TensorFlow runtime precompiled binary for Windows about tensorflowsharp

[jamasu]

the download URL is not working anymore.

[mwk888]

@asimshankar dll seems fine; I re-wrote this post with correct integration steps.

Successful steps:

• download and unzip a 1.2 dll build e.g. I used http://ci.tensorflow.org/view/Nightly/job/nightly-libtensorflow-windows/lastSuccessfulBuild/artifact/lib_package/libtensorflow-cpu-windows-x86_64.zip
• drag tensorflow.dll onto the Visual Studio project (e.g. SampleTest). This will copy the dll file to the project
• In VS rename tensorflow.dll to libtensorflow.dll (it must match the TensorFlowLibrary value in Tensorflow.cs)
• In VS set libtensorflow.dll properties > Advanced > Copy to Output Directory > Copy Always

After this SampleTest successfully loads the dll and runs until the the line helpfully commented as "// Current failing test" for which you can't blame the dll!

fyi my original mistake when integrating with Visual Studio was trying to Add Reference to the dll which gives the error "A reference to ...\tensorflow.dll could not be added. Please make sure that the file is accessible, and that it is a valid assembly or COM component" (this is the wrong method to add a native dll).

[andykernahan]

@MovGP0 That exception is most likely because your platform target is not x64; libtensorflow.dll is 64bit.

[asimshankar]

If anyone is willing to give the .dll mentioned in tensorflow/tensorflow#7258 (comment) a spin and leave any feedback on that issue, that would be great.

[migueldeicaza]

I created a NuGet package that should contain the Windows binary. A simple test shows it works.

[andykernahan]

The JNI builds do not export the C/C++ API, they only export the functions required by JNI:

> dumpbin /EXPORTS tensorflow_jni.dll
Microsoft (R) COFF/PE Dumper Version 14.00.24215.1
Copyright (C) Microsoft Corporation.  All rights reserved.


Dump of file tensorflow_jni.dll

File Type: DLL

  Section contains the following exports for libtensorflow_jni.so

    00000000 characteristics
    58D31E6B time date stamp Thu Mar 23 01:01:31 2017
        0.00 version
           1 ordinal base
          50 number of functions
          50 number of names

    ordinal hint RVA      name

          1    0 0188CAB0 ?DEVICE_CPU@tensorflow@@3QEBDEB
          2    1 0188CAB8 ?DEVICE_GPU@tensorflow@@3QEBDEB
          3    2 0188CAC0 ?DEVICE_SYCL@tensorflow@@3QEBDEB
          4    3 000F4CD0 Java_org_tensorflow_Graph_allocate
          5    4 000F4CE0 Java_org_tensorflow_Graph_delete
          6    5 000F4CF0 Java_org_tensorflow_Graph_importGraphDef
          7    6 000F4E30 Java_org_tensorflow_Graph_operation
          8    7 000F4ED0 Java_org_tensorflow_Graph_toGraphDef
         
         ...

Currently, only the cmake builds export the C/C++ API. If you do not want to build Tensorflow yourself, you can extract it from the CI whl artifact here:

1. download the wheel and unzip it
2. cd into .\*.data\purelib\tensorflow\python\
3. copy _pywrap_tensorflow_internal.pyd as libtensorflow.dll
4. copy libtensorflow.dll to your .NET project and you're all set!

> dumpbin /EXPORTS libtensorflow.dll
Microsoft (R) COFF/PE Dumper Version 14.00.24215.1
Copyright (C) Microsoft Corporation.  All rights reserved.


Dump of file libtensorflow.dll

File Type: DLL

 Section contains the following exports for _pywrap_tensorflow_internal.pyd

   00000000 characteristics
   58E45416 time date stamp Wed Apr  5 03:19:02 2017
       0.00 version
          1 ordinal base
      42374 number of functions
      42374 number of names

