Sparse dense matmul with variable values

[jocampo2]

Describe the feature and the current behavior/state. Currently tensorflow doesn't support a sparse dense matmul with variables as the values of the sparse tensor (Though there are some hacks). i.e y=Wx where W is a sparse matrix and is learnable

I have an implementation see this google colab file, using either autodiff or a custom gradient(which is more efficient though needs a precompute), and is tested against autodiff. Also works with GPU. Can we add the SparseLayer with custom gradients as a feature in tensorflow_addons?

• Are you willing to contribute it: yes
• Are you willing to maintain it going forward?: yes
• Is there a relevant academic paper?: https://arxiv.org/abs/2209.13603 (In the appendix)
• Does the relavent academic paper exceed 50 citations?: not yet, it was put up last month

Which API type would this fall under (layer, metric, optimizer, etc.) layer

Who will benefit with this feature? People have been asking for this on github and stackoverflow frequently. This is a sparse matrix multiplication layer which is widely used, having learnable values should prove useful for many.

[seanpmorgan]

TensorFlow Addons is transitioning to a minimal maintenance and release mode. New features will not be added to this repository. For more information, please see our public messaging on this decision: TensorFlow Addons Wind Down

Please consider sending feature requests / contributions to other repositories in the TF community with a similar charters to TFA: Keras Keras-CV Keras-NLP