Image Classification on ImageNet
90.2Top-1 AccuracyEffNet-L2 + Meta Pseudo Labels
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| EffNet-L2 + Meta Pseudo Labels#Params=480M, TPUv3-core-days=22.5k2021.02 | 90.2 | — | — | — | — | |
| Meta Pseudo Labels (EfficientNet-L2)# Params=480M, Extra Data=300M unlabeled JFT2020.03 | 90.2 | — | — | — | 98.8 | |
| Meta Pseudo Labels (EfficientNet-B6-Wide)# Params=390M, Extra Data=300M unlabeled JFT2020.03 | 90 | — | — | — | 98.7 | |
| NFNet-F4+#FLOPS=367B, #Params=527M, TPUv3-core-days=1.86k2021.02 | 89.2 | — | — | — | — | |
| NFNet-F4#FLOPS=215B, #Params=316M, TPUv3-core-days=3.7k2021.02 | 89.2 | — | — | — | — | |
| EffNet-L2 + NoisyStudent + SAM#Params=480M, TPUv3-core-days=12.3k2021.02 | 88.6 | — | — | — | — | |
| EfficientNet-L2-NoisyStudent + SAM# Params=480M, Extra Data=300M unlabeled JFT2020.03 | 88.6 | — | — | — | 98.6 | |
| ViT-H/14#Params=632M, TPUv3-core-days=2.5k2021.02 | 88.55 | — | — | — | — | |
| Vision Transformer (ViT-H)# Params=632M, Extra Data=300M labeled JFT2020.03 | 88.55 | — | — | — | — | |
| Noisy Student + FixRes# Params=480M, Extra Data=300M unlabeled JFT2020.03 | 88.5 | — | — | — | 98.7 | |
| Noisy Student (EfficientNet-L2)# Params=480M, Extra Data=300M unlabeled JFT2020.03 | 88.4 | — | — | — | 98.7 | |
| ViT-L/16#Params=307M, TPUv3-core-days=0.68k2021.02 | 87.76 | — | — | — | — | |
| BiT-L ResNet152x4#Params=928M, TPUv3-core-days=9.9k2021.02 | 87.54 | — | — | — | — | |
| Big Transfer (BiT-L)# Params=928M, Extra Data=300M labeled JFT2020.03 | 87.5 | — | — | — | 98.5 | |
| DINOv3-LModel Size=Large2026.03 | 86.7 | — | — | — | — | |
| ResNeXt-101 32x48d (IG-940M)#Params=829M2021.02 | 86.4 | — | — | — | — | |
| FixRes ResNeXt-101 WSL# Params=829M, Extra Data=3.5B labeled Instagram2020.03 | 86.4 | — | — | — | 98 | |
| LOUPEevaluation=linear probing2022.08 | 85.7 | — | — | — | — | |
| EfficientNet-L2# Params=480M, Extra Data=-2020.03 | 85.5 | — | — | — | 97.5 | |
| ResNeXt-101 WSL# Params=829M, Extra Data=3.5B labeled Instagram2020.03 | 85.4 | — | — | — | 97.6 | |
| EfficientNet-B7 + FixRes# Params=66M, Extra Data=-2020.03 | 85.3 | — | — | — | 97.4 | |
| EfficientNet-B7# Params=66M, Extra Data=-2020.03 | 85 | — | — | — | 97.2 | |
| ResNeXt-101 Billion-scale SSL# Params=193M, Extra Data=3.5B labeled Instagram2020.03 | 84.8 | — | — | — | — | |
| OmniStream2026.03 | 84.7 | — | — | — | — | |
| EfficientNet-B7#Params=66M, Ratio-to-EfficientNet (Params)=1x, #FLOPs=37B, Ratio-to-EfficientNet (FLOPs)=1x2019.05 | 84.3 | — | — | — | 97 | |
| GPipe#Params=557M, Ratio-to-EfficientNet (Params)=8.4x2019.05 | 84.3 | — | — | — | 97 | |
| GPipe# Params=557M, Extra Data=-2020.03 | 84.3 | — | — | — | 97 | |