   ordinal hint RVA      name

         1    0 02783468 ?$TSS0@?1???$Initialize@$00@?$SparseSlice@Ubfloat16@tensorflow@@@?A0xa2b9e076@tensorflow@@QEAAXAEBV?$TensorMap@V?$Tensor@$$CBUbfloat16@tensorflow@@$01$00_J@Eigen@@$0BA@UMakePointer@2@@Eigen@@H@Z@4HA
         2    1 02783470 ?$TSS0@?1???$Initialize@$0A@@?$SparseSlice@Ubfloat16@tensorflow@@@?A0xa2b9e076@tensorflow@@QEAAXAEBV?$TensorMap@V?$Tensor@$$CBUbfloat16@tensorflow@@$01$00_J@Eigen@@$0BA@UMakePointer@2@@Eigen@@H@Z@4HA
         3    2 027812F0 ?$TSS0@?1???$locale_independent_strtonum@M@?A0x1bfb4b96@tensorflow@@YAMPEBDPEAPEBD@Z@4HA
         ...
     42247 A506 0013FC10 PyInit__pywrap_tensorflow_internal
     42248 A507 001239B0 TF_AbortWhile
     42249 A508 001239F0 TF_AddControlInput
     42250 A509 00123A00 TF_AddInput
     42251 A50A 00123A10 TF_AddInputList
     42252 A50B 00123CA0 TF_AllocateTensor
     42253 A50C 00123D10 TF_CloseDeprecatedSession
     42254 A50D 00123D10 TF_CloseSession
     42255 A50E 00123D60 TF_ColocateWith
     42256 A50F 00123EB0 TF_DataTypeSize
     42257 A510 00123EC0 TF_DeleteBuffer
     42258 A511 00123EF0 TF_DeleteDeprecatedSession
     42259 A512 00123F70 TF_DeleteGraph
     42260 A513 00123FE0 TF_DeleteImportGraphDefOptions
     42261 A514 00124010 TF_DeleteLibraryHandle
     42262 A515 018411A0 TF_DeletePRunHandle
     42263 A516 00124040 TF_DeleteSession
     42264 A517 00124140 TF_DeleteSessionOptions
     42265 A518 00124180 TF_DeleteStatus
     42266 A519 001241D0 TF_DeleteTensor
     42267 A51A 00124230 TF_Dim
     42268 A51B 00124240 TF_ExtendGraph
     ...          

[MovGP0]

that seems overly complicated. there should be a powershell build script which determines the windows version, downloads the proper files and puts the dll in the proper location, builds everything, and executes the tests. shouldn't be too hard.

[migueldeicaza]

@MovGP0 You do not need to manually copy the native files, it should be done for you.

The warnings are normal.

[mwk888]

My suggestion is to build and integrate a lowest-common-denominator x86 dll with TensorFlowSharp until Google distributes an official Windows dll. That missing dll is a big barrier for C# development right now. (I'm assuming the PR to fix python dependencies works)

My guess is the TF dev team will eventually switch to a universal x86 library with runtime switches for the CPU optimizations -- nothing else really works for app distribution. If that happens then the C API dll becomes NuGet and off we go :)

[szdree]

thank you,I will further test.

[andykernahan]

You could run the SampleTest, as per the README.

[andykernahan]

I suggest opening another issue to track those failures.

[goloskokovic]

use:
http://ci.tensorflow.org/view/Nightly/job/nightly-libtensorflow-windows/lastSuccessfulBuild/artifact/lib_package/libtensorflow-cpu-windows-x86_64.zip

deprecated:
you can download [libtensorflow.dll] taken from:
https://storage.googleapis.com/tensorflow/windows/cpu/tensorflow-1.1.0rc1-cp35-cp35m-win_amd64.whl

[JimSEOW]

t.WhileTest (); of SimpleTest will not work until this address????? Anyone as different view?
#26

[szdree]

operating system：window

I can't run, I'll libtensorflow_jni libtensorflow.dll and then put it into the bin folder, I do this right?

Error: untreated exception: System.DllNotFoundException: could not load DLL 'libtensorflow': the specified module could not be found. (exception from HRESULT:0x8007007E).
In TensorFlow.TFCore.TF_Version ()
In TensorFlow.TFCore.get_Version () position E:\Code\TensorFlow\Csharp\TensorFlowSharp-master\TensorFlowSharp\Tensorflow.cs: line number 54
In SampleTest.MainClass.Main (String[] args) location E:\Code\TensorFlow\Csharp\TensorFlowSharp-master\SampleTest\SampleTest.cs: line number 556

[andykernahan]

@szdree the JNI library will not work; follow steps I detailed above to get the correct libtensorflow.dll.

Edit: make sure to copy the library to the bin directory!

[szdree]

@andykernahan
Same result
operating system：window 10 64bit
CPU : Intel Core

Error: untreated exception: System.DllNotFoundException: could not load DLL 'libtensorflow': the specified module could not be found. (exception from HRESULT:0x8007007E).

Use：
HTTPS: //storage.googleapis.com/tensorflow/windows/cpu/tensorflow-1.1.0rc1-cp35-cp35m-win_amd64.whl
(purelib \ tensorflow \ python_pywrap_tensorflow_internal.pyd - > libtensorflow.dll)

copy libtensorflow.dll to bin\Debug
And SampleTest.exe in bin\Debug\

[vsjc91]

@szdree, I had the same issue. Running Dependency Walker revealed missing Python35. I installed Python3.5 and then copied that dll alongside in the Debug\Bin and it resolved the DllNotFoundException.

