CNN-models-comparison

The goal of this repository is:

• To keep on track of state-of-the-art (SoTA) and new CNN architectures
• To see the comparison of famous CNN models at a glance (accuracy, parameters, speed ...)
• To access their research papers and implementations on different frameworks

(This repository will be updated regularly.)

Comparison Table

CNN model comparison table on the ImageNet classification results, reference paper and implementations.

Model	Detail	Input size	Parameters	Mult-Adds	FLOPS	Depth	Top-1 Acc	Top-5 Acc	Paper	TF	Keras	Pytorch	Caffe	Torch	MXNet
AmoebaNet-B	(N=6, F=228)	331x331⁵	155.3M⁵	41.1B⁵			83.10⁵	96.30⁵	Paper	TF	-	-	-	-	-
PNASNet-5_Large_331	(N=4, F=216)	331x331⁵	86.1M⁵	25.0B⁵			82.90⁵	96.20⁵	Paper	TF	-	-	-	-	-
AmoebaNet-B	(N=6, F=190)	331x331⁵	86.7M⁵	23.1B⁵			82.80⁵	96.10⁵	Paper	TF	-	-	-	-	-
SENet-154		320x320⁵	145.8M⁵	42.3B⁵			82.70⁵	96.20⁵	Paper	TF	Keras	Pytorch	Caffe	-	-
NASNet-A_Large_331	(N=6, F=168)	331x331⁵	88.9M⁵	23.8B⁵			82.70⁵	96.20⁵	Paper	TF	Keras	Pytorch	-	-	-
AmoebaNet-B	(N=6, F=190)	331x331⁵	84.0M⁵	22.3B⁵			82.30⁵	96.10⁵	Paper	TF	-	-	-	-	-
Dual-Path-Net-131		320x320⁵	79.5M⁵	32.0B⁵			81.50⁵	95.80⁵	Paper	-	Keras	Pytorch	Caffe	-	MXNet
PolyNet		331x331⁵	92M⁵	34.7B⁵			81.30⁵	95.80⁵	Paper	-	-	Pytorch	Caffe	-	-
ResNeXt-101	(64x4d)	320x320⁵	83.6M⁵	31.5B⁵			80.90⁵	95.60⁵	Paper	TF	Keras	Pytorch	Caffe	Torch	-
PyramidNet-200	α=300	320x320⁷	116.4M⁷				80.80⁷	95.30⁷	Paper	TF	-	Pytorch	Caffe	Torch	-
PyramidNet-200	α=300	320x320⁷	62.1M⁷				80.50⁷	95.20⁷	Paper	TF	-	Pytorch	Caffe	Torch	-
Inception-ResNet-v2		299x299¹	55.8M²		11.75G⁴	572²	80.40¹	95.30¹	Paper	TF	Keras	-	Caffe	-	-
Inception-ResNet-v2+SE		299X299⁴			11.76G⁴		80.20⁴	95.21⁴	Paper	TF	-	Pytorch	Caffe	-	-
Inception V4		299x299¹	46M¹				80.20¹	95.20¹	Paper	TF	Keras	Pytorch	-	-	-
ResNet V2 200		320x320	64.7M		15G		79.90¹	95.20¹	Paper	TF	Keras	-	Caffe	-	-
bl-ResNet-152@256	α=2, β=4	256x256⁸	57.36M⁸		6.58G⁸		79.66⁸	-	Paper	-	-	-	-	-	-
bl-ResNeXt-101@256	α=2, β=4	256x256⁸	41.51M⁸		5.33G⁸		79.59⁸	-	Paper	-	-	-	-	-	-
bl-ResNet-152	α=2, β=2	224x224⁸	59.81M⁸		5.64G⁸		79.04⁸	-	Paper	-	-	-	-	-	-
Xception		299x299²	23M²			126²	79.00²	94.50²	Paper	TF	Keras	Pytorch	Caffe	-	-
bl-ResNet-101@256	α=2, β=4	256x256⁸	41.85M⁸		5.08G⁸		78.96⁸	-	Paper	-	-	-	-	-	-
