H-Liu1997 / Pytorch-Networks

Pytorch Networks

The Pytorch implementations of famous basic networks. Unscheduled update.

Accuracy Report

Model	Dataset	#Params	Epochs	Test Prec(Paper)	Test Prec(This impl)
ResNet-20	CIFAR-10	0.27M	140,180,200	91.25%	92.02%
ResNet-32	CIFAR-10	0.46M	140,180,200	92.49%	92.61%
ResNet-44	CIFAR-10	0.66M	100,150,200	92.83%	92.46%
ResNet-56	CIFAR-10	0.85M	100,150,200	93.03%	93.22%
ResNet-110	CIFAR-10	1.73M	82,123,164	93.57%	93.40%
ResNet-18(A)	CIFAR-10	11.0M	100,150,200	-	93.54%
ResNet-18(B)	CIFAR-10	11.17M	80,110,120	-	94.51%
ResNet-50	CUB-200	23.92M	26,36,40	-	81.74%
ResNeXt-50	CUB-200	23.39M	26,36,40	-	82.70%
RegNetX-4.0	CUB-200	21.03M	26,36,40	-	84.31%
RegNetY-8.0	CUB-200	37.57M	26,36,40	-	84.40%
RegNetY-8.0	CUB-200	37.57M	cos60+wp	-	84.86%
RegNetY-32.0	CUB-200	142.08M	cos60	-	85.23%
EfficientNet-B2	CUB-200	7.98M	26,36,40	-	82.60%
RegY32+EB2+NeXt50	CUB-200	-	-	-	87.31%

For backbone networks

Most of backbone networks already have pytorch official version(ResNet.etc), my implementations have a little diffience with them because of my programming habits

For other networks

Some networks don't have offical pytorch version for several reasons(author didn't public the code.etc), my implementations are totally original reproductions

Network list

Backbone(chronological order)

Common Type:

• AlexNet 2012
• NIN 2013
• VGG 2014
• Inception/GoogLeNet 2014
• InceptionV2/V3 2015
• InceptionV4/Res 2016
• ResNet 2015
• ResNetV2 2016
• ResNeXt 2016
• DenseNet 2016
• SENet 2017
• MnasNet 2018
• EfficientNet 2019

Light Type:

• ShuffleNet 2017
• MobileNet 2017

Other:

• BasicGNN

Pose Estimation(chronological order)

• OpenPose 2015
• Hourglass 2015
• GNNlikeCNN 2015
• SimpleBaseline 2017
• CPN 2017
• OpenPose 2017
• FPNPoseNet 2017
• CPN+GNN 2018

Detection

• Darknet-19 2016
• Darknet-53 2018

Future Work: Pytorch loss

Will release a new repertory in the future (not in current repertory)

• Focusloss 2016
• labelsmooth