| AlexNet ('One weird trick for parallelizing convolutional neural networks') |
62.3M |
1,132.33M |
40.96 |
18.24 |
2014 |
| VGG-16 ('Very Deep Convolutional Networks for Large-Scale Image Recognition') |
138.3M |
? |
26.78 |
8.69 |
2014 |
| ResNet-10 ('Deep Residual Learning for Image Recognition') |
5.5M |
894.04M |
34.69 |
14.36 |
2015 |
| ResNet-18 ('Deep Residual Learning for Image Recognition') |
11.7M |
1,820.41M |
28.53 |
9.82 |
2015 |
| ResNet-34 ('Deep Residual Learning for Image Recognition') |
21.8M |
3,672.68M |
24.84 |
7.80 |
2015 |
| ResNet-50 ('Deep Residual Learning for Image Recognition') |
25.5M |
3,877.95M |
22.28 |
6.33 |
2015 |
| InceptionV3 ('Rethinking the Inception Architecture for Computer Vision') |
23.8M |
? |
21.2 |
5.6 |
2015 |
| PreResNet-18 ('Identity Mappings in Deep Residual Networks') |
11.7M |
1,820.56M |
28.43 |
9.72 |
2016 |
| PreResNet-34 ('Identity Mappings in Deep Residual Networks') |
21.8M |
3,672.83M |
24.89 |
7.74 |
2016 |
| PreResNet-50 ('Identity Mappings in Deep Residual Networks') |
25.6M |
3,875.44M |
22.40 |
6.47 |
2016 |
| DenseNet-121 ('Densely Connected Convolutional Networks') |
8.0M |
2,872.13M |
23.48 |
7.04 |
2016 |
| DenseNet-161 ('Densely Connected Convolutional Networks') |
28.7M |
7,793.16M |
22.86 |
6.44 |
2016 |
| PyramidNet-101 ('Deep Pyramidal Residual Networks') |
42.5M |
8,743.54M |
21.98 |
6.20 |
2016 |
| ResNeXt-14(32x4d) ('Aggregated Residual Transformations for Deep Neural Networks') |
9.5M |
1,603.46M |
30.32 |
11.46 |
2016 |
| ResNeXt-26(32x4d) ('Aggregated Residual Transformations for Deep Neural Networks') |
15.4M |
2,488.07M |
24.14 |
7.46 |
2016 |
| WRN-50-2 ('Wide Residual Networks') |
68.9M |
11,405.42M |
22.53 |
6.41 |
2016 |
| Xception ('Xception: Deep Learning with Depthwise Separable Convolutions') |
22,855,952 |
8,403.63M |
20.97 |
5.49 |
2016 |
| InceptionV4 ('Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning') |
42,679,816 |
12,304.93M |
20.64 |
5.29 |
2016 |
| InceptionResNetV2 ('Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning') |
55,843,464 |
13,188.64M |
19.93 |
4.90 |
2016 |
| PolyNet ('PolyNet: A Pursuit of Structural Diversity in Very Deep Networks') |
95,366,600 |
34,821.34M |
19.10 |
4.52 |
2016 |
| DarkNet Ref ('Darknet: Open source neural networks in C') |
7,319,416 |
367.59M |
38.58 |
17.18 |
2016 |
| DarkNet Tiny ('Darknet: Open source neural networks in C') |
1,042,104 |
500.85M |
40.74 |
17.84 |
2016 |
| DarkNet 53 ('Darknet: Open source neural networks in C') |
41,609,928 |
7,133.86M |
21.75 |
5.64 |
2016 |
| SqueezeResNet1.1 ('SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size') |
1,235,496 |
352.02M |
40.09 |
18.21 |
2016 |
| SqueezeNet1.1 ('SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size') |
1,235,496 |
352.02M |
39.31 |
17.72 |
2016 |
| ResAttNet-92 ('Residual Attention Network for Image Classification') |
