Visual Mesh Getting Started

How to get started with the Visual Mesh.
Joel Wong GitHub avatar
Updated 18 Oct 2023

This guide compliments the quickstart guide on the main Visual Mesh repo by being geared towards those with little experience of the Linux command line (i.e. bash) or of Machine Learning workflows. If you are already familiar with bash and machine learning, you may prefer to work off the Quickstart guide on the Visual Mesh repo as it is briefer than this guide.

To get started, run this line to check the installation and version of the required programs.

python3 --version && pip3 --version && git --version && docker --version

This should provide a result similar to that listed below, possibly with higher version numbers. Run through the relevant installation guides if this doesn't appear

Python 3.8.5
pip 20.0.2 from /usr/lib/python3/dist-packages/pip (python 3.8)
git version 2.25.1
Docker version 20.10.3, build 48d30b5
Installing Python3

The script to build the dataset is written in Python. When using Ubuntu, the simplest way to install Python is to use the APT package manager. Run sudo apt install python3 to have APT install to the most recent version of Python 3 for your system.

Installing pip

Pip is the default package manager for Python. In systems with both Python2 and Python3, the pip3 specifically invokes the pip tied to the Python3. Because Ubuntu packages pip separately from Python, you may have to install it specifically by running sudo apt install python3-pip.

Installing Git

While you don't strictly need to use git to download the Visual Mesh repo (e.g. you could just download it from github's web interface), git is the recommended way to get and manage the code from the repo. The git command line interface is installed on Ubuntu with APT by running sudo apt install git. If you prefer a GUI for git, try GitKraken.

Installing Docker

Docker is the container system that we use for standardised software environments at NUbots, to install it, check out the documentation, as the version packaged by Ubuntu often lags behind the official version. Be sure to ask on Slack if you've got any questions.

Clone the Visual Mesh Repo

Use git to clone the Visual mesh repo. When using git on the commandline, git will default to cloning the repo into a folder under your current working directory. Start by navigating to the folder where you want to store the repo: for example, if you wanted to store the visual mesh under ~/code/, run:

cd ~
mkdir code
cd code

Once you're in the folder you wanted to be in (e.g. ~/code/), clone the repository with the following command:

git clone

Build the docker image

We specify build steps for docker using a file named Dockerfile. This uses similar syntax to a bash script, but is a format specifically used by docker. Open the dockerfile in a text editor/reader, and take note of the last line that starts with WORKDIR (At the time of writing, the last line of the file). This line tells us that the working directory of the container will be the /workspace directory within the container (not the host). Take note of this as it will be the directory from which the program will run.

Having done a quick check of the dockerfile, we can now tell docker to build the image according to the dockerfile with the following command:

docker build . --pull -t visualmesh:latest
What does this command do?
  • docker is the call to docker
  • build is the main command being sent to docker
  • . is telling docker that the build context is the current directory. Docker will look for a file named Dockerfile in the build context telling it what to build (so in this case, the current directory).
  • --pull tells docker to always attempt to pull a newer version of the image
  • -t visualmesh:latest tells docker that we want to tag the built image as visualmesh:latest. We'll see later that when we tell docker to run a container, that we tell docker which image to use by referencing this tag.

Make a dummy dataset

For the sake of this guide, we'll use a dummy dataset containing the same image repeated over and over again. While this will lead to extreme overfitting, it's an easy way to quickly check that the code is running correctly. Download the example image, mask, and lens file to your downloads folder.

Now, we're going to make a dummy dataset by repeating these files. We can do this using a bash loop. Assuming that we have the dummy data in ~/Downloads/dummy_data, and we're going to put the repeated data in ~/Downloads/repeated_data, run the four commands below:

cd ~/Downloads
mkdir repeated_data
for i in {1..1000}; do cp ./dummy_data/image.jpg ./repeated_data/image${i}.jpg && cp ./dummy_data/mask.png ./repeated_data/mask${i}.png && cp ./dummy_data/lens.yaml ./repeated_data/lens${i}.yaml; done