ResNeXt-101+CBAM	(32x4d)	224x224³	49M³		7.519G³		78.93³	94.41³	Paper	TF	Keras	-	-	-	-
bl-ResNet-152	α=2, β=4	224x224⁸	57.36M⁸		5.04G⁸		78.84⁸	-	Paper	-	-	-	-	-	-
ResNeXt-101+SE	(32x4d)	224x224³	49M³		7.512G³		78.83³	94.34³	Paper	TF	-	-	Caffe	-	-
bl-ResNeXt-101	α=2, β=4	224x224⁸	41.51M⁸		4.08G⁸		78.80⁸	-	Paper	-	-	-	-	-	-
bl-ResNeXt-50@256	α=2, β=4	256x256⁸	26.19M⁸		3.95G⁸		78.77⁸	-	Paper	-	-	-	-	-	-
bl-ResNet-101	α=2, β=2	224x224⁸	43.39M⁸		4.27G⁸		78.60⁸	-	Paper	-	-	-	-	-	-
ResNet101+CBAM		224x224³	49M³		7.581G³		78.49³	94.31³	Paper	TF	Keras	-	-	-	-
ResNeXt-101	(32x4d)	224x224³	44.2M³		7.508G³		78.46³	94.25³	Paper	TF	Keras	Pytorch	Caffe	Torch	-
bl-ResNet-101	α=2, β=4	224x224⁸	41.85M⁸		3.89G⁸		78.20⁸	-	Paper	-	-	-	-	-	-
ResNeXt50+SE	(32x4d)	224x224³	27.6M³		3.771G³		78.09³	93.96³	Paper	TF	-	-	Caffe	-	-
ResNeXt50+CBAM	(32x4d)	224x224³	27.6M³		3.774G³		78.08³	94.09³	Paper	TF	Keras	-	-	-	-
bl-ResNeXt-50	α=2, β=4	224x224⁸	26.19M⁸		3.03G⁸		78.00⁸	-	Paper	-	-	-	-	-	-
Inception V3		299x299¹	23.8M²			159²	78.00¹	93.90¹	Paper	TF	Keras	Pytorch	Caffe	-	-
ResNet V2 152		299x299¹					77.80¹	94.10¹	Paper	TF	Keras	-	Caffe	-	-
ResNet101+SE		224x224³	49M³		7.575G³		77.65³	93.81³	Paper	TF	-	Pytorch	Caffe	-	-
ResNet50+CBAM		224x224³	28M³		3.864G³		77.34³	93.69³	Paper	TF	Keras	-	-	-	-
bl-ResNet-50	α=2, β=4	224x224⁸	26.69M⁸		2.85G⁸		77.31⁸	-	Paper	-	-	-	-	-	-
ResNeXt50	(32x4d)	224x224³	25M³		3.768G³		77.15³	94.25³	Paper	TF	Keras	Pytorch	Caffe	Torch	-
DenseNet201		224x224²	20M²			201²	77.00²	93.30²	Paper	TF	Keras	Pytorch	Caffe	-	-
ResNet V2 101		299x299¹					77.00¹	93.70¹	Paper	TF	Keras	-	Caffe	-	-
ResNet50+SE		224x224³	28M³		3.86G³		76.86³	93.30³	Paper	TF	-	Pytorch	Caffe	-	-
bl-ResNet-50	α=4, β=4	224x224⁸	26.24M⁸		2.48G⁸		76.85⁸	-	Paper	-	-	-	-	-	-
ResNet V1 152		224x224¹	60M		11.3G⁴	517	76.80¹	93.20¹	Paper	TF	Keras	Pytorch	Caffe	Torch	-
ResNet V1 101		224x224¹	45M³		7.57G³		76.40¹	92.90¹	Paper	TF	Keras	Pytorch	Caffe	Torch	-
MnasNet-92+SE		224x224⁹	5.1M⁹	391M⁹			76.13⁹	92.85⁹	Paper	TF	Keras	Pytorch	Caffe	-	-
DenseNet169		224x224²	14M²			169²	75.90²	92.80²	Paper	TF	Keras	Pytorch	Caffe	-	-
BN-Inception+SE		224x224⁴			2.04G⁴		75.77⁴	92.86⁴	Paper	TF	-	-	Caffe	-	-
ResNet V2 50		299x299¹					75.60¹	92.80¹	Paper	TF	Keras	-	Caffe	-	-
MnasNet+SE		224x224⁹	4.7M⁹	319M⁹			75.42⁹	92.51⁹	Paper	TF	Keras	Pytorch	Caffe	-	-
