Introduction

We are going to install tensorflow-gpu from scratch on a Windows 10 machine. You will find all the right version numbers. All the traps to dodge. All the solutions to the problems I encountered.

My system is:

• Windows 10 64
• Nvidia GeForce 1070

Installing everything

Python

You need Python3.5.4.

Go to https://www.python.org/downloads/release/python-354rc1/

I downloaded the Windows x86-64 executable installer

This should give you:

$ python --version
Python 3.5.4

$ pip --version
pip 9.0.3 from [...]\local\programs\python\python35\lib\site-packages (python 3.5)

Visual Studio

VB 2017

Visual studio is not really required but it is useful to know if your CUDA installation is working.

• Go to https://www.visualstudio.com/downloads and get the 2017 community edition.
• Install it.

2015 pack

• On this page: https://www.visualstudio.com/downloads/#build-tools-for-visual-studio-2017 get the Microsoft Visual C++ Redistributable for Visual Studio 2017
• From there install the VC++ 2015.3 v140 toolset for desktop (x86,x64).

CUDA

Main documentation: https://docs.nvidia.com/cuda/cuda-installation-guide-microsoft-windows/index.html

Install

Through this process this was the biggest challenge because I installed CUDA-9.1 but it is not compatible with tensorflow 1.7.0.

Uninstalling CUDA-9.1 was not easy and I am still battling with artifact and conflicts.

DO NOT INSTALL CUDA-9.1

Thanks for https://github.com/tensorflow/tensorflow/issues/17101 for the help !

Go to https://developer.nvidia.com/cuda-90-download-archive to download CUDA-9.0 and install it.

If you have an issue with it hanging, try:

• Disabling Avast
• Uncheck visual studio integration.

Issues with NSight

If you had to disable visual studio integration to install CUDA.

For CUDA-9.0, you need NSight-5.3.

Go to https://developer.nvidia.com/nsight-visual-studio-edition-5_3 and download https://developer.nvidia.com/rdp/assets/nsight-visual-studio-5_3-general-win10-64-driver

I am still having issues right now and have an open message here: https://devtalk.nvidia.com/default/topic/1031732/visual-studio-integration-9-0-fail-to-install-after-removing-cuda-9-1/#5249049.

I will update as I get news.

Stay tuned.

Add a path to the environment variable PATH in Windows 10

You need to add the right folder to the PATH environment variable. It is a bit convoluted to do that on windows.

• Right click on Computer
• Properties
• Advanced system settings
• In the tab Advanced, click on Environment Variables...
• Click on the line with Variable is Path
• Click Edit...
• In the new opened windows click New.

• Add C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.0\bin.

Testing if it works

Following the main documentation.

• Open C:\ProgramData\NVIDIA Corporation\CUDA Samples\v9.0\1_Utilities\deviceQuery\deviceQuery_vs2017.sln with visual studio 2017.
• In the Visual Studio UI, you should see debug at the top.
  • 
• Switch it to Release.
• And then in the menu bar: Build -> Build Solution.
• That should produce an exe file at C:\ProgramData\NVIDIA Corporation\CUDA Samples\v9.0\bin\win64\Release.
• Launch the binary through the cmd.
• You can reproduce this step with C:\ProgramData\NVIDIA Corporation\CUDA Samples\v9.0\1_Utilities\bandwidthTest.

Various Issues

Wrong package

Right click on the project in Visual Studio and in the properties where it says VC++ 2017, try using the 2015 one instead.

Python – Install tensorflow-gpu

You need to install tensorflow-gpu, keras and you will be ready to go.

Just do:

$ python -m pip install --upgrade pip

$ pip install tensorflow-gpu

$ pip install keras

And you should be good to go.

Testing

import tensorflow as tf
hello = tf.constant('Hello, TensorFlow!')
sess = tf.Session()
print(sess.run(hello))

Execute that.

Miscellanious

CUBLAS_STATUS_ALLOC_FAILED

Solution found at https://github.com/tensorflow/tensorflow/issues/7072.

Solution:

from keras.backend.tensorflow_backend import set_session

config = tf.ConfigProto(
    gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.8)
    # device_count = {'GPU': 1}
)
config.gpu_options.allow_growth = True
session = tf.Session(config=config)
set_session(session)

This will solve the memory issue on the GPU.