| EfficientNet-B7FLOPs=37.3G, Resolution=224x224, Pre-training=ImageNet2021.03 | 84.3 | — | — | — | — | |
| EfficientNet-B6#Params=43M, Ratio-to-EfficientNet (Params)=1x, #FLOPs=19B, Ratio-to-EfficientNet (FLOPs)=1x2019.05 | 84 | — | — | — | 96.8 | |
| CLIPevaluation=linear probing2022.08 | 83.9 | — | — | — | — | |
| Florence-CoSwin-HBackbone=CoSwin-H, Resolution=384px, Evaluation Protocol=Zero-shot2021.11 | 83.7 | — | — | — | — | |
| EfficientNet-B5#Params=30M, Ratio-to-EfficientNet (Params)=1x, #FLOPs=9.9B, Ratio-to-EfficientNet (FLOPs)=1x2019.05 | 83.6 | — | — | — | 96.7 | |
| EfficientNet-B5FLOPs=10.3G, Resolution=224x224, Pre-training=ImageNet2021.03 | 83.6 | — | — | — | — | |
| EfficientNet-B5#Param=30M, FLOPs=9.5G, Model Type=CNN, Design Type=Auto2021.07 | 83.6 | — | — | — | — | |
| AmoebaNet-C#Params=155M, Ratio-to-EfficientNet (Params)=5.2x, #FLOPs=41B, Ratio-to-EfficientNet (FLOPs)=4.1x2019.05 | 83.5 | — | — | — | 96.5 | |
| AmoebaNet-C + AutoAugment# Params=155M, Extra Data=-2020.03 | 83.5 | — | — | — | 96.5 | |
| TF-TAS-S#Param=23M, Fine-tuned=true, Resolution=384x3842022.03 | 83.5 | — | — | — | — | |
| AutoFormer-S#Param=23M, FLOPs=16.5G, Model Type=Trans, Design Type=Auto, Resolution=384x3842021.07 | 83.4 | — | — | — | — | |
| AutoFormer-S#Param=23M, Fine-tuned=true, Resolution=384x3842022.03 | 83.4 | — | — | — | — | |
| CeiT-SFLOPs=12.9G, Resolution=384x384, Pre-training=ImageNet2021.03 | 83.3 | — | — | — | — | |
| Deit-BFLOPs=52.8G, Resolution=384x384, Pre-training=ImageNet2021.03 | 83.1 | — | — | — | — | |
| DeiT-B#Param=86M, FLOPs=55.4G, Model Type=Trans, Design Type=Manual, Resolution=384x3842021.07 | 83.1 | — | — | — | — | |
| DeiT-B#Param=86M, Fine-tuned=true, Resolution=384x3842022.03 | 83.1 | — | — | — | — | |
| EfficientNet-B4#Params=19M, Ratio-to-EfficientNet (Params)=1x, #FLOPs=4.2B, Ratio-to-EfficientNet (FLOPs)=1x2019.05 | 82.9 | — | — | — | 96.4 | |
| PNASNet#Params=86M, Ratio-to-EfficientNet (Params)=4.5x, #FLOPs=23B, Ratio-to-EfficientNet (FLOPs)=6.0x2019.05 | 82.9 | — | — | — | 96.2 | |
| PNASNet# Params=86M, Extra Data=-2020.03 | 82.9 | — | — | — | 96.2 | |
| AmoebaNet-A#Params=87M, Ratio-to-EfficientNet (Params)=4.6x, #FLOPs=23B, Ratio-to-EfficientNet (FLOPs)=5.5x2019.05 | 82.8 | — | — | — | 96.1 | |
| AmoebaNet-A# Params=87M, Extra Data=-2020.03 | 82.8 | — | — | — | 96.1 | |
| DINOBackbone=ViT-B/16, Pre-training=Self-supervised, Evaluation Protocol=finetuning2021.04 | 82.8 | — | — | — | — | |
| AMoEBudget=0.23TT, Head=Ensemble, Resolution=512x5122025.12 | 82.78 | — | — | — | — | |
| SENet#Params=146M, Ratio-to-EfficientNet (Params)=7.7x, #FLOPs=42B, Ratio-to-EfficientNet (FLOPs)=10x2019.05 | 82.7 | — | — | — | 96.2 | |