51.3M |
? |
19.5 |
4.8 |
2017 |
| CondenseNet (G=C=8) ('CondenseNet: An Efficient DenseNet using Learned Group Convolutions') |
4.8M |
? |
26.2 |
8.3 |
2017 |
| DPN-68 ('Dual Path Networks') |
12,611,602 |
2,351.84M |
23.24 |
6.79 |
2017 |
| ShuffleNet x1.0 (g=1) ('ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices') |
1,531,936 |
148.13M |
34.93 |
13.89 |
2017 |
| DiracNetV2-18 ('DiracNets: Training Very Deep Neural Networks Without Skip-Connections') |
11,511,784 |
1,796.62M |
31.47 |
11.70 |
2017 |
| DiracNetV2-34 ('DiracNets: Training Very Deep Neural Networks Without Skip-Connections') |
21,616,232 |
3,646.93M |
28.75 |
9.93 |
2017 |
| SENet-16 ('Squeeze-and-Excitation Networks') |
31,366,168 |
5,081.30M |
25.65 |
8.20 |
2017 |
| SENet-154 ('Squeeze-and-Excitation Networks') |
115,088,984 |
20,745.78M |
18.62 |
4.61 |
2017 |
| MobileNet ('MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications') |
4,231,976 |
579.80M |
26.61 |
8.95 |
2017 |
| NASNet-A 4@1056 ('Learning Transferable Architectures for Scalable Image Recognition') |
5,289,978 |
584.90M |
25.68 |
8.16 |
2017 |
| NASNet-A 6@4032('Learning Transferable Architectures for Scalable Image Recognition') |
88,753,150 |
23,976.44M |
18.14 |
4.21 |
2017 |
| DLA-34 ('Deep Layer Aggregation') |
15,742,104 |
3,071.37M |
25.36 |
7.94 |
2017 |
| AirNet50-1x64d (r=2) ('Attention Inspiring Receptive-Fields Network for Learning Invariant Representations') |
27.43M |
? |
22.48 |
6.21 |
2018 |
| BAM-ResNet-50 ('BAM: Bottleneck Attention Module') |
25.92M |
? |
23.68 |
6.96 |
2018 |
| CBAM-ResNet-50 ('CBAM: Convolutional Block Attention Module') |
28.1M |
? |
23.02 |
6.38 |
2018 |
| 1.0-SqNxt-23v5 ('SqueezeNext: Hardware-Aware Neural Network Design') |
921,816 |
285.82M |
40.77 |
17.85 |
2018 |
| 1.5-SqNxt-23v5 ('SqueezeNext: Hardware-Aware Neural Network Design') |
1,953,616 |
550.97M |
33.81 |
13.01 |
2018 |
| 2.0-SqNxt-23v5 ('SqueezeNext: Hardware-Aware Neural Network Design') |
3,366,344 |
897.60M |
29.63 |
10.66 |
2018 |
| ShuffleNetV2 ('ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design') |
2,278,604 |
149.72M |
31.44 |
11.63 |
2018 |
| 456-MENet-24×1(g=3) ('Merging and Evolution: Improving Convolutional Neural Networks for Mobile Applications') |
5.3M |
? |
28.4 |
9.8 |
2018 |
| FD-MobileNet ('FD-MobileNet: Improved MobileNet with A Fast Downsampling Strategy') |
2,901,288 |
147.46M |
34.23 |
13.38 |
2018 |
| MobileNetV2 ('MobileNetV2: Inverted Residuals and Linear Bottlenecks') |
3,504,960 |
329.36M |
26.97 |
8.87 |
2018 |
| IGCV3 ('IGCV3: Interleaved Low-Rank Group Convolutions for Efficient Deep Neural Networks') |
3.5M |
? |
28.22 |
9.54 |
2018 |
| DARTS ('DARTS: Differentiable Architecture Search') |
4.9M |
? |
26.9 |
9.0 |
2018 |
| PNASNet-5 ('Progressive Neural Architecture Search') |
5.1M |
? |
25.8 |
8.1 |
