Using MeshGraphNets on own data?

We have training sets with time steps of triangular meshes of the surface of volumes that deform over time and are hoping that your MeshGraphNets approach for Lagrangian systems may be a big win over our current finite element approach which takes days to generate predictions. The meshing is dynamic but it seems that you do not require the same nodes to exist in each timestep.

I have had some success in trying your code. I know some Python but not TensorFlow. I managed to get the older Python 3.6 and numpy 1.18.5 installed so I could run your code with TensorFlow 1, though your `run.sh` seems to do things then delete it all without showing anything, even if I also install `tk` into the virtual environment. TensorFlow record datasets are new to me but, even if I learn about that, I don't know if it's clear what the input and output formats of your software are within those datasets. At least the code itself seems somewhat clear, much appreciated.

I guess, I'm faced with a dilemma with which I'd appreciate your thoughts. We have limited time and funds and could:

* If you think that your code would be a good fit for our surface deformation prediction problem, we could try to understand it all better and apply it directly and let you know how that goes. We'd certainly appreciate any clues along the way in terms of input or output formats, what kind of model to apply (cloth?), if we must do something special to account for the dynamic meshing, etc. I don't know if any of your demo datasets are already very similar, you seem to provide more than are described in the paper.

* I think that your approach may be great, especially as we can get meshes but, depending on the difficulty of our applying your software, it's also possible that it might be easiest for us to not try to use it directly, instead coding something based on your ideas and reusing specific parameters (network size and depth, etc.) wherever you have left us some low-hanging hard-won fruit. We'd possibly use PyTorch instead as we are already using it in trying a voxel/image-based approach in applying U-Nets to the same problem.

Any advice or pointers are appreciated. I should note that this issue may be related to #199 and #204.