Image Classification on CIFAR-10 (Top-1 Accuracy)
99.3Top-1 AccuracyMugs
Evaluation results
| Method | Links | |
|---|---|---|
| MugsBackbone=ViT-B/162022.03 | 99.3 | |
| MugsBackbone=ViT-S/162022.03 | 99.2 | |
| iBOTBackbone=ViT-B/162022.03 | 99.2 | |
| DearKD-BStudent Network=DeiT-B2022.04 | 99.2 | |
| DeiT-B#Param=86M, FLOPs=55.4G, Model Type=Trans, Design Type=Manual, Resolution=384x3842021.07 | 99.1 | |
| AutoFormer-S#Param=23M, FLOPs=16.5G, Model Type=Trans, Design Type=Auto, Resolution=384x3842021.07 | 99.1 | |
| iBOTBackbone=ViT-S/162022.03 | 99.1 | |
| DINOBackbone=ViT-B/162022.03 | 99.1 | |
| DeiT-Bvariant=12022.04 | 99.1 | |
| DeiT-Bvariant=22022.04 | 99.1 | |
| Sup.Backbone=ViT-S/162022.03 | 99 | |
| DINOBackbone=ViT-S/162022.03 | 99 | |
| Sup.Backbone=ViT-B/162022.03 | 99 | |
| BEITBackbone=ViT-B/162022.03 | 99 | |
| MoCo-v3Backbone=ViT-B/162022.03 | 98.9 | |
| EfficientNet-B5#Param=30M, FLOPs=9.5G, Model Type=CNN, Design Type=Auto2021.07 | 98.7 | |
| BEITBackbone=ViT-S/162022.03 | 98.6 | |
| DearKD-SStudent Network=DeiT-S2022.04 | 98.4 | |
| T2T-ViT-142022.04 | 98.3 | |
| ViT-B/16#Param=86M, FLOPs=55.4G, Model Type=Trans, Design Type=Manual2021.07 | 98.1 | |
| ViT-B/162022.04 | 98.1 | |
| EffiNet-B52022.04 | 98.1 | |
| TWISTEvaluation Protocol=Fine-tune, Backbone=ResNet-50, Pre-trained=ImageNet2021.10 | 97.9 | |
| ViT-L/322022.04 | 97.9 | |
| ViT-L/162022.04 | 97.9 | |
| BYOLEvaluation Protocol=Fine-tune, Backbone=ResNet-50, Pre-trained=ImageNet2021.10 | 97.8 | |
| ViT-B/322022.04 | 97.8 | |
| SUPEvaluation Protocol=Fine-tune, Backbone=ResNet-50, Pre-trained=ImageNet2021.10 | 97.5 | |
| DearKD-TiStudent Network=DeiT-Ti2022.04 | 97.5 | |
| SimCLREvaluation Protocol=Fine-tune, Backbone=ResNet-50, Pre-trained=ImageNet2021.10 | 97.4 | |
| NASNet-AGPU (days)=2000, Params (M)=3.32026.02 | 97.35 | |
| NNCLRArch=ViT-B/82021.04 | 96.8 | |
| NNCLR + SupArch=ViT-B/82021.04 | 96.2 | |
| Sup. INArch=ViT-B/82021.04 | 95.9 | |
| RandomEvaluation Protocol=Fine-tune, Backbone=ResNet-50, Pre-trained=ImageNet2021.10 | 95.9 | |
| NNCLRArch=ViT-B/162021.04 | 95.7 | |
| Sup. INArch=ViT-B/162021.04 | 95.2 | |
| PyramidCLIPPretrain Dataset=143M, Backbone=ResNet-50, Evaluation Protocol=End-to-end fine-tuning2022.04 | 95.2 | |
| CLIP*Pretrain Dataset=400M, Backbone=ResNet-50, Evaluation Protocol=End-to-end fine-tuning2022.04 | 95 | |
