ImageNet Model zoo overview

August 18, 2022 · View on GitHub

Catalogue

1. Model library overview diagram

Based on the ImageNet-1k classification dataset, the 37 classification network structures supported by PaddleClas and the corresponding 217 image classification pretrained models are shown below. Training trick, a brief introduction to each series of network structures, and performance evaluation will be shown in the corresponding chapters. The evaluation environment is as follows.

  • Arm CPU evaluation environment is based on Snapdragon 855 (SD855).
  • Intel CPU evaluation environment is based on Intel(R) Xeon(R) Gold 6148.
  • The GPU evaluation speed is measured by running 2100 times under the FP32+TensorRT configuration (excluding the warmup time of the first 100 times).
  • FLOPs and Params are calculated by paddle.flops() (PaddlePaddle version is 2.2)

Curves of accuracy to the inference time of common server-side models are shown as follows.

Curves of accuracy to the inference time of common mobile-side models are shown as follows.

Curves of accuracy to the inference time of some VisionTransformer models are shown as follows.

2. SSLD pretrained models

Accuracy and inference time of the prtrained models based on SSLD distillation are as follows. More detailed information can be refered to SSLD distillation tutorial.

2.1 Server-side knowledge distillation model

ModelTop-1 AccReference
Top-1 Acc		Acc gaintime(ms)
bs=1		time(ms)
bs=4		time(ms)
bs=8		FLOPs(G)Params(M)Pretrained Model Download AddressInference Model Download Address
ResNet34_vd_ssld0.7970.7600.0372.003.285.843.9321.84Download link  Download link  
ResNet50_vd_ssld0.8300.7920.0392.604.867.634.3525.63Download linkDownload link
ResNet101_vd_ssld0.8370.8020.0354.438.2512.608.0844.67Download linkDownload link
Res2Net50_vd_26w_4s_ssld0.8310.7980.0333.596.359.504.2825.76Download linkDownload link
Res2Net101_vd_
26w_4s_ssld		0.8390.8060.0336.3411.0216.138.3545.35Download linkDownload link
Res2Net200_vd_
26w_4s_ssld		0.8510.8120.04911.4519.7728.8115.7776.44Download linkDownload link
HRNet_W18_C_ssld0.8120.7690.0436.668.9411.954.3221.35Download linkDownload link
HRNet_W48_C_ssld0.8360.7900.04611.0717.0627.2817.3477.57Download linkDownload link
SE_HRNet_W64_C_ssld0.848--17.1126.8743.2429.00129.12Download linkDownload link

2.2 Mobile-side knowledge distillation model

ModelTop-1 AccReference
Top-1 Acc		Acc gainSD855 time(ms)
bs=1, thread=1		SD855 time(ms)
bs=1, thread=2		SD855 time(ms)
bs=1, thread=4		FLOPs(M)Params(M)Model大小(M)Pretrained Model Download AddressInference Model Download Address
MobileNetV1_ssld0.7790.7100.06930.2417.8610.30578.884.2516Download linkDownload link
MobileNetV2_ssld0.7670.7220.04520.7412.718.10327.843.5414Download linkDownload link
MobileNetV3_small_x0_35_ssld0.5560.5300.0262.231.661.4314.561.676.9Download linkDownload link
MobileNetV3_large_x1_0_ssld0.7900.7530.03616.5510.096.84229.665.5021Download linkDownload link
MobileNetV3_small_x1_0_ssld0.7130.6820.0315.633.652.6063.672.9512Download linkDownload link
GhostNet_x1_3_ssld0.7940.7570.03719.1612.259.40236.897.3829Download linkDownload link

2.3 Intel-CPU-side knowledge distillation model

ModelTop-1 AccReference
Top-1 Acc		Acc gainIntel-Xeon-Gold-6148 time(ms)
bs=1		FLOPs(M)Params(M)Pretrained Model Download AddressInference Model Download Address
PPLCNet_x0_5_ssld0.6610.6310.0302.0547.281.89Download linkDownload link
PPLCNet_x1_0_ssld0.7440.7130.0332.46160.812.96Download linkDownload link
PPLCNet_x2_5_ssld0.8080.7660.0425.39906.499.04Download linkDownload link
  • Note: Reference Top-1 Acc means the accuracy of the pre-trained model obtained by PaddleClas based on ImageNet1k dataset training.

3. PP-LCNet series [28]

The accuracy and speed indicators of the PP-LCNet series models are shown in the following table. For more information about this series of models, please refer to: PP-LCNet series model documents。

ModelTop-1 AccTop-5 AccIntel-Xeon-Gold-6148 time(ms)
bs=1		FLOPs(M)Params(M)Pretrained Model Download AddressInference Model Download Address
PPLCNet_x0_250.51860.75651.6178518.251.52Download linkDownload link
PPLCNet_x0_350.58090.80832.1134429.461.65Download linkDownload link
PPLCNet_x0_50.63140.84662.7297447.281.89Download linkDownload link
PPLCNet_x0_750.68180.88304.5121698.822.37Download linkDownload link
PPLCNet_x1_00.71320.90036.49276160.812.96Download linkDownload link
PPLCNet_x1_50.73710.915312.2601341.864.52Download linkDownload link
PPLCNet_x2_00.75180.922720.16675906.54Download linkDownload link
PPLCNet_x2_50.76600.930029.5959069.04Download linkDownload link

