Open RoraChen opened 2 months ago
Hi Rora,
Thank you for your interest. The N:M pruner of torch2chip is a starting from the GraNet (NeurIPS'21), which is designed for element-wise pruning. But I modified the implementation to enable the prune-and-regrow for the structured fine-grained sparsity.
Specifically, we gradually increase the probability/percentage of the N:M groups throughout the training process.
Hi, I came across the NMPruner class in your repository, particularly interested in its structured fine-grained sparsity methods like _groupdeath and _grp_gradregrow. Could you please let me know which research paper or reference this implementation is based on?Thank you!