setup.txt

Install Anaconda (Python 3.6 version)
https://www.anaconda.com/download/#linux

Install CUDA:
http://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html

Install keras and tensorflow-gpu backend (GPU backend for efficient training):
conda update --all
conda install tensorflow-gpu keras

If you ONLY want to do inference (i.e. make predictions) the CPU backend is sufficient:
conda install tensorflow keras


Edit ~./keras/keras.json to specify preferred backend.
Models were created and trained with the theano backend but should work 
fine with the tensorflow backend as well. Unfortunately, saving training 
histories does not seem to work with the tensorflow backend anymore.

theano backend (development of theano is discontinued):
{
    "epsilon": 1e-07,
    "backend": "theano",
    "floatx": "float32",
    "image_dim_ordering": "th",
    "image_data_format": "channels_first"
}

to use the tensorflow-gpu backend:
{
    "floatx": "float32",
    "backend": "tensorflow",
    "image_dim_ordering": "th",
    "image_data_format": "channels_first",
    "epsilon": 1e-07
}

The image data format can also be set directly in Python if you don't want to change the tensorflow configuration files:

    # somewhere after initial imports
    import keras.backend as K
    K.set_image_data_format("channels_first")


run nvidia-smi to check CUDA status:
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 387.26                 Driver Version: 387.26                    |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce GTX 1070    Off  | 00000000:09:00.0  On |                  N/A |
|  0%   45C    P2    64W / 200W |   7883MiB /  8105MiB |     98%      Default |
+-------------------------------+----------------------+----------------------+
                                                                               
+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID   Type   Process name                             Usage      |
|=============================================================================|
|    0      1264      G   /usr/lib/xorg/Xorg                           110MiB |
|    0      4747      C   /home/js/anaconda3/bin/python               7685MiB |
+-----------------------------------------------------------------------------+

Open the Jupyter notebook:
jupyter notebook mnist.ipynb