Question about fowward of GCNMLPDecoder in DMTET

[colorful-liyu]

Dear Kaolin group, 

Thank you for your great work of Kaolin !

I am a PhD student in China, and want to follow Deep Marching Tetrahedra: a Hybrid Representation for High-Resolution 3D Shape Synthesis(DMTet). I'm trying to modify the codebase of DefTet to achieve the TSDF supervision and volume subdivision claimed in DMTet. 

However, a problem confused me for a while.  The DMTet paper claimed:

'We first identify surface tetrahedra Tsurf  based on the current s(v) value. We then build a graph G = (Vsurf; Esurf), where Vsurf; Esurf correspond to the vertices and edges in Tsurf.'  

I've read DefTet and Kaolin carefully, including kaolin.ops.conversions.marching_tetrahedra(). But different meshes have different numbers of surface tetrahedra Tsurf. I was wondering how the meshes with different Tsurf can be synthesized in one batch, and how the different graphs G can be used in one forward process. 

So I sincerely hope to get help from you. Thanks for your great work again!