How to use TensorBoard in a Docker container (on Windows)

Question

I have installed tensorflow on windows through docker toolbox. Everything goes well except I can't use tensorboard. The command line shows 'Starting Tensorboard 29 on port 6006. You can navigate to http://localhost:6006/'.However, when I opened this address on my webbrowser, it just can not connect to it. Does anyone know how to solve this problem?

Answer 1

If you're running TensorBoard inside a Docker container, and trying to use a web browser in Windows to view it, you will need to set up port forwarding from the container to your Windows machine. See this answer for a longer discussion about port forwarding for TensorBoard, but you should be able to make progress by using the following command:

docker run -p 0.0.0.0:6006:6006 -it b.gcr.io/tensorflow/tensorflow

However, it may be easier to install TensorFlow directly on Windows, and run TensorBoard there. If you install Python 3.5 for Windows, you can install TensorFlow and TensorBoard by running:

pip install tensorflow

You can then run TensorBoard directly from the command prompt, and you will not need to worry about port forwarding. See the Windows installation instructions for more details.

Answer 2

Maybe you should map your volumes to the folder with the logs and enter with bash well:

docker run -v //c/pathto/tf_logs:/tf_logs  
-p 0.0.0.0:6006:6006 -p 8888:8888 -it b.gcr.io/tensorflow/tensorflow bash
cd ..
tensorboard --logdir tf_logs/

hit the mapping in your browser

http://192.168.99.100:6006

Answer 3

On Windows 10 + WSL2 + Docker using the offical tensorflow/tensorflow:latest-gpu-py3-jupyter image, I had to tell TB to bind to the wildcard address. That is, in the Jupyter notebook, I called:

%tensorboard --logdir logs/ --host 0.0.0.0

After this, I was able to see the embedded dashboard in my notebook.

Answer 4

I'd like to update the answer here, since I just ran into the same problem on Ubuntu 20.04 and the latest-gpu tensorflow docker image (03e706e09b04).

What worked for me was the following docker run:

docker run -p 8888:8888 -p 6006:6006 --rm -v <path_to_logdir>:/logdir tensorflow/tensorflow tensorboard --logdir /logdir --bind_all

The server is then accessible at localhost:6006 as one would expect. The main difference here is, I guess, adding the --bind_all flag to the tensorboard call which exposes the server to external networks, thus allowing the host machine access.