ResNet V1 50		224x224¹	25.6M²		3.858G³	168²	75.20¹	92.20¹	Paper	TF	Keras	Pytorch	Caffe	Torch	-
WideResNet18+CBAM	widen=2.0	224x224³	45.97M³		6.697G³		75.16³	92.37³	Paper	-	-	-	-	-	-
WideResNet18+SE	widen=2.0	224x224³	45.97M³		6.696G³		75.07³	92.35³	Paper	-	-	-	Caffe	-	-
MobileNet_v2	α=1.4	224x224¹	6M				74.90¹	92.50¹	Paper	TF	Keras	Pytorch	Caffe	-	-
MnasNet-92		224x224⁹	4.4M⁹	388M⁹			74.79⁹	92.05⁹	Paper	TF	Keras	Pytorch	Caffe	-	-
MnasNet-65+SE		224x224⁹	4.1M⁹	272M⁹			74.62⁹	91.93⁹	Paper	TF	Keras	Pytorch	Caffe	-	-
BN-Inception		224x224⁴			2.03G⁴		74.62⁴	92.21⁴	Paper	-	-	-	Caffe	-	-
AmoebaNet-A		224x224⁹	5.1M⁹	555M⁹			74.50⁹	92.00⁹	Paper	TF	-	-	-	-	-
DenseNet121		224x224²	8M²			121²	74.50²	91.80²	Paper	TF	Keras	Pytorch	Caffe	-	-
WideResNet18	widen=2.0	224x224³	45.62M³		6.696G³		74.37³	91.80³	Paper	TF	Keras	Pytorch	Caffe	-	-
PNASNet-5_Mobile		224x224⁹	5.1M⁹	588M⁹			74.20⁹	91.90⁹	Paper	TF	-	-	-	-	-
ResNet34+CBAM		224x224³	22M³		3.665G³		74.01³	91.76³	Paper	TF	Keras	-	-	-	-
MnasNet		224x224⁹	4.2M⁹	317M⁹			74.00⁹	91.78⁹	Paper	TF	Keras	Pytorch	Caffe	-	-
NASNet-A_Mobile_224		224x224¹					74.00¹	91.60¹	Paper	TF	Keras	Pytorch	-	-	-
WideResNet18+CBAM	widen=1.5	224x224³	26.08M³		3.868G³		73.90³	91.57³	Paper	-	-	-	-	-	-
Inception V2		224x224¹					73.90¹	91.80¹	Paper	TF	-	Pytorch	-	-	-
ResNet34+SE		224x224³	22M³		3.664G³		73.87³	91.65³	Paper	TF	-	Pytorch	Caffe	-	-
CondenseNet	G=C=8	224x224⁹	4.8M⁹	529M⁹			73.80⁹	91.70⁹	Paper	TF	-	Pytorch	-	-	-
WideResNet18+SE	widen=1.5	224x224³	26.07M³		3.867G³		73.79³	91.53³	Paper	-	-	-	Caffe	-	-
ShuffleNet	x2	224x224⁹	5.4M⁹	524M⁹			73.70⁹	-	Paper	TF	Keras	Pytorch	Caffe	-	-
ResNet34		224x224³	21.8M³		3.664G³		73.31³	91.40³	Paper	TF	-	Pytorch	Caffe	Torch	-
WideResNet18	widen=1.5	224x224³	25.88M³		3.866G³		73.15³	91.12³	Paper	TF	Keras	Pytorch	Caffe	-	-
DARTS		224x224⁹	4.9M⁹	595M⁹			73.10⁹	91.00⁹	Paper	-	-	Pytorch	-	-	-
MnasNet-65		224x224⁹	3.6M⁹	270M⁹			73.02⁹	91.14⁹	Paper	TF	Keras	Pytorch	Caffe	-	-
MobileNet_v2	α=1.0	224x224¹	3.47M				71.90¹	91.00¹	Paper	TF	Keras	Pytorch	Caffe	-	-
ShuffleNet	1.5	224x224⁹	3.4M⁹	292M⁹			71.50⁹	-	Paper	TF	Keras	Pytorch	Caffe	-	-
VGG 16		224x224¹	138M²		15.47⁴	23²	71.50¹	89.80¹	Paper	TF	Keras	Pytorch	Caffe	-	-
VGG 19		224x224¹	143M²			26²	71.10¹	89.80¹	Paper	TF	Keras	Pytorch	Caffe	-	-
