Model Conversion and Compression Statistics

Conversion

name	net	baseline	pretrained	retrained
AlexNet	PyTorch	56.55%/79.09%	55.86%/78.60%	56.47%/79.05%
SqueezeNet v1.0	PyTorch	58.10%/80.42%	56.45%/79.24%	57.75%/80.36% [3]
SqueezeNet v1.1	PyTorch	58.19%/80.62%	56.99%/79.78%	57.90%/80.28% [3]
Inception V3	TF-Slim	78.0%/93.9%	77.98%/93.94%	Not necessary
MobileNet V1	TF-Slim	70.7%/89.5%	70.53%/89.57%, 70.75%/89.53% [1]	Not necessary
VGG-16	TF-Slim	71.5%/89.8%	65.69%/86.61% [2]	70.82%/89.93%
VGG-16bn	Pytorch	73.37%/91.5%	72.56%/91.09%	Not necessary
ResNet18	Pytorch	69.76%/89.08%	68.98%/88.68%	Not necessary

[1] With fill=True and fill=False respectively.

[2] Result from #8.

[3] Retrained for around 25 epochs starting from the pretrained models.

Our Results

Network	Size	CR	top1 error (%)	top5 error (%)
LeNet5	1.11%	90.1x	0.7%	-
CifarNet	4.16%	24.0x	18.81%	1.28%
CifarNet-baseline	-	-	18.35%	1.35%
AlexNet	4.58%	21.83x	45.41%	21.6%
AlexNet-baseline	-	-	44.14%	21.4%
SqueezeNet10	48.25%	2.07x	43.82%	20.87%
SqueezeNet10-baseline	-	-	43.55%	20.76%
SqueezeNet11	46.35%	2.16x	43.74%	20.58%
SqueezeNet11-baseline	-	-	43.01%	20.22%
Mobilenet	-	-	29.47%	10.05%
Mobilenet-baseline	-	-	-	-
Resnet18	19.50%	5.13x	30.74%	10.87%
Resnet18-baseline	-	-	31.02%	11.32%

Deep Compression

name	baseline	baseline accuracy	density	accuracy
AlexNet	11%	57.22%/80.27% [1]
VGG-16	7.7%	68.83%/89.09% [1]

[1]: After all compression pipeline stages.

Network	Size	CR	top1 (%)	top5 (%)
LeNet5	-	108x	0.91%	-
AlexNet	-	17.7x	43.09%	-