tensorflow-in-windows (2/2)

tensorflow-in-windows (2/2)

• previous post covers
  • install python using anaconda
  • install tensorflow-cpu, which supports cpu only
  • install vs-code, which can be used for editor
• this post covers
  • install cuda
  • install tensorflow-gpu, which supports gpu too
• environment
  • Windows / 10
  • Anaconda / 4.6.11

---

anaconda-and-python

• I regard you have already followed previous post

• let us create new virtual environment for installing tensorflow-gpu

cuda

• tensorflow supports gpu for the faster matrix calculation
  • been gpu specialized for 3d graphics operations, and many portions of the operations are the matrix calculation
• especially, google's tensorflow team has cooperated with NVIDIA for several years
  • so that is why you should use CUDA for tensorflow-gpu
  • IOW, AMD gpu and Intel IGP cannot be used for tensorflow-gpu

• visit here to download the installer for CUDA
  • I recommend you to install CUDA its version >= 10.0
  • in this post, I will cover the CUDA 10.0
• select the proper installer fits your system
  • local installer would be better for general purposes

• open the executable with compression program
  • ex) bandizip
  • decompress the executable into folder

• when the decompression is over, execute the setup
  • in this way, each time execution of installer you can avoid generating temporary files for installation
  • while the installing, follow default options

 

• if you have some errors while installing, follow the instructions below
  • if you have multiple versions of CUDA, uninstall all
  • select custom at the installation options
  • uncheck the options except for CUDA
  • expand the CUDA option and uncheck Visual Studio Integration
  • and now good to go

• launch the cmd
  • type nvcc and check whether nvcc executes
  • nvcc is used for compiling CUDA codes

cudnn

• visit here to download the cudnn library as zip file
  • cudnn is acronym of CUda for Deep Neural Network
  • this library includes useful and "essential" features for using tensorflow-gpu
• "essential" means, without the cudnn you cannot use tensorflow-gpu
  • some import errors would happen

• decompress the zip and follow the guide
  • in windows, you should place the library files at proper path
  • the proper path would be different upon your CUDA version

• copy dll file into the path

• copy header file into the path

• copy lib file into the path

tensorflow

• activate the environment we installed
  • type pip install tensorflow-gpu>=1.13.0
  • CUDA 10.0 is supported with 1.13.0 or over

• you can check whether tensorflow-gpu is successfully installed with python
  • type the command we used in previous post
  • now you can use tensorflow accelerated by gpu

 

• if you encounter this error, visit here to download latest nvidia graphics driver
  • follow the instructions when we used to install CUDA

• execute the setup file and keep default settings
  • and retry the test process