CondenseNet	G=C=4	224x224⁹	2.9M⁹	274M⁹			71.00⁹	90.00⁹	Paper	TF	-	Pytorch	-	-	-
MobileNet_v1+CBAM	α=1.0	224x224³	5.07M³		0.576G³		70.99³	90.01³	Paper	-	Keras	-	-	-	-
MobileNet_v1	α=1.0	224x224¹					70.90¹	89.90¹	Paper	TF	Keras	Pytorch	Caffe	-	-
ResNet18+CBAM		224x224³	11.8M³		1.815G³		70.73³	89.91³	Paper	TF	Keras	-	-	-	-
ResNet18+SE		224x224³	11.8M³		1.814G³		70.59³	89.78³	Paper	TF	-	Pytorch	Caffe	-	-
ResNet18		224x224³	11.7M³		1.814G³		70.40³	89.45³	Paper	TF	-	Pytorch	Caffe	Torch	-
MobileNet_v1+SE	α=1.0	224x224³	5.07M³		0.570G³		70.03³	89.37³	Paper	-	-	-	Caffe	-	-
Inception V1		224x224¹					69.80¹	89.60¹	Paper	TF	Keras	Pytorch	-	-	-
MobileNet_v1	α=1.0	224x224³	4.23M³		0.569G³		68.61³	88.49³	Paper	TF	Keras	Pytorch	Caffe	-	-
MobileNet_v1+CBAM	α=0.7	224x224³	2.71M³		0.289G³		68.49³	88.52³	Paper	-	Keras	-	-	-	-
ShuffleNet+SE	1x(g=3)	224x224⁴	2.4M⁴		0.142G⁴		68.30⁴	88.30⁴	Paper	-	-	-	Caffe	-	-
SqueezeNext		224x224⁹	3.2M⁹	708M⁹			67.50⁹	88.20⁹	Paper	TF	-	Pytorch	Caffe	-	-
MobileNet_v1+SE	α=0.7	224x224³	2.71M³		0.283G³		67.50³	87.51³	Paper	-	-	-	Caffe	-	-
ShuffleNet	1x(g=3)	224x224⁴	1.8M⁴		0.14G⁴		66.10⁴	86.40⁴	Paper	TF	Keras	Pytorch	Caffe	-	-
MobileNet_v1	α=0.7	224x224³	2.3M³		0.283G³		65.14³	86.31³	Paper	TF	Keras	Pytorch	Caffe	-	-
MobileNet_v1	α=0.5	160x160¹					59.10¹	81.90¹	Paper	TF	Keras	Pytorch	Caffe	-	-
MobileNet_v1	α=0.25	128x128¹					41.50¹	66.30¹	Paper	TF	Keras	Pytorch	Caffe	-	-

• Mult-Adds: The number of multiply-add operations
• FLOPS: The floating point operations

Value References

Superscript numbers on each value indicate the reference number of each value from.

1. TF-Slim
2. Keras: Applications
3. CBAM: Convolutional Block Attention Module
4. Squeeze-and-Excitation Networks
5. Progressive Neural Architecture Search
6. Residual Attention Network for Image Classification
7. Deep Pyramidal Residual Networks
8. Big-Little Net: An Efficient Multi-Scale Feature Representation for Visual and Speech Recognition
9. MnasNet: Platform-Aware Neural Architecture Search for Mobile

Related Resources

Check the other good resources about CNN models

• Caffe-model
• TensorNets
• DeepDetect : Open Source Deep Learning Server & API
• Pretrained models for Pytorch
• Netscope CNN Analyzer
• Memory consumption and FLOP count

Author

Byung Soo Ko / [email protected]