[szdree]

@vsjc91 Mine is Python3.5 I haven't solved the problem yet.

[goloskokovic]

try installing Anaconda:
https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/cmake/README.md#known-good-configurations

[szdree]

@goloskokovic I was originally installed Anaconda

and
I installed the tensorflow python version, which can be run on the window.

[avrshawn]

Running SampleTest (1.1.0-rc1) on Windows 10 / VS2015 using the above instructions (thanks!) I'm getting a System.AccessViolationException when it hits TestOuputShape->GetShape->TF_GraphGetTensorShape (TensorFlow.cs line 546) for the second constant.

I'm not seeing any obvious nulls, and the prior tests complete - has anybody else stubbed their toe on this one yet?

[kangchihlun]

@avrshawn same issue here , have no clue.

[andykernahan]

Indeed. Unfortunately there isn't an official Windows build which can be downloaded; the best option is still to build it yourself using cmake.

Note that the Python related dependency issues above may be mitigated by this PR (although I've not tried it).

[MovGP0]

Can't get it to work. I did the following:

• Creating a new console project and setting .NET target version to 4.6.1
• adding the TensorFlowSharp 0.96.0 nuget package
• adding a reference to the packages\TensorFlowSharp.0.96.0\lib\net461\libtensorflow.dll
• including all files in packages\TensorFlowSharp.0.96.0\native\ and setting build action to Content and Copy if newer

When starting the project, I get the following error:

System.BadImageFormatException occurred
  HResult=0x8007000B
  Message=Could not load file or assembly 'TensorFlowSharp, Version=1.0.0.0, Culture=neutral, PublicKeyToken=null' or one of its dependencies. An attempt was made to load a program with an incorrect format.

Windows version 10.0.15063.0

[MovGP0]

@andykernahan that worked. now only get a lot of warnings. nothing problematic:

2017-05-21 10:22:25.310807: W C:\tmp\_bazel_system\6fx6hpz4\execroot\nightly-libtensorflow-windows\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE instructions, but these are available on your machine and could speed up CPU computations.
2017-05-21 10:22:25.311268: W C:\tmp\_bazel_system\6fx6hpz4\execroot\nightly-libtensorflow-windows\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE2 instructions, but these are available on your machine and could speed up CPU computations.
2017-05-21 10:22:25.313023: W C:\tmp\_bazel_system\6fx6hpz4\execroot\nightly-libtensorflow-windows\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE3 instructions, but these are available on your machine and could speed up CPU computations.
2017-05-21 10:22:25.313090: W C:\tmp\_bazel_system\6fx6hpz4\execroot\nightly-libtensorflow-windows\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.1 instructions, but these are available on your machine and could speed up CPU computations.
2017-05-21 10:22:25.313154: W C:\tmp\_bazel_system\6fx6hpz4\execroot\nightly-libtensorflow-windows\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.2 instructions, but these are available on your machine and could speed up CPU computations.
2017-05-21 10:22:25.313218: W C:\tmp\_bazel_system\6fx6hpz4\execroot\nightly-libtensorflow-windows\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX instructions, but these are available on your machine and could speed up CPU computations.
2017-05-21 10:22:25.313279: W C:\tmp\_bazel_system\6fx6hpz4\execroot\nightly-libtensorflow-windows\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX2 instructions, but these are available on your machine and could speed up CPU computations.
2017-05-21 10:22:25.313332: W C:\tmp\_bazel_system\6fx6hpz4\execroot\nightly-libtensorflow-windows\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use FMA instructions, but these are available on your machine and could speed up CPU computations.

[andykernahan]

@MovGP0 Yes, they are expected as the CI builds are not built with march=native.

[lostmsu]

GPU one?

[pablonautic]

I've just tried running SampleTest under Win10 / MS .NET Framework 4.6.1 and I get the following runtime error:

TensorFlow version: 1.2.0-rc0
Testing Parameters with indexes
2017-05-24 22:16:50.519124: W C:\tmp\_bazel_system\6fx6hpz4\execroot\nightly-libtensorflow-windows\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE instructions, but these are available on your machine and could speed up CPU computations.
2017-05-24 22:16:50.519645: W C:\tmp\_bazel_system\6fx6hpz4\execroot\nightly-libtensorflow-windows\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE2 instructions, but these are available on your machine and could speed up CPU computations.
2017-05-24 22:16:50.519684: W C:\tmp\_bazel_system\6fx6hpz4\execroot\nightly-libtensorflow-windows\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE3 instructions, but these are available on your machine and could speed up CPU computations.
2017-05-24 22:16:50.519719: W C:\tmp\_bazel_system\6fx6hpz4\execroot\nightly-libtensorflow-windows\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.1 instructions, but these are available on your machine and could speed up CPU computations.
2017-05-24 22:16:50.519754: W C:\tmp\_bazel_system\6fx6hpz4\execroot\nightly-libtensorflow-windows\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.2 instructions, but these are available on your machine and could speed up CPU computations.
success
Testing AddControlInput for assertions
Success, got the expected exception when using tensorflow control inputs to assert
Handle: 2007385971504
FinishHandle: 2007384924392
Handle: 2007385985616
FinishHandle: 2007384924392
Variables
Result of variable read 0 -> 1.5
Result of variable read 1 -> 2
Result of variable read 2 -> 2.5
Result of variable read 3 -> 3
Result of variable read 4 -> 3.5

