Running TensorBoard in a Dockerfile

Tenosrboard is one of the most useful tools you can use when training deep neural networks. If you run your TensorFlow experiments from a Dockerfile it’s also possible to run Tensorboard in the same dockerfile. Things to keep in mind are:

  • Make sure your port 6006 is open. If you are running a dockerfile directly, either with a locally defined dockerfile or using an image from the DockerHub, you have to pass the argument -p 6006:6006 to the command line. You will normally pass -p 8888:888 already, so simply extend that command so you get -p 8888:888 -p 6006:6006.
  • If you are using docker-compose you should also add port 6006 to the ports in your docker-compose.yml file.
  • You don’t have to make a separate Dockerfile for TensorBoard. I used to make this mistake for many years, where I defined a separate Dockerfile for Tensorboard. Turns out you can start tensorboard from a Jupyter notebook or command line later. To do this in a jupyter notebook you have to use the load_ext tensorboard command in a cell to start using it. Then visualise TensorBoard in a Jupyter notebook cell using the %tensorboard --logdir logs --bind_all command. You only have to execute this command once. Every next time you use this command you will get the Reusing TensorBoard on port 6006 message, which will just show your current existing tensorboard session. For me killing tensorboard doesn’t work, and it required me to restart the whole docker container.
  • If you get 127.0.0.1 didn’t send any data. you likely forgot the --bind_all command at some point in time. As I mention at my previous point: I simply had to restart my docker container by killing it. Even using the python tools by importing tensorboard (using from tensorboard import notebook) I could not make it work. Normally you might be able to run it with a different port, but then you would have to open that one first in your docker container.

TensorBoard example

TensorFlow has a great example on how to use their tool, which is described here: https://www.tensorflow.org/tensorboard/tensorboard_in_notebooks. I decided not to copy paste their example code so you won’t have to deal with an outdated piece of code in a few years.