Run Keras with CNTK backend from CentOS

Keras now supports CNTK. https://docs.microsoft.com/en-us/cognitive-toolkit/Using-CNTK-with-Keras

I installed it and tried it, so I will introduce the procedure.

environment

CentOS 7.3 and Python 3.5, CPU only. Looking at the Microsoft site, only Ubuntu is written about how to build a Linux environment, but Centos 7.3 also worked fine. (Boyaki: Deep learning has become the strongest in Ubuntu.)

CNTK installation

The official procedure is below. https://docs.microsoft.com/en-us/cognitive-toolkit/setup-linux-python?tabs=cntkpy21

Here, we will show you how to install it on Centos 7.3.

First, install Anaconda3. Select the version from the following and install it. https://www.continuum.io/downloads

If you want to install from the command line, you can install it as follows.

wget https://repo.continuum.io/archive/Anaconda3-4.3.1-Linux-x86_64.sh
bash Anaconda3-4.3.1-Linux-x86_64.sh -b -p /opt/anaconda3
echo 'export PATH="/opt/anaconda3/bin:$PATH"' >> /etc/profile
source /etc/profile

CNTK requires OpenMPI. Install it on CentOS7.3 with the following and set environment variables as well.

yum -y install openmpi openmpi-devel
export PATH=/usr/lib64/openmpi/bin:$PATH
export LD_LIBRARY_PATH=/usr/lib64/openmpi/lib:$LD_LIBRARY_PATH

It's finally time to install CNTK. Specify the URL that suits your environment with pip install. For Linux, CPUonly, Python3.5, it will be as follows.

pip install https://cntk.ai/PythonWheel/CPU-Only/cntk-2.1-cp35-cp35m-linux_x86_64.whl

Below is a list of URLs to specify. https://docs.microsoft.com/en-us/cognitive-toolkit/setup-linux-python?tabs=cntkpy21

If the installation is successful, you can see the CNTK version below.

python -c "import cntk; print(cntk.__version__)"

Sample programs and tutorials are available below.

python -m cntk.sample_installer

Use with Keras

You need to change the backend to use CNTK with Keras. https://keras.io/ja/backend/

The backend is changed in /User'sHOME/.keras/keras.json, but at this stage the .keras directory is not yet there. You need to call Keras from Python once and create it.

python -c "import keras"

I think TensorFlow is the backend by default. Edit /User'sHOME/.keras/keras.json.

#Before editing
{
    "floatx": "float32",
    "image_data_format": "channels_last",
    "epsilon": 1e-07,
    "backend": "cntk"
}

#After editing
{
    "floatx": "float32",
    "image_data_format": "channels_last",
    "epsilon": 1e-07,
    "backend": "cntk"
}

You can now use Keras with the CNTK backend.

If you do import Keras on Jupyter Notebook, you can see that the backend is CNTK.

For the time being, MNIST

For the time being, I tried running the MNIST MLP sample. https://github.com/fchollet/keras/blob/master/examples/mnist_mlp.py

It is such a network.

The program works as it is from the existing Keras. No editing required.

Overview and results --Training data: 60,000 28x28 images --Test data: 10000 28x28 images --Batch size: 128 --Epoch: 20 --CNTK backend training time: 314 seconds --CNTK backend test results: Loss 0.106430094829, Accuracy 0.9835

Digression

There are a lot of deep learning frameworks, but DL4J makes a comparison. It is surprisingly well organized. https://deeplearning4j.org/ja/compare-dl4j-torch7-pylearn