Thursday, November 23, 2017

TensorFlow - Getting Started with Docker Container and Jupyter Notebook

I'm studying Machine Learning and would like to share some intro experience working with TensorFlow. To get started with TensorFlow you need to install it, easiest way (at least for me) was to run TensorFlow using Docker. Read installation instructions - Installing TensorFlow.

Once TensorFlow Docker image is installed. I suggest to create container in detach mode (--detach=true) and provide port for Jupyter UI. Make sure to provide meaningful name for Docker container:

docker run --detach=true --name RedSamuraiTensorFlowUI -it -p 8888:8888 gcr.io/tensorflow/tensorflow

Make sure to start and stop container using Docker start/stop commands, don't run and create container each time (this way you will loose your work, since new container will be created each time):

docker start RedSamuraiTensorFlowUI (docker stop RedSamuraiTensorFlowUI)

Once container is running in detached mode, you can access logs by executing docker logs command and specifying container name:

docker logs -f RedSamuraiTensorFlowUI

At this point you should see output in Docker container log, copy URL to Jupyter UI with token and paste it to the browser (for example: http://localhost:8888/?token=d0f617a4c719c40ea39a3732447d67fd40ff2028bb335823):


This will give you access to Jupyter UI. Is possible to run TensorFlow Python scripts directly through command line in Docker environment, but is more convenient to do the same through UI:


UI gives option to create new Terminal session:


Terminal allows to run Python code using command line:


Instead of using command line, more convenient is to create new notebook:


Notebook environment allows to type in Python code and execute math calculations. In the example below I multiply two arrays (1x5, 2x6, 3x7, 4x8) in Python code through TensorFlow library. Result is printed through TensorFlow session object right below and prompt for the next command is displayed - very convenient:


Jupyter UI allows to track running notebooks and terminals:


Whatever action you do in Jupyter UI, it can be tracked using log printed in Docker container log. Jupyter UI is client side JS application:


To double check Docker config, I have TensorFlow Docker image:


And Docker container, which can be started/stopped by name (see command listed above), without running new Docker container every time during restart:

No comments: