Setting Up Tensorflow (with CUDA) for Windows 10

Below are some of my rough notes on how I've setup my Windows 10 laptop to use Tensorflow with CUDA.

My reference:

1. http://www.heatonresearch.com/2017/01/01/tensorflow-windows-gpu.html
2. https://www.tensorflow.org/install/

Install the following NVidia drivers:

1. CUDA Drivers (http://docs.nvidia.com/cuda/cuda-installation-guide-microsoft-windows). Currently it's CUDA Toolkit 8.0. Can install using the installer downloaded.
2. CUDNN - CUDA for Deep Neural Networks (https://developer.nvidia.com/cudnn). Currently it's 5.1. Once extracted, place the files in the respective directory along with the other CUDA files in the NVIDIA Toolkit folder

Setting up Tensorflow (CPU)

	conda create --name tensorflow python=3.5
	activate tensorflow
	conda install jupyter (this might fail due to :PaddingError: Placeholder of length '30' too short in package qt-5.6.2-vc14_0.% The package must be rebuilt with conda-build > 2.0. Try running 'conda update --all' in root env, and running it again)
	conda install scipy
	pip install tensorflow

view raw install_tf_cpu hosted with ❤ by GitHub

Setting up Tensorflow (GPU)

	conda create --name tensorflow-gpu python=3.5
	activate tensorflow-gpu
	conda install jupyter
	conda install scipy
	pip install tensorflow-gpu

view raw install_tf_gpu hosted with ❤ by GitHub

Note: It's 22/2/2017 now and Google have recently released their Tensorflow 1.0, which might've rendered the above guide obsolote (i've haven't tested them yet).

Update (23/2/2017):
The above basically creates 2 new anaconda instance for you to play with. I like this approach since it's cleaner and does not mess up my original Anaconda installation - easier to troubleshoot (and destroy if need be). To select the instance that you wish to run, and run jupyter - you may use the following steps:

	C:\Users\UserName>activate tensorflow-gpu
	(tensorflow-gpu) C:\Users\UserName>jupyter
	usage: jupyter-script.py [-h] [--version] [--config-dir] [--data-dir]
	[--runtime-dir] [--paths] [--json]
	[subcommand]
	jupyter-script.py: error: one of the arguments --version subcommand --config-dir --data-dir --runtime-dir --paths is required
	(tensorflow-gpu) C:\Users\UserName>jupyter notebook
	[I 01:02:35.990 NotebookApp] [nb_conda_kernels] enabled, 4 kernels found
	[I 01:02:41.129 NotebookApp] [nb_anacondacloud] enabled
	[I 01:02:41.176 NotebookApp] [nb_conda] enabled
	[I 01:02:42.015 NotebookApp] \u2713 nbpresent HTML export ENABLED
	[W 01:02:42.015 NotebookApp] \u2717 nbpresent PDF export DISABLED: No module named 'nbbrowserpdf'
	[I 01:02:42.147 NotebookApp] Serving notebooks from local directory: c:/Hafidz/Playground/
	[I 01:02:42.147 NotebookApp] 0 active kernels
	[I 01:02:42.147 NotebookApp] The Jupyter Notebook is running at: http://localhost:8888/?token=38564ad6bc871a295d6f9f0ddcbb6c2e9a459ff9cf8860fb
	[I 01:02:42.147 NotebookApp] Use Control-C to stop this server and shut down all kernels (twice to skip confirmation).
	[C 01:02:42.162 NotebookApp]
	Copy/paste this URL into your browser when you connect for the first time,
	to login with a token:
	http://localhost:8888/?token=38564ad6bc871a295d6f9f0ddcbb6c2e9a459ff9cf8860fb
	[I 01:02:42.581 NotebookApp] Accepting one-time-token-authenticated connection from ::1

view raw run_tensorflowgpu_jupyter.txt hosted with ❤ by GitHub

And your Jupyter should've already loaded by now.