Why?

The original BRepNet architecture described in BRepNet: A topological message passing system for solid models used only very simple input features like curve and surface types, face areas, edges lengths and B-Rep data structure flags. A better way to get curve and surface data into a neural network is described in UV-Net: Learning from Boundary Representations. This PR makes use of these preferred input features in BRepNet.

What?

The BRepNet data extraction pipeline has been enhanced to add UV-grids for faces and edges
The BRepNet dataloader now passes this information through to the network
The BRepNet model now optionally uses the face and edge grid information
The pre-trained model has been updated to operate with the UV grid features
Embeddings for each face can now be saved from the model
A jupyter notebook shows how these embeddings can be used to search for similar faces and edges
Dropout is added, yielding improved results
Reproducing the results in the paper can still be achieved using the train/reproduce_paper_results.sh script to set the hyper-parameters