Tensorflow 1.7 GPU can be compiled in Windows and used in TensorflowSharp

[edwardyehuang]

Just let you guys know. I have succeeded compiled the Tensorflow 1.7 GPU version, and also successfully used in TensorflowSharp with the large graph model (Before the Tensorflow GPU 1.5, the size of the pb graph is limited at 64mb)

In order to compile the lastest Tensorflow dll, you need to use the CMake:
https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/cmake

Be sure enable the GPU support and shared library support when using the CMake to generate the project files

Be sure using the Visual Studio 2015 to compile. The Cuda version is 9.0 and cudnn version is the lastest.

Be sure to change " tf_core_kernels.vcxproj" properties:
Properties->Configuration Properties->VC++ Directories->Executable Directories
$(VC_ExecutablePath_x64) to $(VC_ExecutablePath_x64_x64)

After compiling, make sure you added your cudnn/bin folder to "PATH". And rename the tensorflow.dll to libtensorflow.dll

[migueldeicaza]

This is excellent - could you provide build instructions?

It looks like I am going to have to provide those builds myself.

[edwardyehuang]

1, Download the Tensorflow 1.7 source code.

2, Be sure you have Cuda 9.0, cudnn latest version and Visual Studio 2015 installed

3, Follow the tutorial in https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/cmake
Make sure you change your CMake command like below (Change to your path)
cmake -G "Visual Studio 14" .. -A x64 -DCMAKE_BUILD_TYPE=Release ^
-DSWIG_EXECUTABLE=C:/swigwin/swig.exe ^
-DPYTHON_EXECUTABLE=C:/ProgramData/Anaconda3/python.exe ^
-DPYTHON_LIBRARIES=C:/ProgramData/Anaconda3/libs/python36.lib ^
-Dtensorflow_ENABLE_GPU=ON ^
-Dtensorflow_BUILD_SHARED_LIB=ON ^
-DCUDNN_HOME="C:/cuda7"

4, Open the generated project. Change " tf_core_kernels.vcxproj" properties:
Properties->Configuration Properties->VC++ Directories->Executable Directories
$(VC_ExecutablePath_x64) to $(VC_ExecutablePath_x64_x64)

5, Build

6, Make sure you copy (Or add to your project) the cudnn dll used in build

`

[edwardyehuang]

Btw, the dll will contain some private information about the machine you built it, including the path that you placed the Tensorflow source code, just be careful about that.

[YanYas]

No luck for me. Anyone else give this a try? I kept getting an error that regarding the tf_stream_executor.dir usually to do with being unable to open a file.
If someone wouldn't mind leaving a compiled binary somewhere on the internet, it would be much appreciated.

[edwardyehuang]

@YanYas Could you tell me which version of Tensorflow you are trying to build?

[YanYas]

Hi @edwardyehuang.
I'm trying to build TF 1.7 for Windows 10 PC, Nvidia 1070 Ti with Cuda 9.0.176 and the matching CuDNN installed.
3 Errors similar to this:

No such file or directory (compiling source file C:\TFbuild\tensorflow\tensorflow\stream_executor\cuda\cudnn_ version.cc) [C:\TFbuild\tensorflow\tensorflow\contrib\cmake\build\tf_stream_executor.vcxproj]

[YanYas]

To add, before hand I installed Tensorflow-gpu for python to be sure the Cuda paths were configured properly.

[edwardyehuang]

@YanYas I do not have this file either, in "tensorflow\stream_executor\cuda".

I just checked the Tensorflow repo. This file is created was 11 days ago (Mar 29, 2018).

The source code I downloaded is from https://github.com/tensorflow/tensorflow/releases

So maybe is not included in the 1.7 release.

Firstly, make sure the one you downloaded is from 1.7 released version (The link above)

Then, could you provide more exact detail about the error?

[YanYas]

I see, yeah I cloned the git repo instead. Thanks! Okay, building...

[YanYas]

Completed without errors! Thanks very very much!

[ivarsdzalbs]

Any chance someone could upload the compiled dll?
Thanks.

[jacano]

Hi guys, I've just compiled TF with GPU support for Windows in x64/Release using the instructions provided by @edwardyehuang. Thank you so much!
I used the commit (4fb125264c5394c9e4295ed437adb1d9711bd456) from master branch on (May 09, 2018)
The 1.7.0 version of TensorFlowSharp nuget package is able to correctly PInvoke on it. It's a pretty heavy binary (~270MB) :S

I get the following output when initializing my TF graph in C# from the inference model:

Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX AVX2

Found device 0 with properties:
name: GeForce GTX 650 major: 3 minor: 0 memoryClockRate(GHz): 1.0715
pciBusID: 0000:01:00.0
totalMemory: 2.00GiB freeMemory: 1.65GiB
Adding visible gpu devices: 0
....
Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 1436 MB memory) -> physical GPU (device: 0, name: GeForce GTX 650, pci bus id: 0000:01:00.0, compute capability: 3.0)

So far, so good. I'm uploading my binary just in case someone want to try it.
Link: libtensorflow_gpu.dll

[freiz]

There is a repo host pre built windows Tensorflow binary may help, https://github.com/fo40225/tensorflow-windows-wheel