Table 2. Experimental results of different SE-ResNet architectures on Cifar.
Method | Module | Cifar-10 | Cifar-100 | ||||
---|---|---|---|---|---|---|---|
#params | #model size | Accuracy | #params | #model size | Accuracy | ||
SE-ResNet-18 | basic | 11, 272K | 90.4Mb | 94.85 ± 0.14 | 11, 312K | 90.8Mb | 75.86 ± 0.22 |
SE-ResNet-26 | bottleneck | 15, 383K | 123.4Mb | 93.90 ± 0.18 | 15, 567K | 124.8Mb | 75.40 ± 0.27 |
SE-ResNet-34 | small | 8, 145K | 65.6Mb | 94.79 ± 0.17 | 8, 191K | 65.9Mb | 75.81 ± 0.13 |
SE-ResNet-50 | bottleneck | 26, 100K | 209.4Mb | 94.67 ± 0.09 | 26, 285K | 210.8Mb | 78.02 ± 0.25 |
SE-ResNet-66 | small | 14, 986K | 120.7Mb | 95.26 ± 0.18 | 15, 033K | 121.1Mb | 77.02 ± 0.20 |
BHCNet-3 | small | 198K | 2.1Mb | 92.30 ± 0.14 | 204K | 2.2Mb | 68.36 ± 0.23 |
BHCNet-6 | small | 401K | 4.2Mb | 93.18 ± 0.14 | 407K | 4.2Mb | 70.33 ± 0.25 |