Unhandled Exception: System.EntryPointNotFoundException: Unable to find an entry point named 'TF_NewWhile' in DLL 'tensorflow.dll'.
   at TensorFlow.TFGraph.TF_NewWhile(IntPtr g, TFOutput[] inputs, Int32 ninputs, IntPtr status)
   at TensorFlow.TFGraph.While(TFOutput[] inputs, WhileConstructor constructor, TFStatus status) in C:\Users\pawelgorczynski\Repos\TensorFlowSharp\TensorFlowSharp\Tensorflow.cs:line 846
   at SampleTest.MainClass.WhileTester.Init(Int32 ninputs, WhileConstructor constructor) in C:\Users\pawelgorczynski\Repos\TensorFlowSharp\SampleTest\LowLevelTests.cs:line 189
   at SampleTest.MainClass.WhileTest() in C:\Users\pawelgorczynski\Repos\TensorFlowSharp\SampleTest\LowLevelTests.cs:line 220
   at SampleTest.MainClass.Main(String[] args) in C:\Users\pawelgorczynski\Repos\TensorFlowSharp\SampleTest\SampleTest.cs:line 319
Press any key to continue . . .

The behavior is the same for both
https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow-cpu-windows-x86_64-1.2.0-rc0.zip
and
http://ci.tensorflow.org/view/All/job/nightly-libtensorflow-windows/

[asimshankar]

Ah, it seems that the TF_While related symbols aren't being exported in the .dll. I'll try to get that fixed for 1.2-rc1 or 1.3.

In the mean time, any while loop functionality will not be available on Windows.

[FiveTechSoft]

@asimshankar would it be possible to have a tensorflow.dll 32 bits too ?

[FiveTechSoft]

Properly using TensorFlow.dll from Harbour:

http://forums.fivetechsupport.com/viewtopic.php?p=202354#p202354

https://harbour.github.io/

many thanks :-)

[FiveTechSoft]

I get this when using the TensorFlow.dll:

2017-06-06 12:48:43.269929: W C:\tmp_bazel_system\6fx6hpz4\execroot\nightly-libtensorflow-windows\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE instructions, but these are available on your machine and could speed up CPU computations.
2017-06-06 12:48:43.270149: W C:\tmp_bazel_system\6fx6hpz4\execroot\nightly-libtensorflow-windows\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE2 instructions, but these are available on your machine and could speed up CPU computations.
2017-06-06 12:48:43.272532: W C:\tmp_bazel_system\6fx6hpz4\execroot\nightly-libtensorflow-windows\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE3 instructions, but these are available on your machine and could speed up CPU computations.
2017-06-06 12:48:43.273426: W C:\tmp_bazel_system\6fx6hpz4\execroot\nightly-libtensorflow-windows\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.1 instructions, but these are available on your machine and could speed up CPU computations.
2017-06-06 12:48:43.273932: W C:\tmp_bazel_system\6fx6hpz4\execroot\nightly-libtensorflow-windows\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.2 instructions, but these are available on your machine and could speed up CPU computations.

[andykernahan]

@FiveTechSoft That is expected.

[anguoyang]

how to compile the dll from the latest tf source code?

[Digital2Slave]

how to compile the 32bit dll from the latest tf source code? I want to use the 32bit dll in my own c/c++ project on windows7. @andykernahan thanks very much.

[migueldeicaza]

I am downloading the binaries from the daily build hosts from TensorFlow.

I feel that questions on how to download the native bits should go into the tensorflow github repo, not here.

[fastlater]

@mwk888 Thanks for your post. It solved my problem.
When I was using tensorflowSharp 1.3, I had no problems but when I updated to version 1.5, I got this error.
1 year later, using tensorflowSharp 1.5 downloaded using nuGet packages manager in vs2015 and still need to do this? Why it was working tensorflowSharp 1.3 without have to do this but when I upgraded, the error came out? just curious about this.