We’ve designed a pruning methodology embedded to the
MASE system with pruning and retraining that able to be
called from the command line, and then introduce a set
of experiments to explore the effects of different pruning
methods, including tensor-element-wise, tensor-channel-wise,
layer-element-wise, layer-channel-wise, and global pruning
with l1 and l2 norm to decide the threshold respectively.
Meanwhile, we also introduce a retraining process after
pruning is completed for recovering the performance of
the pruned model. We use the accuracy retention rate and
parameters compression rate as the main evaluation metrics to
measure the robustness of the pruning methods, and change
one pruning/retraining configure each time and keep other
variables consistent
We’ve designed a pruning methodology embedded to the MASE system with pruning and retraining that able to be called from the command line, and then introduce a set of experiments to explore the effects of different pruning methods, including tensor-element-wise, tensor-channel-wise, layer-element-wise, layer-channel-wise, and global pruning with l1 and l2 norm to decide the threshold respectively. Meanwhile, we also introduce a retraining process after pruning is completed for recovering the performance of the pruned model. We use the accuracy retention rate and parameters compression rate as the main evaluation metrics to measure the robustness of the pruning methods, and change one pruning/retraining configure each time and keep other variables consistent