4. ResNet series [1]

The accuracy and speed indicators of ResNet and ResNet_vd series models are shown in the following table. For more information about this series of models, please refer to: ResNet and ResNet_vd series model documents。

ModelTop-1 AccTop-5 Acctime(ms)
bs=1		time(ms)
bs=4		time(ms)
bs=8		FLOPs(G)Params(M)Pretrained Model Download AddressInference Model Download Address
ResNet180.70980.89921.222.193.631.8311.70Download linkDownload link
ResNet18_vd0.72260.90801.262.283.892.0711.72Download linkDownload link
ResNet340.74570.92141.973.255.703.6821.81Download linkDownload link
ResNet34_vd0.75980.92982.003.285.843.9321.84Download linkDownload link
ResNet34_vd_ssld0.79720.94902.003.285.843.9321.84Download linkDownload link
ResNet500.76500.93002.544.797.404.1125.61Download linkDownload link
ResNet50_vc0.78350.94032.574.837.524.3525.63Download linkDownload link
ResNet50_vd0.79120.94442.604.867.634.3525.63Download linkDownload link
ResNet1010.77560.93644.378.1812.387.8344.65Download linkDownload link
ResNet101_vd0.80170.94974.438.2512.608.0844.67Download linkDownload link
ResNet1520.78260.93966.0511.4117.3311.5660.34Download linkDownload link
ResNet152_vd0.80590.95306.1111.5117.5911.8060.36Download linkDownload link
ResNet200_vd0.80930.95337.7014.5722.1615.3074.93Download linkDownload link
ResNet50_vd_
ssld		0.83000.96402.604.867.634.3525.63Download linkDownload link
ResNet101_vd_
ssld		0.83730.96694.438.2512.608.0844.67Download linkDownload link

5. Mobile series [3][4][5][6][23]

The accuracy and speed indicators of the mobile series models are shown in the following table. For more information about this series, please refer to: Mobile series model documents。

ModelTop-1 AccTop-5 AccSD855 time(ms)
bs=1, thread=1		SD855 time(ms)
bs=1, thread=2		SD855 time(ms)
bs=1, thread=4		FLOPs(M)Params(M)Model大小(M)Pretrained Model Download AddressInference Model Download Address
MobileNetV1_
x0_25		0.51430.75462.881.821.2643.560.481.9Download linkDownload link
MobileNetV1_
x0_5		0.63520.84738.745.263.09154.571.345.2Download linkDownload link
MobileNetV1_
x0_75		0.68810.882317.8410.616.21333.002.6010Download linkDownload link
MobileNetV10.70990.896830.2417.8610.30578.884.2516Download linkDownload link
MobileNetV1_
ssld		0.77890.939430.2417.8610.30578.884.2516Download linkDownload link
MobileNetV2_
x0_25		0.53210.76523.462.512.0334.181.536.1Download linkDownload link
MobileNetV2_
x0_5		0.65030.85727.694.923.5799.481.987.8Download linkDownload link
MobileNetV2_
x0_75		0.69830.890113.698.605.82197.372.6510Download linkDownload link
MobileNetV20.72150.906520.7412.718.10327.843.5414Download linkDownload link
MobileNetV2_
x1_5		0.74120.916740.7924.4915.50702.356.9026Download linkDownload link
MobileNetV2_
x2_0		0.75230.925867.5040.0325.551217.2511.3343Download linkDownload link
MobileNetV2_
ssld		0.76740.933920.7412.718.10327.843.5414Download linkDownload link
MobileNetV3_
large_x1_25		0.76410.929524.5214.769.89362.707.4729Download linkDownload link
MobileNetV3_
large_x1_0		0.75320.923116.5510.096.84229.665.5021Download linkDownload link
MobileNetV3_
large_x0_75		0.73140.910811.537.064.94151.703.9316Download linkDownload link
MobileNetV3_
large_x0_5		0.69240.88526.504.223.1571.832.6911Download linkDownload link
MobileNetV3_
large_x0_35		0.64320.85464.433.112.4140.902.118.6Download linkDownload link
MobileNetV3_
small_x1_25		0.70670.89517.884.913.45100.073.6414Download linkDownload link
MobileNetV3_
small_x1_0		0.68240.88065.633.652.6063.672.9512Download linkDownload link
MobileNetV3_
small_x0_75		0.66020.86334.502.962.1946.022.389.6Download linkDownload link
MobileNetV3_
small_x0_5		0.59210.81522.892.041.6222.601.917.8Download linkDownload link
MobileNetV3_
small_x0_35		0.53030.76372.231.661.4314.561.676.9Download linkDownload link
MobileNetV3_
small_x0_35_ssld		0.55550.77712.231.661.4314.561.676.9Download linkDownload link
MobileNetV3_
large_x1_0_ssld		0.78960.944816.5510.096.84229.665.5021Download linkDownload link
MobileNetV3_small_
x1_0_ssld		0.71290.90105.633.652.6063.672.9512Download linkDownload link
ShuffleNetV20.68800.88459.725.974.13148.862.299Download linkDownload link
ShuffleNetV2_
x0_25		0.49900.73791.941.531.4318.950.612.7Download linkDownload link
ShuffleNetV2_
x0_33		0.53730.77052.231.701.7924.040.652.8Download linkDownload link
ShuffleNetV2_
x0_5		0.60320.82263.672.632.0642.581.375.6Download linkDownload link
ShuffleNetV2_
x1_5		0.71630.901517.2110.566.81301.353.5314Download linkDownload link
ShuffleNetV2_
x2_0		0.73150.912031.2118.9811.65571.707.4028Download linkDownload link
ShuffleNetV2_
swish		0.70030.891731.219.065.74148.862.299.1Download linkDownload link
GhostNet_
x0_5		0.66880.86955.283.953.2946.152.6010Download linkDownload link
GhostNet_
x1_0		0.74020.916512.898.666.72148.785.2120Download linkDownload link
GhostNet_
x1_3		0.75790.925419.1612.259.40236.897.3829Download linkDownload link
GhostNet_
x1_3_ssld		0.79380.944919.1612.259.40236.897.3829Download linkDownload link
ESNet_x0_250.62480.83464.122.972.5130.852.8311Download linkDownload link
ESNet_x0_50.68820.88046.454.423.3567.313.2513Download linkDownload link
ESNet_x0_750.72240.90459.596.284.52123.743.8715Download linkDownload link
ESNet_x1_00.73920.914013.678.715.97197.334.6418Download linkDownload link

