In this project I built a simple Graph Neural Network model to attack the problem of 3D data classification. The datasets employed in this project are graph-representations of triangular meshes of facial data, obtained with structured-light based 3D system.

The repository is structured as following:

• config.py : Python dictionary containing the hyperparameters for the training procedure.
• data.py: Input data pipeline (loading, pre-processing, ...)
• main.py: Script that handles the training phase
• models.py: Contains the code for a few GNN models
• test.py: Script that handles the testing phase
• graph_data/: Directory containing the datasets
• runs/: Directory containing the experiments

I cannot share my data but, if you want to build a new dataset, follow this procedure: assuming that you have a set of 3D meshes in .obj format you have to group them by class into their relative sub-directories. Let' say you have a dataset like this:

./data/my_dataset/

where every subdirectory of my_dataset/ represents an identity, which contains a number of meshes, e.g.

./data/my_dataset/id000/id000_0.obj
./data/my_dataset/id000/id000_1.obj
...

You just have to run ./data/to_graph.py and a new directory will be created under ./graph_data/, containing the freshly build dataset of graphs derived by their relative meshes.

This is an example of the data I've dealt with

You can look at my slides to gain insights about the (simple) model(s) I've coded.

Once you're set up with the data you just have to set your hyperparameters under ./config.py and run python main.py.

A Tensorboard log containing the hyperparameters, learning curves and histograms will be logged under ./runs/

Every 100 epochs both the model state and optimizer state dictionaries will be saved inside the same sub-directory.

If you want to test a model you just need to pass the relative ./runs/ sub-directory to the ./test.py script. Let's say you want to test the model logged under ./runs/exp_000/, then you just need to run

python test.py --exp_dir ./runs/exp000

Running

conda env create -f torchgeom.yml

will create an environment containing all the packages needed for this project (CUDA 10.2, cudnn 7.6).

The main dependencies are

python 3.7.7
torch 1.5.1
torch_geometric 1.6.0
tensorboard 2.2.2
scipy 1.5.1
numpy 1.18.5
scikit-learn 0.23.1
trimesh 3.8.8
tqdm 4.47.0