SetNN

This is a program to train and compare different Set Neural Networks on different datasets.

How to use:

Arguments:

--dataset={maximum, max4, min2max2, sum, mean, cardinality, mode, eq2, pointcloud100, pointcloud1000, pointcloud5000}
--net={deepset, set_transformer, pointnet, repset}
Note: pointnet only works on the point cloud variations
--num_epochs={For how many epochs you want to train the model}
--test_freq={After how many epochs should the Network be tested}
--pool={max, mean, sum, fspool, default}

Example:

python run.py --net=deepset --dataset=pointcloud100 --num_epochs=10

How to set up point cloud:

The datasets called pointcloud{100,1000,5000} are based on the ModelNet40 dataset. You will need to download the models (http://modelnet.cs.princeton.edu/) and extract the zip file to Datasets\ModelNet40. You can then run python Datasets\createPointCloud.py (takes a few minutes), which creates Datasets\ModelNet40_cloud.h5. After that you can delete the downloaded files again and use the dataset in the program.

Datasets:

Maximum:

Return the maximum of a set of numbers

Max4:

Return the biggest four elements of a set of numbers

Min2max2:

Return the two smallest and the two biggest elements of a set of numbers

Pointcloud:

Object classification with 40 classes from a point cloud with 100, 1000, 5000 points, by editing the code, the point cloud can be perturbed as well

Cardinality:

Return the number of elements of a set

Mode:

Return the value of the most common element of a multiset

Eq2:

Return 1 if the set is a multiset, 0 otherwise