2018 |
| AmoebaNet-C ('Regularized Evolution for Image Classifier Architecture Search') |
5.1M |
? |
24.3 |
7.6 |
2018 |
| MnasNet ('MnasNet: Platform-Aware Neural Architecture Search for Mobile') |
4,308,816 |
317.67M |
31.58 |
11.74 |
2018 |
| IBN-Net50-a ('Two at Once: Enhancing Learning andGeneralization Capacities via IBN-Net') |
? |
? |
22.54 |
6.32 |
2018 |
| MarginNet ('Large Margin Deep Networks for Classification') |
? |
? |
22.0 |
? |
2018 |
| A^2 Net ('A^2-Nets: Double Attention Networks') |
? |
? |
23.0 |
6.5 |
2018 |
| FishNeXt-150 ('FishNet: A Versatile Backbone for Image, Region, and Pixel Level Prediction') |
26.2M |
? |
21.5 |
? |
2018 |
| Shape-ResNet ('IMAGENET-TRAINED CNNS ARE BIASED TOWARDS TEXTURE; INCREASING SHAPE BIAS IMPROVES ACCURACY AND ROBUSTNESS') |
25.5M |
? |
23.28 |
6.72 |
2019 |
| SimCNN(k=3 train) ('Greedy Layerwise Learning Can Scale to ImageNet') |
? |
? |
28.4 |
10.2 |
2019 |
| SKNet-50 ('Selective Kernel Networks') |
27.5M |
? |
20.79 |
? |
2019 |
| SRM-ResNet-50 ('SRM : A Style-based Recalibration Module for Convolutional Neural Networks') |
25.62M |
? |
22.87 |
6.49 |
2019 |
| EfficientNet-B0 ('EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks') |
5,288,548 |
414.31M |
24.77 |
7.52 |
2019 |
| EfficientNet-B7b ('EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks') |
66,347,960 |
39,010.98M |
15.94 |
3.22 |
2019 |
| ProxylessNAS ('PROXYLESSNAS: DIRECT NEURAL ARCHITECTURE SEARCH ON TARGET TASK AND HARDWARE') |
? |
? |
24.9 |
7.5 |
2019 |
| MixNet-L ('MixNet: Mixed Depthwise Convolutional Kernels') |
7.3M |
? |
21.1 |
5.8 |
2019 |
| ECA-Net50 ('ECA-Net: Efficient Channel Attention for Deep Convolutional Neural Networks') |
24.37M |
3.86G |
22.52 |
6.32 |
2019 |
| ECA-Net101 ('ECA-Net: Efficient Channel Attention for Deep Convolutional Neural Networks') |
7.3M |
7.35G |
21.35 |
5.66 |
2019 |
| ACNet-Densenet121 ('ACNet: Strengthening the Kernel Skeletons for Powerful CNN via Asymmetric Convolution Blocks') |
? |
? |
24.18 |
7.23 |
2019 |
| LIP-ResNet-50 ('LIP: Local Importance-based Pooling') |
23.9M |
5.33G |
21.81 |
6.04 |
2019 |
| LIP-ResNet-101 ('LIP: Local Importance-based Pooling') |
42.9M |
9.06G |
20.67 |
5.40 |
2019 |
| LIP-DenseNet-BC-121 ('LIP: Local Importance-based Pooling') |
8.7M |
4.13G |
23.36 |
6.84 |
2019 |
| MuffNet_1.0 ('MuffNet: Multi-Layer Feature Federation for Mobile Deep Learning') |
2.3M |
146M |
30.1 |
? |
2019 |
| MuffNet_1.5 ('MuffNet: Multi-Layer Feature Federation for Mobile Deep Learning') |
3.4M |
300M |
26.9 |
? |
2019 |
| ResNet-34-Bin-5 ('Making Convolutional Networks Shift-Invariant Again') |
21.8M |
3,672.68M |
25.80 |
? |
2019 |
| ResNet-50-Bin-5 ('Making Convolutional Networks Shift-Invariant Again') |
25.5M |
3,877.95M |
22.96 |
? |
2019 |
| MobileNetV2-Bin-5 ('Making Convolutional Networks Shift-Invariant Again') |
3,504,960 |
