Research Article

Differentiable Network Pruning via Polarization of Probabilistic Channelwise Soft Masks

Table 4

Comparison between probabilistic and deterministic mask on the CIFAR-10 dataset. Acc is the accuracy drop of pruned model compared to the baseline model. FLOPs represent the pruning rate of FLOPs.

NetworkMaskBaceline acc (%)Pruned acc (%)Acc (%)FLOPs (%)

ResNet32Probabilistic93.1993.31āˆ’0.1253.27
93.110.0868.60
93.000.1974.40
Deterministic93.1993.120.0753.35
92.930.2669.59
92.810.3872.70

ResNet56Probabilistic93.4493.57āˆ’0.1354.60
93.350.0966.84
93.220.2275.62
Deterministic93.4493.240.2054.30
93.070.3769.59
92.910.5375.16