details.md

Training Results

VGGNet

# vggnet16_bn default training
total -  top1 acc: 81.060  top5 acc: 95.770
# vggnet16_bn slim 1e-4 pruning training
total -  top1 acc: 81.120  top5 acc: 95.750

VGGNet Pruning

arch	flops/G	model size/MB	slim	predict pruning ratio	true pruning ratio	Flops after pruning	Model size after pruning	top1	top5
vggnet16_bn	15.51	134.68	1e-4	20%	18.75%	8.36	130.31	80.670	95.160
vggnet16_bn	15.51	134.68	1e-4	40%	39.39%	4.52	125.72	79.960	95.030
vggnet16_bn	15.51	134.68	1e-4	60%	58.71%	2.45	120.73	77.620	93.970

ResNet

# resnet50  default training
total -  top1 acc: 83.850  top5 acc: 96.400
# resnet50 slim 1e-5 pruning training
total -  top1 acc: 83.940  top5 acc: 96.340

ResNet Pruning

arch	flops/G	model size/MB	slim	predict pruning ratio	true pruning ratio	Flops after pruning	Model size after pruning	top1	top5
resnet50	4.11	23.72	1e-5	20%	00.09%	4.08	23.67	83.680	96.260
resnet50	4.11	23.72	1e-5	40%	05.99%	2.84	19.44	83.010	95.780
resnet50	4.11	23.72	1e-5	60%	20.09%	1.12	7.42	74.580	92.720

MobileNet_v2

# mobilenet_v2  default training
total -  top1 acc: 80.030  top5 acc: 95.380
# mobilenet_v2 slim 1e-5  pruning training
total -  top1 acc: 80.320  top5 acc: 95.050

MobileNet_v2 Pruning

arch	flops/G	model size/MB	slim	predict pruning ratio	true pruning ratio	Flops after pruning	Model size after pruning	top1	top5
mobilenet_v2	0.313	2.352	1e-5	5%	/	/	/	75.260	92.830
mobilenet_v2	0.313	2.352	1e-5	20%	29.08%	0.224	1.780	83.680	96.260
mobilenet_v2	0.313	2.352	1e-5	40%	54.60%	0.153	1.206	83.010	95.780
mobilenet_v2	0.313	2.352	1e-5	60%	73.03%	0.096	0.728	74.580	92.720