Before installing tensorflow, install CUDA 8.0 and Cudnn 5.1 first.
CUDA 7.5 is not compatible! https://github.com/tensorflow/tensorflow/issues/9469
Uninstall 7.5 and reboot if you have it previously installed.
Install TensorFlow as official guide
Download and install anaconda, in prompt
conda create --name tensorflow python=3.5 activate tensorflow
conda create --name tf python=3.6 activate tf (tensorflow)C:>conda install anaconda tensorflow-gpu
For old versions:
(tensorflow)C:> pip install --ignore-installed --upgrade https://storage.googleapis.com/tensorflow/windows/gpu/tensorflow_gpu-1.1.0rc1-cp35-cp35m-win_amd64.whl
conda and pip are two separate package management tools. However, for some unknown reasons, install tensorflow using pip will not let virtual environment find how to initiate tensorflow, resulting following errors.
(tensorflow)C:>pip install --ignore-installed --upgrade tensorflow-gpu python import tensorflow ImportError: No module named pywrap_tensorflow_internal Failed to load the native TensorFlow runtime.
So we need to install tensorflow using conda install.
conda install tensorflow-gpu
However, when you install tensorflow using pip, the files under Lib/site-packages will not be replaced by the conda installed tensorflow. By installing this second tensorflow is only letting virtual environment realize the existence of a package called tensorflow-gpu
For using Anaconda3 (or Miniconda3) you need to install nb_conda into the conda environment (in addition to jupyter):
conda install nb_conda jupyter
Then starting jupyter applications, for example
jupyter notebook
If there is an error about zmq package
pip uninstall pyzmq pip install pyzmq
even though there is a warning.
To add new ipykernel into jupyter
C:\Users\Li\AppData\Local\conda\conda\envs\tf\python.exe -m ipykernel install --prefix=C:\Users\Li\AppData\Local\conda\conda\envs\tf --name tf
D:\LiChen\Anaconda\envs\tf\python -m ipykernel install --user --name tf
To list all conda envs
conda env list
If you want to delete conda environment
conda remove --name tf --all
To test tensoflow installed correctly
import tensorflow as tf #check which version is actually working print(tf.__version__) #1.3.0 hello = tf.constant('Hello, TensorFlow!') sess = tf.Session() print(sess.run(hello))
To change jupyter default path, edit the jupyter notebook config file.
jupyter notebook --generate-config
This writes a file to C:\Users\username\.jupyter\jupyter_notebook_config
Change line 179 #c.NotebookApp.notebook_dir = ''
to c.NotebookApp.notebook_dir = 'your path'
Make sure you use forward slashes in your path.
list all kernels in jupyter
jupyter-kernelspec list
If the "tensorflow" environment is not shown in jupyter notebook
Install kernels
python -m ipykernel install --user
veryfy the install
python -c "import tensorflow as tf; tf.enable_eager_execution(); print(tf.reduce_sum(tf.random_normal([1000, 1000])))"
Optionally, other useful modules to install
pip install keras matplotlib scikit-image tifffile sklearn opencv-python pydicom imageio interpolation mysql-connector SimpleITK medpy pip install git+https://www.github.com/farizrahman4u/keras-contrib.git #python server pip install h5py flask