Concept软剪枝方法
Soft Pruning
A pruning method that multiplies each functional unit by a scaling coefficient between 0 and 1. The model remains end-to-end trainable and gradients flow smoothly through the scaling factors. When a coefficient approaches 0, the corresponding unit is effectively removed, making soft pruning a differentiable relaxation of hard structural pruning.
Timeline (last 90 days)
Oct 15
The optimization objective is to simultaneously minimize the model's loss under soft-pruned states across multiple sparsity targets, enabling the model to learn at once to degrade smoothly under multiple compression ratios.
Unverified50%