google-deepmind
Code for the next steps
Hi and congratulations for the nice work ! I was wondering whether you intended to also give access to the code that does operations with the latent modulations, i.e classification or generation tasks once the INR is trained. Besides, do you have recommendations to efficiently train a classifier or generator based on the modulations ? Thanks for you help !
Hi, thanks! The TLDR is unfortunately we don't have plans to open source code for classification/generation, but here below is some explanation of why this is: So the code for the classification of voxels used in the paper is just an MLP, but unfortunately we couldn't opensource code related to voxels because of some licensing issues of the ShapeNet dataset that clashes with opensoucing DeepMind code. As for the generation, we relied on an DeepMind internal jax/haiku reproduction of the official DDPM codebase, which was again difficult to opensource from DeepMind. However all the details you need to implement this should be in the appendix of the paper. In terms of recommendations for training classifiers/generators, I'd say use MLPs for classification and DDPM for generartion, and stick to the implementation details described in the paper! I hope this answers your questions.