6. SEResNeXt and Res2Net series [7][8][9]

The accuracy and speed indicators of the SEResNeXt and Res2Net series models are shown in the following table. For more information about the models of this series, please refer to: SEResNeXt and Res2Net series model documents.

ModelTop-1 AccTop-5 Acctime(ms)
bs=1		time(ms)
bs=4		time(ms)
bs=8		FLOPs(G)Params(M)Pretrained Model Download AddressInference Model Download Address
Res2Net50_
26w_4s		0.79330.94573.526.239.304.2825.76Download linkDownload link
Res2Net50_vd_
26w_4s		0.79750.94913.596.359.504.5225.78Download linkDownload link
Res2Net50_
14w_8s		0.79460.94704.397.2110.384.2025.12Download linkDownload link
Res2Net101_vd_
26w_4s		0.80640.95226.3411.0216.138.3545.35Download linkDownload link
Res2Net200_vd_
26w_4s		0.81210.957111.4519.7728.8115.7776.44Download linkDownload link
Res2Net200_vd_
26w_4s_ssld		0.85130.974211.4519.7728.8115.7776.44Download linkDownload link
ResNeXt50_
32x4d		0.77750.93825.078.4912.024.2625.10Download linkDownload link
ResNeXt50_vd_
32x4d		0.79560.94625.298.6812.334.5025.12Download linkDownload link
ResNeXt50_
64x4d		0.78430.94139.3913.9720.568.0245.29Download linkDownload link
ResNeXt50_vd_
64x4d		0.80120.94869.7514.1420.848.2645.31Download linkDownload link
ResNeXt101_
32x4d		0.78650.941911.3416.7822.808.0144.32Download linkDownload link
ResNeXt101_vd_
32x4d		0.80330.951211.3617.0123.078.2544.33Download linkDownload link
ResNeXt101_
64x4d		0.78350.945221.5728.0839.4915.5283.66Download linkDownload link
ResNeXt101_vd_
64x4d		0.80780.952021.5728.2239.7015.7683.68Download linkDownload link
ResNeXt152_
32x4d		0.78980.943317.1425.1133.7911.7660.15Download linkDownload link
ResNeXt152_vd_
32x4d		0.80720.952016.9925.2933.8512.0160.17Download linkDownload link
ResNeXt152_
64x4d		0.79510.947133.0742.0559.1323.03115.27Download linkDownload link
ResNeXt152_vd_
64x4d		0.81080.953433.3042.4159.4223.27115.29Download linkDownload link
SE_ResNet18_vd0.73330.91381.482.704.322.0711.81Download linkDownload link
SE_ResNet34_vd0.76510.93202.423.696.293.9322.00Download linkDownload link
SE_ResNet50_vd0.79520.94753.115.999.344.3628.16Download linkDownload link
SE_ResNeXt50_
32x4d		0.78440.93966.3911.0114.944.2727.63Download linkDownload link
SE_ResNeXt50_vd_
32x4d		0.80240.94897.0411.5716.015.6427.76Download linkDownload link
SE_ResNeXt101_
32x4d		0.79390.944313.3121.8528.778.0349.09Download linkDownload link
SENet154_vd0.81400.954834.8351.2269.7424.45122.03Download linkDownload link

7. DPN and DenseNet series [14][15]

The accuracy and speed indicators of the DPN and DenseNet series models are shown in the following table. For more information about the models of this series, please refer to: DPN and DenseNet series model documents.

