Table 5.

Configurations for auto compress pruner.

ADMM-Based Structured Pruning	Initial penalty parameter ρ = 1 × 10−4
	Training epochs for the optimization of ADMMPruner: 5
	Number of iterations of ADMM Pruner: 5
	Optimizer: gradient centralization (GC)
Simulated Annealing (SA)	Cooling factor: η = 0.9
	Initial perturbation magnitude to the sparsities: 0.35
	Start temperature of the simulated annealing process: 100
	Stop temperature of the simulated annealing process: 20