '//tensorflow/python:_pywrap_tensorflow.so' failed: msvc_link.bat failed: error executing command
  cd C:/tmp/Bazel/44EoGNou/execroot/tensorflow
  SET PATH=external/local_config_cc/wrapper/bin
  external/local_config_cc/wrapper/bin/msvc_link.bat /DLL /WHOLEARCHIVE -pthread -m64 -Xcompilation-mode=opt -Wl,@bazel-out/vc_14_0_x64-py3-opt/bin/tensorflow/python/_pywrap_tensorflow.so-2.params: com.google.devtools.build.lib.shell.BadExitStatusException: Process exited with status 1169
Warning: Unmatched arguments: -ldl -lm -ldl -lm
Warning: Unmatched arguments: -pthread
libversion_lib.a(version_info.o) : error LNK2005: "char const * __cdecl tf_compiler_version(void)" (?tf_compiler_version@@YAPEBDXZ) already defined in libframework_internal.lo(version_info.o)
libversion_lib.a(version_info.o) : error LNK2005: "char const * __cdecl tf_git_version(void)" (?tf_git_version@@YAPEBDXZ) already defined in libframework_internal.lo(version_info.o)
   Creating library bazel-out/vc_14_0_x64-py3-opt/bin/tensorflow/python/_pywrap_tensorflow.lib and object bazel-out/vc_14_0_x64-py3-opt/bin/tensorflow/python/_pywrap_tensorflow.exp
bazel-out/vc_14_0_x64-py3-opt/bin/tensorflow/python/_pywrap_tensorflow.so : fatal error LNK1169: one or more multiply defined symbols found
ERROR: C:/libs/tensorflow/tensorflow/python/BUILD:639:1: output 'tensorflow/python/gen_control_flow_ops_py_wrappers_cc.exe' was not created
ERROR: C:/libs/tensorflow/tensorflow/python/BUILD:698:1: output 'tensorflow/python/gen_resource_variable_ops_py_wrappers_cc.exe' was not created
Target //tensorflow/tools/pip_package:build_pip_package failed to build

Hi folks, now that rudimentary Windows support is in place, I am going to lock this issue for new comments, but keep it open until we have a stable release. Going forward:

• If you have an issue, please open a new issue against TensorFlow to describe your problem, and fill out the standard issue template. This will help us to track and assign the remaining work that we have to do to get TensorFlow-on-Windows released.
• We will continue to use this issue for (rare) announcements about Windows support, so you can continue to subscribe to the issue for a low-traffic feed about improvements to TensorFlow-on-Windows.

Once again, thanks for all your interest in TensorFlow-on-Windows, and now we're going to get back to work!

Today we announced native Windows support in TensorFlow 0.12, with packages for Python 3.5. If you have installed the 64-bit version of Python 3.5 (either from Python.org or Anaconda), you can install TensorFlow with a single command:

C:\> pip install tensorflow

For GPU support, if you have CUDA 8.0 installed, you can install the following package instead:

C:\> pip install tensorflow-gpu

Now that Windows is a supported configuration, I'm going to close this issue; please open a new issue if you have problems with these packages. There are still some discrepancies between the Windows package and the other platforms, which are outlined in the release notes. We'll be working over the next few weeks to close these gaps.

Thanks for all of your interest in TensorFlow on Windows!