ModelTop-1 AccTop-5 Acctime(ms)
bs=1		time(ms)
bs=4		time(ms)
bs=8		FLOPs(G)Params(M)Pretrained Model Download AddressInference Model Download Address
DenseNet1210.75660.92583.406.949.172.878.06Download linkDownload link
DenseNet1610.78570.94147.0614.3719.557.7928.90Download linkDownload link
DenseNet1690.76810.93315.0010.2912.843.4014.31Download linkDownload link
DenseNet2010.77630.93666.3813.7217.174.3420.24Download linkDownload link
DenseNet2640.77960.93859.3420.9525.415.8233.74Download linkDownload link
DPN680.76780.93438.1811.4014.822.3512.68Download linkDownload link
DPN920.79850.948012.4820.0425.106.5437.79Download linkDownload link
DPN980.80590.951014.7025.5535.1211.72861.74Download linkDownload link
DPN1070.80890.953219.4635.6250.2218.3887.13Download linkDownload link
DPN1310.80700.951419.6434.6047.4216.0979.48Download linkDownload link

8. HRNet series [13]

The accuracy and speed indicators of the HRNet series models are shown in the following table. For more information about the models of this series, please refer to: HRNet series model documents.

ModelTop-1 AccTop-5 Acctime(ms)
bs=1		time(ms)
bs=4		time(ms)
bs=8		FLOPs(G)Params(M)Pretrained Model Download AddressInference Model Download Address
HRNet_W18_C0.76920.93396.668.9411.954.3221.35Download linkDownload link
HRNet_W18_C_ssld0.811620.958046.668.9411.954.3221.35Download linkDownload link
HRNet_W30_C0.78040.94028.6111.4015.238.1537.78Download linkDownload link
HRNet_W32_C0.78280.94248.5411.5815.578.9741.30Download linkDownload link
HRNet_W40_C0.78770.94479.8315.0220.9212.7457.64Download linkDownload link
HRNet_W44_C0.79000.945110.6216.1825.9214.9467.16Download linkDownload link
HRNet_W48_C0.78950.944211.0717.0627.2817.3477.57Download linkDownload link
HRNet_W48_C_ssld0.83630.968211.0717.0627.2817.3477.57Download linkDownload link
HRNet_W64_C0.79300.946113.8221.1535.5128.97128.18Download linkDownload link
SE_HRNet_W64_C_ssld0.84750.972617.1126.8743.2429.00129.12Download linkDownload link

9. Inception series [10][11][12][26]

The accuracy and speed indicators of the Inception series models are shown in the following table. For more information about this series of models, please refer to: Inception series model documents.

ModelTop-1 AccTop-5 Acctime(ms)
bs=1		time(ms)
bs=4		time(ms)
bs=8		FLOPs(G)Params(M)Pretrained Model Download AddressInference Model Download Address
GoogLeNet0.70700.89661.413.255.001.4411.54Download linkDownload link
Xception410.79300.94533.588.7616.618.5723.02Download linkDownload link
Xception41_deeplab0.79550.94383.819.1617.209.2827.08Download linkDownload link
Xception650.81000.95495.4512.7824.5313.2536.04Download linkDownload link
Xception65_deeplab0.80320.94495.6513.0824.6113.9640.10Download linkDownload link
Xception710.81110.95456.1915.3429.2116.2137.86Download linkDownload link
InceptionV30.79140.94594.788.5312.285.7323.87Download linkDownload link
InceptionV40.80770.95268.9315.1721.5612.2942.74Download linkDownload link

10. EfficientNet and ResNeXt101_wsl series [16][17]

The accuracy and speed indicators of the EfficientNet and ResNeXt101_wsl series models are shown in the following table. For more information about this series of models, please refer to: EfficientNet and ResNeXt101_wsl series model documents.

ModelTop-1 AccTop-5 Acctime(ms)
bs=1		time(ms)
bs=4		time(ms)
bs=8		FLOPs(G)Params(M)Pretrained Model Download AddressInference Model Download Address
ResNeXt101_
32x8d_wsl		0.82550.967413.5523.3936.1816.4888.99Download linkDownload link
ResNeXt101_
32x16d_wsl		0.84240.972621.9638.3563.2936.26194.36Download linkDownload link
ResNeXt101_
32x32d_wsl		0.84970.975937.2876.50121.5687.28469.12Download linkDownload link
ResNeXt101_
32x48d_wsl		0.85370.976955.07124.39205.01153.57829.26Download linkDownload link
Fix_ResNeXt101_
32x48d_wsl		0.86260.979755.01122.63204.66313.41829.26Download linkDownload link
EfficientNetB00.77380.93311.963.715.560.405.33Download linkDownload link
EfficientNetB10.79150.94412.885.407.630.717.86Download linkDownload link
EfficientNetB20.79850.94743.266.209.171.029.18Download linkDownload link
EfficientNetB30.81150.95414.528.8513.541.8812.324Download linkDownload link
EfficientNetB40.82850.96236.7815.4724.954.5119.47Download linkDownload link
EfficientNetB50.83620.967210.9727.2445.9310.5130.56Download linkDownload link
EfficientNetB60.84000.968817.0943.3276.9019.4743.27Download linkDownload link
EfficientNetB70.84300.968925.9171.23128.2038.4566.66Download linkDownload link
EfficientNetB0_
small		0.75800.92581.242.593.920.404.69Download linkDownload link

11. ResNeSt and RegNet series [24][25]

The accuracy and speed indicators of the ResNeSt and RegNet series models are shown in the following table. For more information about the models of this series, please refer to: ResNeSt and RegNet series model documents.

