All major Deep learning frameworks make heavy use of GEMM operations through BLAS libraries. These libraries are very efficient in computing dense matrix multiplications. But, when, we start removing weights by pruning, it creates sparsity in matrices. Even after 95% fewer calculations, sparse matrices takes more time for matrix multiplication. So, there is no very efficient way to do this yet. But there are many ongoing projects for efficient sparse matrix multiplication.
If you still want to prune network but in efficient way, look for structured pruning.
I have a question regarding the pruning rates for CNN regarding CIFAR-10 dataset. In the paper, it’s mentioned in “Figure 2: Architectures testing in this paper”. It’s mentioned that for Conv-2 CNN, the pruning rate is 10% for conv layers and 20% for fully connected layers.
I have the following questions:
How does this relate to each round pruning (p^(1/n))% of the surviving weights? I thought this value of ‘p’ was a user input
According to the Figure 2 table, each round prunes 10% conv layer’s surviving weights and 20% for fully connected layer’s surviving weights. If this be true, then the ‘p’ from above point doesn’t make sense.