To ensure that the data was copied sucessfully run the command below which will print all the filenames to the terminal window. If you would like to clean up your display, you can use the clear command.

ls ~/Downloads/repeated_data

Install the prerequisites to make a tfrecord

We want to convert our dataset from image and text files to the tfrecord format that the visual mesh expects, but to do this we need to install some python libraries to allow python to work with tfrecords. Feel free to use something like conda or virtualenv if you're concerned about ending up with conflicting library versions.

pip3 install tensorflow tqdm

Make a set of tfrecord

Having made the repeated dataset and installed the libraries required to work with tfrecords, we can now put it into the tfrecord format the visual mesh wants using the script. By default, the visual mesh will look for the dataset inside a folder named dataset, so we'll navigate back to the visual mesh repo, make the dataset folder, then make the .tfrecord files required using the script:

cd ~/code/VisualMesh
mkdir dataset
./training/ ~/Downloads/repeated_data ./dataset

The script won't complain if you give it an empty folder, so make sure that the created tfrecord files are not empty.

Configure the training

The visual mesh looks for a config.yaml file in the output folder to set the training parameters. An example file example_net.yaml can be found in the root of the repo. We'll create an output folder and copy the example parameters into it:

cd ~/code/VisualMesh
mkdir output
cp example_net.yaml ./output/config.yaml

With that, we're almost ready to train the network. To save time, change the configuration to run less training epochs (500 epochs can easily take 30 hours to complete on CPU). Open the config file in a text editor, (e.g. nano output/config.yaml) and change the line that says epochs: 500 to say epochs: 5. Also note that the paths for the training, testing and validation tfrecords are set in the config file. If you have accidentally started training 500 epochs hit ctrl+c to keyboard interrupt training, if that doesn't work, open another bash terminal, type xkill then click on the training terminal window

Train the network

The command to train the network is as below. It's a long command, but it's straightforward if we break it down into the individual parts:

cd ~/code/VisualMesh
docker run --gpus all -u $(id -u):$(id -g) -it --rm --volume $(pwd):/workspace visualmesh:latest ./ train output
What does this command do?
  • docker is the call to docker
  • run is the main command being sent to docker
  • --gpus all tells docker to use the gpus from your system (remove --gpus all if you're running on a system without a GPU)
  • -it tells docker to run in interactive mode with a terminal
  • --rm tells docker to automatically clean up the temporary files associated with this container
  • --volume $(pwd):/workspace tells docker to mount the current working directory as /workspace inside the container
  • visualmesh:latest tells docker which image we want to run (note that we built this image and gave it this name in an earlier step)
  • ./ train output is the command to run inside the container, so we're passing the command train with the path output to the script. Note that the filesystem inside the container is different to the filesystem outside of the container, but that due to the --volume flag, the current working directory is mounted at /workspace, so in this example, the output folder ends up being the one in the code directory.

Test the network

Testing the network is a very similar command to training, we just replace "train" with "test":

cd ~/code/VisualMesh
docker run --gpus all -u $(id -u):$(id -g) -it --rm --volume $(pwd):/workspace visualmesh:latest ./ test output

Once the test has finished running, you can inspect the output in output/test/metrics.

Using Tensorboard to inspect the process

Tensorboard is a tool that provides a dashboard for visualising of the training process that runs in the browser. For tensorflow, we just pass the output directory to tensorboard: (Note: You may have to install TensorBoard separately if your TensorFlow install didn't bundle it.)

tensorboard --logdir output

You should see something like the following:

Serving TensorBoard on localhost; to expose to the network, use a proxy or pass --bind_all
TensorBoard 2.4.1 at http://localhost:6006/ (Press CTRL+C to quit)

Go to the url specified on your terminal to see the tensorboard dashboard.

NUbots acknowledges the traditional custodians of the lands within our footprint areas: Awabakal, Darkinjung, Biripai, Worimi, Wonnarua, and Eora Nations. We acknowledge that our laboratory is situated on unceded Pambalong land. We pay respect to the wisdom of our Elders past and present.
Copyright © 2023 NUbots - CC-BY-4.0
Deploys by Netlify