ModelTop-1 AccTop-5 Acctime(ms)
bs=1		time(ms)
bs=4		time(ms)
bs=8		FLOPs(G)Params(M)Pretrained Model Download AddressInference Model Download Address
ResNeSt50_
fast_1s1x64d		0.80350.95282.735.338.244.3626.27Download linkDownload link
ResNeSt500.80830.95427.3610.2313.845.4027.54Download linkDownload link
RegNetX_4GF0.7850.94166.468.4811.454.0022.23Download linkDownload link

12. ViT and DeiT series [31][32]

The accuracy and speed indicators of ViT (Vision Transformer) and DeiT (Data-efficient Image Transformers) series models are shown in the following table. For more information about this series of models, please refer to: ViT_and_DeiT series model documents.

ModelTop-1 AccTop-5 Acctime(ms)
bs=1		time(ms)
bs=4		time(ms)
bs=8		FLOPs(G)Params(M)Pretrained Model Download AddressInference Model Download Address
ViT_small_
patch16_224		0.77690.93423.719.0516.729.4148.60Download linkDownload link
ViT_base_
patch16_224		0.81950.96176.1214.8428.5116.8586.42Download linkDownload link
ViT_base_
patch16_384		0.84140.971714.1548.3895.0649.3586.42Download linkDownload link
ViT_base_
patch32_384		0.81760.96134.9413.4324.0812.6688.19Download linkDownload link
ViT_large_
patch16_224		0.83230.965015.5349.5094.0959.65304.12Download linkDownload link
ViT_large_
patch16_384		0.85130.973639.51152.46304.06174.70304.12Download linkDownload link
ViT_large_
patch32_384		0.81530.960811.4436.0970.6344.24306.48Download linkDownload link
ModelTop-1 AccTop-5 Acctime(ms)
bs=1		time(ms)
bs=4		time(ms)
bs=8		FLOPs(G)Params(M)Pretrained Model Download AddressInference Model Download Address
DeiT_tiny_
patch16_224		0.7180.9103.613.946.101.075.68Download linkDownload link
DeiT_small_
patch16_224		0.7960.9493.616.2410.494.2421.97Download linkDownload link
DeiT_base_
patch16_224		0.8170.9576.1314.8728.5016.8586.42Download linkDownload link
DeiT_base_
patch16_384		0.8300.96214.1248.8097.6049.3586.42Download linkDownload link
DeiT_tiny_
distilled_patch16_224		0.7410.9183.514.056.031.085.87Download linkDownload link
DeiT_small_
distilled_patch16_224		0.8090.9533.706.2010.534.2622.36Download linkDownload link
DeiT_base_
distilled_patch16_224		0.8310.9646.1714.9428.5816.9387.18Download linkDownload link
DeiT_base_
distilled_patch16_384		0.8510.97314.1248.7697.0949.4387.18Download linkDownload link

13. RepVGG series [36]

The accuracy and speed indicators of RepVGG series models are shown in the following table. For more introduction, please refer to: RepVGG series model documents.

ModelTop-1 AccTop-5 Acctime(ms)
bs=1		time(ms)
bs=4		time(ms)
bs=8		FLOPs(G)Params(M)Pretrained Model Download AddressInference Model Download Address
RepVGG_A00.71310.90161.368.31Download linkDownload link
RepVGG_A10.73800.91462.3712.79Download linkDownload link
RepVGG_A20.75710.92645.1225.50Download linkDownload link
RepVGG_B00.74500.92133.0614.34Download linkDownload link
RepVGG_B10.77730.938511.8251.83Download linkDownload link
RepVGG_B20.78130.941018.3880.32Download linkDownload link
RepVGG_B1g20.77320.93598.8241.36Download linkDownload link
RepVGG_B1g40.76750.93357.3136.13Download linkDownload link
RepVGG_B2g40.78810.944811.3455.78Download linkDownload link
RepVGG_B3g40.79650.948516.0775.63Download linkDownload link

14. MixNet series [29]

The accuracy and speed indicators of the MixNet series models are shown in the following table. For more introduction, please refer to: MixNet series model documents.

ModelTop-1 AccTop-5 Acctime(ms)
bs=1		time(ms)
bs=4		time(ms)
bs=8		FLOPs(M)Params(M)Pretrained Model Download AddressInference Model Download Address
MixNet_S0.76280.92992.313.635.20252.9774.167Download linkDownload link
MixNet_M0.77670.93642.844.606.62357.1195.065Download linkDownload link
MixNet_L0.78600.94373.165.558.03579.0177.384Download linkDownload link

15. ReXNet series [30]

The accuracy and speed indicators of ReXNet series models are shown in the following table. For more introduction, please refer to: ReXNet series model documents.

ModelTop-1 AccTop-5 Acctime(ms)
bs=1		time(ms)
bs=4		time(ms)
bs=8		FLOPs(G)Params(M)Pretrained Model Download AddressInference Model Download Address
ReXNet_1_00.77460.93703.084.155.490.4154.84Download linkDownload link
ReXNet_1_30.79130.94643.544.876.540.687.61Download linkDownload link
ReXNet_1_50.80060.95123.685.317.380.909.79Download linkDownload link
ReXNet_2_00.81220.95364.306.549.191.5616.45Download linkDownload link
ReXNet_3_00.82090.96125.749.4913.623.4434.83Download linkDownload link

16. SwinTransformer series [27]

The accuracy and speed indicators of SwinTransformer series models are shown in the following table. For more introduction, please refer to: SwinTransformer series model documents.

ModelTop-1 AccTop-5 Acctime(ms)
bs=1		time(ms)
bs=4		time(ms)
bs=8		FLOPs(G)Params(M)Pretrained Model Download AddressInference Model Download Address
SwinTransformer_tiny_patch4_window7_2240.80690.95346.599.6816.324.3528.26Download linkDownload link
SwinTransformer_small_patch4_window7_2240.82750.961312.5417.0728.088.5149.56Download linkDownload link
SwinTransformer_base_patch4_window7_2240.83000.962613.3723.5339.1115.1387.70Download linkDownload link
SwinTransformer_base_patch4_window12_3840.84390.969319.5264.56123.3044.4587.70Download linkDownload link
SwinTransformer_base_patch4_window7_224[1]0.84870.974613.5323.4639.1315.1387.70Download linkDownload link
SwinTransformer_base_patch4_window12_384[1]0.86420.980719.6564.72123.4244.4587.70Download linkDownload link
SwinTransformer_large_patch4_window7_224[1]0.85960.978315.7438.5771.4934.02196.43Download linkDownload link
SwinTransformer_large_patch4_window12_384[1]0.87190.982332.61116.59223.2399.97196.43Download linkDownload link

[1]：It is pre-trained based on the ImageNet22k dataset, and then transferred and learned from the ImageNet1k dataset.

17. LeViT series [33]

The accuracy and speed indicators of LeViT series models are shown in the following table. For more introduction, please refer to: LeViT series model documents.

ModelTop-1 AccTop-5 Acctime(ms)
bs=1		time(ms)
bs=4		time(ms)
bs=8		FLOPs(M)Params(M)Pretrained Model Download AddressInference Model Download Address
LeViT_128S0.75980.92692817.42Download linkDownload link
LeViT_1280.78100.93713658.87Download linkDownload link
LeViT_1920.79340.944659710.61Download linkDownload link
LeViT_2560.80850.9497104918.45Download linkDownload link
LeViT_3840.81910.9551223438.45Download linkDownload link

Note: The accuracy difference with Reference is due to the difference in data preprocessing and the use of no distilled head as output.

18. Twins series [34]

The accuracy and speed indicators of Twins series models are shown in the following table. For more introduction, please refer to: Twins series model documents.

ModelTop-1 AccTop-5 Acctime(ms)
bs=1		time(ms)
bs=4		time(ms)
bs=8		FLOPs(G)Params(M)Pretrained Model Download AddressInference Model Download Address
pcpvt_small0.80820.95527.3210.5115.273.6724.06Download linkDownload link
pcpvt_base0.82420.961912.2016.2223.166.4443.83Download linkDownload link
pcpvt_large0.82730.965016.4722.9032.739.5060.99Download linkDownload link
alt_gvt_small0.81400.95466.949.0112.272.8124.06Download linkDownload link
alt_gvt_base0.82940.96219.3715.0224.548.3456.07Download linkDownload link
alt_gvt_large0.83310.964211.7622.0835.1214.8199.27Download linkDownload link

Note: The accuracy difference with Reference is due to the difference in data preprocessing.

19. HarDNet series [37]

The accuracy and speed indicators of HarDNet series models are shown in the following table. For more introduction, please refer to: HarDNet series model documents.

ModelTop-1 AccTop-5 Acctime(ms)
bs=1		time(ms)
bs=4		time(ms)
bs=8		FLOPs(G)Params(M)Pretrained Model Download AddressInference Model Download Address
HarDNet39_ds0.71330.89981.402.303.330.443.51Download linkDownload link
HarDNet68_ds0.73620.91522.263.345.060.794.20Download linkDownload link
HarDNet680.75460.92653.588.5311.584.2617.58Download linkDownload link
HarDNet850.77440.93556.2414.8520.579.0936.69Download linkDownload link

20. DLA series [38]

The accuracy and speed indicators of DLA series models are shown in the following table. For more introduction, please refer to: DLA series model documents.

ModelTop-1 AccTop-5 Acctime(ms)
bs=1		time(ms)
bs=4		time(ms)
bs=8		FLOPs(G)Params(M)Pretrained Model Download AddressInference Model Download Address
DLA1020.78930.94524.958.0812.407.1933.34Download linkDownload link
DLA102x20.78850.944519.5823.9731.379.3441.42Download linkDownload link
DLA102x0.7810.940011.1215.6020.375.8926.40Download linkDownload link
DLA1690.78090.94097.7012.2518.9011.5953.50Download linkDownload link
DLA340.76030.92981.833.375.983.0715.76Download linkDownload link
DLA46_c0.63210.8531.062.083.230.541.31Download linkDownload link
DLA600.76100.92922.785.368.294.2622.08Download linkDownload link
DLA60x_c0.66450.87541.793.685.190.591.33Download linkDownload link
DLA60x0.77530.93785.989.2412.523.5417.41Download linkDownload link

21. RedNet series [39]

The accuracy and speed indicators of RedNet series models are shown in the following table. For more introduction, please refer to: RedNet series model documents.

