backpropagation for sparse semi-structured #126420
Labels
module: sparse
Related to torch.sparse
triaged
This issue has been looked at a team member, and triaged and prioritized into an appropriate module
馃殌 The feature, motivation and pitch
I am trying to train ultra-sparse linear layers with as low as 0.1% of nonzero elements. Forward propagation is successful, however propagating the loss backward fails.
I understand that this data structure is a work in progress and also that the main use-case is training dense linear modules but replace them with sparse ones for a skinnier evaluation model.
Nonetheless I see this as a logical next step in research an implementation. The current next best thing is masking dense tensors, which is a significant decrease in training performance and limitation in model size.
Here's a snippet of code for what I am trying to achieve but fail:
This results in the following error:
Alternatives
No response
Additional context
No response
cc @alexsamardzic @nikitaved @pearu @cpuhrsch @amjames @bhosmer @jcaip
The text was updated successfully, but these errors were encountered: