96lives / gca Goto Github PK

I really appreciate this interesting work, which nicely bridges up two seemingly different areas. I'm currently interested in applying the GCA (not cGCA) model to a new dataset (actually, it's ShapeNet with more categories). It would be great if you can give me some help regarding how to custom a new dataset. I've tried to read the source codes of TransitionShapenetDataset, but I still feel a bit confused.

It seems that the dataset requires embeddings to train the GCA model, and the read context is then used to output embedding_coord, embedding_feat, input_pc, state_coord, respectively. What are the default values for embedding['coord'], embedding['feat'] and embedding['translation']? I think it's important because in the training script, it takes the embedding information as input, and therefore we have to give them some default values.

Besides, I think the current released version may still have some room for small improvements, including but not limited to, the actual dataset folder is named cgca_shapenet rather than shapenet_sdf; config file configs/gca-shapenet-vox=64.yaml provides a None value for the embedding_root: key, which certainly causes problem when trying to compute the self.embedding_root for TransitionShapenetDataset.

Many thanks!!