ModelTop-1 AccTop-5 Acctime(ms)
bs=1		time(ms)
bs=4		time(ms)
bs=8		FLOPs(G)Params(M)Pretrained Model Download AddressInference Model Download Address
RedNet260.75950.93194.4515.1629.031.699.26Download linkDownload link
RedNet380.77470.93566.2421.3941.262.1412.43Download linkDownload link
RedNet500.78330.94178.0427.7153.732.6115.60Download linkDownload link
RedNet1010.78940.943613.0744.1283.284.5925.76Download linkDownload link
RedNet1520.79170.944018.6663.27119.486.5734.14Download linkDownload link

22. TNT series [35]

The accuracy and speed indicators of TNT series models are shown in the following table. For more introduction, please refer to: TNT series model documents.

ModelTop-1 AccTop-5 Acctime(ms)
bs=1		time(ms)
bs=4		FLOPs(G)Params(M)Pretrained Model Download AddressInference Model Download Address
TNT_small0.81210.95634.8323.68Download linkDownload link

Note: Both mean and std in the data preprocessing part of the TNT model NormalizeImage are 0.5.

23. CSWinTransformer series [40]

The accuracy and speed indicators of CSWinTransformer series models are shown in the following table. For more introduction, please refer to: CSWinTransformer series model documents。

ModelTop-1 AccTop-5 Acctime(ms)
bs=1		time(ms)
bs=4		time(ms)
bs=8		FLOPs(G)Params(M)Pretrained Model Download AddressInference Model Download Address
CSWinTransformer_tiny_2240.82810.9628---4.122Download linkDownload link
CSWinTransformer_small_2240.83580.9658---6.435Download linkDownload link
CSWinTransformer_base_2240.84200.9692---14.377Download linkDownload link
CSWinTransformer_large_2240.86430.9799---32.2173.3Download linkDownload link
CSWinTransformer_base_3840.85500.9749---42.277Download linkDownload link
CSWinTransformer_large_3840.87480.9833---94.7173.3Download linkDownload link

24. PVTV2 series [41]

The accuracy and speed indicators of PVTV2 series models are shown in the following table. For more introduction, please refer to: PVTV2 series model documents。

ModelTop-1 AccTop-5 Acctime(ms)
bs=1		time(ms)
bs=4		time(ms)
bs=8		FLOPs(G)Params(M)Pretrained Model Download AddressInference Model Download Address
PVT_V2_B00.7050.902---0.533.7Download linkDownload link
PVT_V2_B10.7870.945---2.014.0Download linkDownload link
PVT_V2_B20.8210.960---3.925.4Download linkDownload link
PVT_V2_B2_Linear0.8210.961---3.822.6Download linkDownload link
PVT_V2_B30.8310.965---6.745.2Download linkDownload link
PVT_V2_B40.8360.967---9.862.6Download linkDownload link
PVT_V2_B50.8370.966---11.482.0Download linkDownload link

25. MobileViT series [42]

The accuracy and speed indicators of MobileViT series models are shown in the following table. For more introduction, please refer to:MobileViT series model documents

ModelTop-1 AccTop-5 Acctime(ms)
bs=1		time(ms)
bs=4		time(ms)
bs=8		FLOPs(M)Params(M)Pretrained Model Download AddressInference Model Download Address
MobileViT_XXS0.68670.8878---337.241.28Download linkDownload link
MobileViT_XS0.74540.9227---930.752.33Download linkDownload link
MobileViT_S0.78140.9413---1849.355.59Download linkDownload link

26. Other models

The accuracy and speed indicators of AlexNet [18], SqueezeNet series [19], VGG series [20], DarkNet53 [21] and other models are shown in the following table. For more information, please refer to: Other model documents.

ModelTop-1 AccTop-5 Acctime(ms)
bs=1		time(ms)
bs=4		time(ms)
bs=8		FLOPs(G)Params(M)Pretrained Model Download AddressInference Model Download Address
AlexNet0.5670.7920.811.502.330.7161.10Download linkDownload link
SqueezeNet1_00.5960.8170.681.642.620.781.25Download linkDownload link
SqueezeNet1_10.6010.8190.621.302.090.351.24Download linkDownload link
VGG110.6930.8911.724.157.247.61132.86Download linkDownload link
VGG130.7000.8942.025.289.5411.31133.05Download linkDownload link
VGG160.7200.9072.486.7912.3315.470138.35Download linkDownload link
VGG190.7260.9092.938.2815.2119.63143.66Download linkDownload link
DarkNet530.7800.9412.796.4210.899.3141.65Download linkDownload link

Reference

[1] He K, Zhang X, Ren S, et al. Deep residual learning for image recognition[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 770-778.

[2] He T, Zhang Z, Zhang H, et al. Bag of tricks for image classification with convolutional neural networks[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2019: 558-567.

[3] Howard A, Sandler M, Chu G, et al. Searching for mobilenetv3[C]//Proceedings of the IEEE International Conference on Computer Vision. 2019: 1314-1324.

[4] Sandler M, Howard A, Zhu M, et al. Mobilenetv2: Inverted residuals and linear bottlenecks[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 4510-4520.

[5] Howard A G, Zhu M, Chen B, et al. Mobilenets: Efficient convolutional neural networks for mobile vision applications[J]. arXiv preprint arXiv:1704.04861, 2017.

[6] Ma N, Zhang X, Zheng H T, et al. Shufflenet v2: Practical guidelines for efficient cnn architecture design[C]//Proceedings of the European Conference on Computer Vision (ECCV). 2018: 116-131.

