DISCLAIMER: This guide extends upon the Tensorflow Docker Guide so following their installation instructions and requirements is mandatory!
Assuming you have installed docker and cloned this repository, you can build the container using
docker build -t iannwtf .
Once the container is done building run it with
docker run -dp 8888:8888 -v "$PWD":/tf iannwtf
We use the volume flag -v "$PWD":/tf to allow the container access to the current directory with all the homeworks.
If you have an NVIDIA graphics card and installed the respective toolkit you can run the container with GPU support using
docker run --gpus all -dp 8888:8888 -v "$PWD":/tf iannwtf
In week 08 Tensorboard was introduced which requires additional setup for our container. By default Tensorboard uses port 6006 so we need to forward that port as well.
docker run -dp 8888:8888 -p 6006:6006 -v "$PWD":/tf iannwtf
or
docker run --gpus all -dp 8888:8888 -p 6006:6006 -v "$PWD":/tf iannwtf
for the GPU variant.
You can find the server running at localhost:8888
Get the ID of the currently running container via
docker ps
Then execute
docker exec <your-container id> jupyter notebook list
and copy the token.
You might notice that git will complain when you try to add files to the repo. This is because files created in the container belong to the root user and you will need elevated privilages to work with them. A simple chown can help, but it might get tedious to do that everytime you used the container. So we pass the our current user:usergroup to use in the container.
docker run --gpus all -dp 8888:8888 -p 6006:6006 -v "$PWD":/tf -u $(id -u ${USER}):$(id -g ${USER}) iannwtf
Make sure the required port is exposed as described above. If you plan on using multiple Tensorboard instances in the same notebook, you will need to forward more ports in the container.
When calling the tensorboard magic, make sure to pass the --bind_all flag. Otherwise the notebook server will not be able to connect to the tensorboard server.
The Homework08 folder contains an example notebook taken from tensorflows documentation to illustrate the process.
If you want to set custom password for your server instead of having to grab the randomized token everytime, you can pass a token via an docker environment variable when running the container
-e JUPYTER_TOKEN=my_super_secure_password
Once you shutdown the notebook server, the container will close as well.
Alternatively you can shutdown the container via docker stop <your-container id>