329.36M |
27.50 |
? |
2019 |
| FixRes ResNeXt101 WSL ('Fixing the train-test resolution discrepancy') |
829M |
? |
13.6 |
2.0 |
2019 |
| Noisy Student*(L2) ('Self-training with Noisy Student improves ImageNet classification') |
480M |
? |
12.6 |
1.8 |
2019 |
| TResNet-M ('TResNet: High Performance GPU-Dedicated Architecture') |
29.4M |
5.5G |
19.3 |
? |
2020 |
| DA-NAS-C ('DA-NAS: Data Adapted Pruning for Efficient Neural Architecture Search') |
? |
467M |
23.8 |
? |
2020 |
| ResNeSt-50 ('ResNeSt: Split-Attention Networks') |
27.5M |
5.39G |
18.87 |
? |
2020 |
| ResNeSt-101 ('ResNeSt: Split-Attention Networks') |
48.3M |
10.2G |
17.73 |
? |
2020 |
| ResNet-50-FReLU ('Funnel Activation for Visual Recognition') |
25.5M |
3.87G |
22.40 |
? |
2020 |
| ResNet-101-FReLU ('Funnel Activation for Visual Recognition') |
44.5M |
7.6G |
22.10 |
? |
2020 |
| ResNet-50-MEALv2 ('MEAL V2: Boosting Vanilla ResNet-50 to 80%+ Top-1 Accuracy on ImageNet without Tricks') |
25.6M |
? |
19.33 |
4.91 |
2020 |
| ResNet-50-MEALv2 + CutMix ('MEAL V2: Boosting Vanilla ResNet-50 to 80%+ Top-1 Accuracy on ImageNet without Tricks') |
25.6M |
? |
19.02 |
4.65 |
2020 |
| MobileNet V3-Large-MEALv2 ('MEAL V2: Boosting Vanilla ResNet-50 to 80%+ Top-1 Accuracy on ImageNet without Tricks') |
5.48M |
? |
23.08 |
6.68 |
2020 |
| EfficientNet-B0-MEALv2 ('MEAL V2: Boosting Vanilla ResNet-50 to 80%+ Top-1 Accuracy on ImageNet without Tricks') |
5.29M |
? |
21.71 |
6.05 |
2020 |
| T2T-ViT-7 ('Tokens-to-Token ViT: Training Vision Transformers from Scratch on ImageNet') |
4.2M |
0.6G |
28.8 |
? |
2021 |
| T2T-ViT-14 ('Tokens-to-Token ViT: Training Vision Transformers from Scratch on ImageNet') |
19.4M |
4.8G |
19.4 |
? |
2021 |
| T2T-ViT-19 ('Tokens-to-Token ViT: Training Vision Transformers from Scratch on ImageNet') |
39.0M |
8.0G |
18.8 |
? |
2021 |
| NFNet-F0 ('High-Performance Large-Scale Image Recognition Without Normalization') |
71.5M |
12.38G |
16.4 |
3.2 |
2021 |
| NFNet-F1 ('High-Performance Large-Scale Image Recognition Without Normalization') |
132.6M |
35.54G |
15.4 |
2.9 |
2021 |
| NFNet-F6+SAM ('High-Performance Large-Scale Image Recognition Without Normalization') |
438.4M |
377.28G |
13.5 |
2.1 |
2021 |
| EfficientNetV2-S ('EfficientNetV2: Smaller Models and Faster Training') |
24M |
8.8G |
16.1 |
? |
2021 |
| EfficientNetV2-M ('EfficientNetV2: Smaller Models and Faster Training') |
55M |
24G |
14.9 |
? |
2021 |
| EfficientNetV2-L ('EfficientNetV2: Smaller Models and Faster Training') |
121M |
53G |
14.3 |
? |
2021 |
| EfficientNetV2-S (21k) ('EfficientNetV2: Smaller Models and Faster Training') |
24M |
8.8G |
15.0 |
? |
2021 |
| EfficientNetV2-M (21k) ('EfficientNetV2: Smaller Models and Faster Training') |
55M |
24G |
13.9 |
? |
2021 |
| EfficientNetV2-L (21k) ('EfficientNetV2: Smaller Models and Faster Training') |
121M |
53G |
13.2 |
? |
2021 |