[7] Xie S, Girshick R, Dollár P, et al. Aggregated residual transformations for deep neural networks[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 1492-1500.

[8] Hu J, Shen L, Sun G. Squeeze-and-excitation networks[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 7132-7141.

[9] Gao S, Cheng M M, Zhao K, et al. Res2net: A new multi-scale backbone architecture[J]. IEEE transactions on pattern analysis and machine intelligence, 2019.

[10] Szegedy C, Liu W, Jia Y, et al. Going deeper with convolutions[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2015: 1-9.

[11] Szegedy C, Ioffe S, Vanhoucke V, et al. Inception-v4, inception-resnet and the impact of residual connections on learning[C]//Thirty-first AAAI conference on artificial intelligence. 2017.

[12] Chollet F. Xception: Deep learning with depthwise separable convolutions[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 1251-1258.

[13] Wang J, Sun K, Cheng T, et al. Deep high-resolution representation learning for visual recognition[J]. arXiv preprint arXiv:1908.07919, 2019.

[14] Chen Y, Li J, Xiao H, et al. Dual path networks[C]//Advances in neural information processing systems. 2017: 4467-4475.

[15] Huang G, Liu Z, Van Der Maaten L, et al. Densely connected convolutional networks[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 4700-4708.

[16] Tan M, Le Q V. Efficientnet: Rethinking model scaling for convolutional neural networks[J]. arXiv preprint arXiv:1905.11946, 2019.

[17] Mahajan D, Girshick R, Ramanathan V, et al. Exploring the limits of weakly supervised pretraining[C]//Proceedings of the European Conference on Computer Vision (ECCV). 2018: 181-196.

[18] Krizhevsky A, Sutskever I, Hinton G E. Imagenet classification with deep convolutional neural networks[C]//Advances in neural information processing systems. 2012: 1097-1105.

[19] Iandola F N, Han S, Moskewicz M W, et al. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and< 0.5 MB model size[J]. arXiv preprint arXiv:1602.07360, 2016.

[20] Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition[J]. arXiv preprint arXiv:1409.1556, 2014.

[21] Redmon J, Divvala S, Girshick R, et al. You only look once: Unified, real-time object detection[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 779-788.

[22] Ding X, Guo Y, Ding G, et al. Acnet: Strengthening the kernel skeletons for powerful cnn via asymmetric convolution blocks[C]//Proceedings of the IEEE International Conference on Computer Vision. 2019: 1911-1920.

[23] Han K, Wang Y, Tian Q, et al. GhostNet: More features from cheap operations[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020: 1580-1589.

[24] Zhang H, Wu C, Zhang Z, et al. Resnest: Split-attention networks[J]. arXiv preprint arXiv:2004.08955, 2020.

[25] Radosavovic I, Kosaraju R P, Girshick R, et al. Designing network design spaces[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020: 10428-10436.

[26] C.Szegedy, V.Vanhoucke, S.Ioffe, J.Shlens, and Z.Wojna. Rethinking the inception architecture for computer vision. arXiv preprint arXiv:1512.00567, 2015.

[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin and Baining Guo. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows.

[28]Cheng Cui, Tingquan Gao, Shengyu Wei, Yuning Du, Ruoyu Guo, Shuilong Dong, Bin Lu, Ying Zhou, Xueying Lv, Qiwen Liu, Xiaoguang Hu, Dianhai Yu, Yanjun Ma. PP-LCNet: A Lightweight CPU Convolutional Neural Network.

[29]Mingxing Tan, Quoc V. Le. MixConv: Mixed Depthwise Convolutional Kernels.

[30]Dongyoon Han, Sangdoo Yun, Byeongho Heo, YoungJoon Yoo. Rethinking Channel Dimensions for Efficient Model Design.

[31]Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby. AN IMAGE IS WORTH 16X16 WORDS: TRANSFORMERS FOR IMAGE RECOGNITION AT SCALE.

[32]Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, Herve Jegou. Training data-efficient image transformers & distillation through attention.

[33]Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Herve Jegou, Matthijs Douze. LeViT: a Vision Transformer in ConvNet’s Clothing for Faster Inference.

[34]Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, Chunhua Shen. Twins: Revisiting the Design of Spatial Attention in Vision Transformers.

[35]Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, Yunhe Wang. Transformer in Transformer.

[36]Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, Jian Sun. RepVGG: Making VGG-style ConvNets Great Again.

[37]Ping Chao, Chao-Yang Kao, Yu-Shan Ruan, Chien-Hsiang Huang, Youn-Long Lin. HarDNet: A Low Memory Traffic Network.

[38]Fisher Yu, Dequan Wang, Evan Shelhamer, Trevor Darrell. Deep Layer Aggregation.

[39]Duo Lim Jie Hu, Changhu Wang, Xiangtai Li, Qi She, Lei Zhu, Tong Zhang, Qifeng Chen. Involution: Inverting the Inherence of Convolution for Visual Recognition.

[40]Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, Baining Guo. CSWin Transformer: A General Vision Transformer Backbone with Cross-Shaped Windows.

[41]Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao. PVTv2: Improved Baselines with Pyramid Vision Transformer.

[42]Sachin Mehta, Mohammad Rastegari. MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer.