thanks for reading.
I have been attempting to train a simple VAE on very sparse 2D and 3D data. So far I have been training using dense tensors which - I think - is resulting in horrible training due to the submanifold dialation (so coined in this paper: https://ui.adsabs.harvard.edu/abs/2017arXiv170601307G/abstract). I have been really struggling to implement this however, and keep getting errors. I lack an understanding of what Tensorflow can actually achieve using sparse tensors - how supported are they by tf.keras? How much funtionality actually is there for training on sparse tensors? I am familiar with a few libraries in Pytorch that are designed for sparse input; Minkowski Engine by Nvidia (https://github.com/NVIDIA/MinkowskiEngine) seems to be one such. It would take me a fair bit of work to transition to Pytorch however. I have struggled to find resources detailing the use of sparse tensors in tensorflow - every paper I read seems to brush over the sparsity problem or address it by pre-processing their data.
In summary, could someone please point me to a resource explaining something akin to the implementation of a simple model with sparse tensor input. Also, if anybody has experience, could you explain to me the capabilities of sparse tensors in Tensorflow as pertains to training models.
Thank you very much.