| NASNet-A#Params=89M, Ratio-to-EfficientNet (Params)=4.7x, #FLOPs=24B, Ratio-to-EfficientNet (FLOPs)=5.7x2019.05 | 82.7 | — | — | — | 96.2 | |
| SENet# Params=146M, Extra Data=-2020.03 | 82.7 | — | — | — | 96.2 | |
| NASNet-A# Params=89M, Extra Data=-2020.03 | 82.7 | — | — | — | 96.2 | |
| RADIOv2.5-H (0.6B)Budget=1.1TT, Head=Ensemble, Resolution=512x5122025.12 | 82.22 | — | — | — | — | |
| CeiT-SFLOPs=4.5G, Resolution=224x224, Pre-training=ImageNet2021.03 | 82 | — | — | — | — | |
| Deit-BFLOPs=17.3G, Resolution=224x224, Pre-training=ImageNet2021.03 | 81.8 | — | — | — | — | |
| SupervisedBackbone=ViT-B/16, Pre-training=Supervised, Evaluation Protocol=finetuning2021.04 | 81.8 | — | — | — | — | |
| EfficientNet-B3#Params=12M, Ratio-to-EfficientNet (Params)=1x, #FLOPs=1.8B, Ratio-to-EfficientNet (FLOPs)=1x2019.05 | 81.6 | — | — | — | 95.7 | |
| DINOBackbone=ViT-S/16, Pre-training=Self-supervised, Evaluation Protocol=finetuning2021.04 | 81.5 | — | — | — | — | |
| PolyNet#Params=92M, Ratio-to-EfficientNet (Params)=7.7x, #FLOPs=35B, Ratio-to-EfficientNet (FLOPs)=19x2019.05 | 81.3 | — | — | — | 95.8 | |
| PolyNet# Params=92M, Extra Data=-2020.03 | 81.3 | — | — | — | 95.8 | |
| ViT-S/8-SAM#params=22M, Throughput (img/sec/core)=333, Resolution=2242021.06 | 81.3 | — | — | — | — | |
| ResNet-50 Billion-scale SSL# Params=26M, Extra Data=3.5B labeled Instagram2020.03 | 81.2 | — | — | — | 96 | |
| ResNet-152x2-SAM#params=236M, Throughput (img/sec/core)=356, Resolution=2242021.06 | 81.1 | — | — | — | — | |
| ResNeXt-101#Params=84M, Ratio-to-EfficientNet (Params)=7.0x, #FLOPs=32B, Ratio-to-EfficientNet (FLOPs)=18x2019.05 | 80.9 | — | — | — | 95.6 | |
| ResNeXt-101# Params=84M, Extra Data=-2020.03 | 80.9 | — | — | — | 95.6 | |
| ResNet-101x2-SAM#params=173M, Throughput (img/sec/core)=519, Resolution=2242021.06 | 80.9 | — | — | — | — | |
| EfficientNet-B2#Params=9.2M, Ratio-to-EfficientNet (Params)=1x, #FLOPs=1.0B, Ratio-to-EfficientNet (FLOPs)=1x2019.05 | 80.1 | — | — | — | 94.9 | |
| Inception-resnet-v2#Params=56M, Ratio-to-EfficientNet (Params)=6.1x, #FLOPs=13B, Ratio-to-EfficientNet (FLOPs)=13x2019.05 | 80.1 | — | — | — | 95.1 | |
| Inception-resnet-v2# Params=56M, Extra Data=-2020.03 | 80.1 | — | — | — | 95.1 | |
| Inception-v4#Params=48M, Ratio-to-EfficientNet (Params)=5.2x, #FLOPs=13B, Ratio-to-EfficientNet (FLOPs)=13x2019.05 | 80 | — | — | — | 95 | |
| Inception-v4# Params=48M, Extra Data=-2020.03 | 80 | — | — | — | 95 | |
| SupervisedBackbone=ViT-S/16, Pre-training=Supervised, Evaluation Protocol=finetuning2021.04 | 79.9 | — | — | — | — | |
| ViT-B/16-SAM#params=87M, Throughput (img/sec/core)=863, Resolution=2242021.06 | 79.9 | — | — | — | — | |
