Skip to content

Latest commit

 

History

History
67 lines (52 loc) · 5.21 KB

details.md

File metadata and controls

67 lines (52 loc) · 5.21 KB

Training Results

VGGNet

cifar100 + e100 + sgd + mslr + ssl(1e-5)

# vggnet16_bn default training
total -  top1 acc: 80.940  top5 acc: 95.550
# vggnet16_bn channel_wise pruning training
total -  top1 acc: 80.790  top5 acc: 95.310
# vggnet16_bn filter_and_channel_wise pruning training
total -  top1 acc: 80.770  top5 acc: 95.580
# vggnet16_bn filter_and_channel_wise pruning training
total -  top1 acc: 80.790  top5 acc: 95.310

ResNet

cifar100 + e100 + sgd + mslr + ssl(1e-5)

# resnet50 default training
total -  top1 acc: 84.070  top5 acc: 96.290
# resnet50 depth_wise pruning training
total -  top1 acc: 84.070  top5 acc: 96.290

Filter_wise

arch prune type prune way pruning ratio actual pruning ratio flops/G model size/MB Flops after pruning Model size after pruning top1 top5
vggnet16_bn filter_wise group_lasso 20% 18.56% 15.51 134.68 7.67 130.88 80.810 95.090
vggnet16_bn filter_wise mean_abs 20% 19.13% 15.51 134.68 9.20 129.51 80.470 94.940
vggnet16_bn filter_wise mean 20% 19.13% 15.51 134.68 10.92 69.38 79.650 94.800
vggnet16_bn filter_wise sum_abs 20% 19.13% 15.51 134.68 6.75 132.22 79.440 94.880
vggnet16_bn filter_wise sum 20% 19.32% 15.51 134.68 13.85 55.88 79.580 95.100
vggnet16_bn filter_wise mean_abs 40% 39.01% 15.51 134.68 6.42 112.47 78.900 94.470
vggnet16_bn filter_wise mean_abs 40% 58.71% 15.51 134.68 4.60 74.06 75.880 93.090

Channel_wise

arch prune type prune way pruning ratio actual pruning ratio flops/G model size/MB Flops after pruning Model size after pruning top1 top5
vggnet16_bn channel_wise group_lasso 20% 18.95% 15.51 134.68 8.21 131.26 80.800 95.070
vggnet16_bn channel_wise mean_abs 20% 19.38% 15.51 134.68 9.57 130.53 80.660 95.280
vggnet16_bn channel_wise mean_abs 40% 38.98% 15.51 134.68 6.15 126.91 79.900 94.910
vggnet16_bn channel_wise mean_abs 60% 59.00% 15.51 134.68 4.11 123.30 78.620 94.480

Filter_and_Channel_wise

arch prune type prune way pruning ratio actual pruning ratio flops/G model size/MB Flops after pruning Model size after pruning top1 top5
vggnet16_bn filter_and_channel_wise group_lasso 20% 13.91% 15.51 134.68 9.56 131.94 80.53 95.320
vggnet16_bn filter_and_channel_wise mean_abs 20% 11.29% 15.51 134.68 11.36 131.84 80.720 95.130
vggnet16_bn filter_and_channel_wise mean_abs 40% 22.98% 15.51 134.68 8.98 128.81 80.040 95.180
vggnet16_bn filter_and_channel_wise mean_abs 60% 35.07% 15.51 134.68 7.45 125.67 79.570 94.920

Depth_wise

arch prune type prune way N flops/G model size/MB Flops after pruning Model size after pruning top1 top5
resnet50 depth_wise group_lasso 1 4.11 23.72 3.89 23.64 83.610 95.970
resnet50 depth_wise mean_abs 1 4.11 23.72 3.89 19.25 83.630 96.090
resnet50 depth_wise mean_abs 2 4.11 23.72 3.67 14.79 82.990 95.710
resnet50 depth_wise mean_abs 4 4.11 23.72 3.23 13.39 82.280 95.400