| Grafit ResNet-50FLOPs=4.1G, Resolution=224x224, Pre-training=ImageNet2021.03 | 79.6 | — | — | — | — | |
| Grafit ResNet-50#Param=25M, FLOPs=12.1G, Model Type=CNN, Design Type=Manual2021.07 | 79.6 | — | — | — | — | |
| Grafit RegNetY-8GF#Param=39M, FLOPs=23.4G, Model Type=CNN, Design Type=Manual2021.07 | 79.6 | — | — | — | — | |
| ResNet-50x2-SAM#params=98M, Throughput (img/sec/core)=891, Resolution=2242021.06 | 79.6 | — | — | — | — | |
| ResNet-152-SAM#params=60M, Throughput (img/sec/core)=935, Resolution=2242021.06 | 79.3 | — | — | — | — | |
| EfficientNet-B1#Params=7.8M, Ratio-to-EfficientNet (Params)=1x, #FLOPs=0.70B, Ratio-to-EfficientNet (FLOPs)=1x2019.05 | 79.1 | — | — | — | 94.4 | |
| DK + DCNBackbone=ResNet-50-DW, 1×1 Deformable Kernels=w/o, 3×3 Deformable Kernels=local, scope size 4×4, #P (M)=26.1, GFLOPS=4.602019.10 | 79 | — | — | — | — | |
| Xception#Params=23M, Ratio-to-EfficientNet (Params)=3.0x, #FLOPs=8.4B, Ratio-to-EfficientNet (FLOPs)=12x2019.05 | 79 | — | — | — | 94.5 | |
| Xception# Params=23M, Extra Data=-2020.03 | 79 | — | — | — | 94.5 | |
| CutMix + MoExBackbone=ResNet-502020.10 | 79 | — | — | — | — | |
| Mixer-B/8-SAM#params=64M, Throughput (img/sec/core)=466, Resolution=2242021.06 | 79 | — | — | — | — | |
| DK + SCCBackbone=ResNet-50-DW, 1×1 Deformable Kernels=w/o, 3×3 Deformable Kernels=local, scope size 4×4, #P (M)=43.7, GFLOPS=7.612019.10 | 78.9 | — | — | — | — | |
| DSGModel Architecture=InceptionV3, No D=true, No FT=true, W-bit=8, A-bit=82021.03 | 78.81 | — | — | — | — | |
| ResNeXt-101#.Param.=46.66M, FLOPS=7.33G2019.10 | 78.8 | — | — | — | 94.4 | |
| Inception-v3#Params=24M, Ratio-to-EfficientNet (Params)=3.0x, #FLOPs=5.7B, Ratio-to-EfficientNet (FLOPs)=8.1x2019.05 | 78.8 | — | — | — | 94.4 | |
| Inception-v3# Params=24M, Extra Data=-2020.03 | 78.8 | — | — | — | 94.4 | |
| CeiT-TFLOPs=3.6G, Resolution=384x384, Pre-training=ImageNet2021.03 | 78.8 | — | — | — | — | |
| BaselineModel Architecture=InceptionV3, W-bit=32, A-bit=322021.03 | 78.8 | — | — | — | — | |
| ViT-S/14-SAM#params=22M, Throughput (img/sec/core)=1234, Resolution=2242021.06 | 78.8 | — | — | — | — | |
| Real DataModel Architecture=InceptionV3, No D=false, No FT=true, W-bit=8, A-bit=82021.03 | 78.78 | — | — | — | — | |
| ZeroQModel Architecture=InceptionV3, No D=true, No FT=true, W-bit=8, A-bit=82021.03 | 78.7 | — | — | — | — | |
| ECA-Net101#.Param.=42.49M, FLOPS=7.35G2019.10 | 78.65 | — | — | — | 94.31 | |
| ResNet-101-SAM#params=44M, Throughput (img/sec/core)=1334, Resolution=2242021.06 | 78.6 | — | — | — | — | |
| FLIP-ViT-L/14Backbone=ViT-L/14, Evaluation Protocol=Zero-shot2021.11 | 78.3 | — | — | — | — | |
| ResNet-200#.Param.=74.45M, FLOPS=14.10G2019.10 | 78.2 | — | — | — | 94 |