| Supervised(IN1K)Pretrain Dataset=1.2M, Backbone=ResNet-50, Evaluation Protocol=End-to-end fine-tuning2022.04 | 94 | |
| NNCLRArch=R502021.04 | 93.7 | |
| Sup. INArch=R502021.04 | 93.6 | |
| SUPEvaluation Protocol=Linear evaluation, Backbone=ResNet-50, Pre-trained=ImageNet2021.10 | 93.6 | |
| QimeraModel=ResNet-20, FP32 Acc.=93.89, Bits=5w5a2021.11 | 93.46 | |
| GDFQModel=ResNet-20, FP32 Acc.=93.89, Bits=5w5a2021.11 | 93.38 | |
| ZAQModel=ResNet-20, FP32 Acc.=93.89, Bits=5w5a2021.11 | 93.36 | |
| GDFQ+MixupModel=ResNet-20, FP32 Acc.=93.89, Bits=5w5a2021.11 | 92.79 | |
| GDFQ+CutmixModel=ResNet-20, FP32 Acc.=93.89, Bits=5w5a2021.11 | 92.75 | |
| ZAQModel=ResNet-20, FP32 Acc.=93.89, Bits=4w4a2021.11 | 92.13 | |
| CPV-5/4# Params=1.3 M, GPU (GB)=3.2, MACs=1.37 G, Batch Size=322022.01 | 91.42 | |
| ZeroQModel=ResNet-20, FP32 Acc.=93.89, Bits=5w5a2021.11 | 91.34 | |
| BYOLArch=R502021.04 | 91.3 | |
| BYOLEvaluation Protocol=Linear evaluation, Backbone=ResNet-50, Pre-trained=ImageNet2021.10 | 91.3 | |
| QimeraModel=ResNet-20, FP32 Acc.=93.89, Bits=4w4a2021.11 | 91.26 | |
| TWISTEvaluation Protocol=Linear evaluation, Backbone=ResNet-50, Pre-trained=ImageNet2021.10 | 91.2 | |
| WaveMix-128/7#Param (Million)=2.42, GPU (GB)=1.32022.03 | 91.08 | |
| WaveMix-256/7#Param (Million)=9.62, GPU (GB)=2.32022.03 | 90.72 | |
| SimCLREvaluation Protocol=Linear evaluation, Backbone=ResNet-50, Pre-trained=ImageNet2021.10 | 90.6 | |
| SimCLRArch=R502021.04 | 90.5 | |
| GDFQModel=ResNet-20, FP32 Acc.=93.89, Bits=4w4a2021.11 | 90.25 | |
| GDFQ+CutmixModel=ResNet-20, FP32 Acc.=93.89, Bits=4w4a2021.11 | 89.58 | |
| CNV-5/4# Params=1.3 M, GPU (GB)=3.1, MACs=1.39 G, Batch Size=322022.01 | 89.56 | |
| GDFQ+MixupModel=ResNet-20, FP32 Acc.=93.89, Bits=4w4a2021.11 | 88.69 | |
| ConvMixer-16# Params=1.3 M, GPU (GB)=3.5, MACs=1.37 G, Batch Size=322022.01 | 88.46 | |
| ConvMixer-256/16#Param (Million)=1.3, GPU (GB)=7.02022.03 | 88.46 | |
| ResNet-34# Params=21.3 M, GPU (GB)=0.7, MACs=1.16 G, Batch Size=322022.01 | 87.97 | |
| ResNet-34#Param (Million)=21.3, GPU (GB)=1.42022.03 | 87.97 | |
| CLTV-5/4# Params=1.3 M, GPU (GB)=1.8, MACs=1.38 G, Batch Size=322022.01 | 86.99 | |
| ResNet-18# Params=11.2 M, GPU (GB)=0.6, MACs=0.56 G, Batch Size=322022.01 | 86.29 | |
| ResNet-18#Param (Million)=11.2, GPU (GB)=1.22022.03 | 86.29 | |
| ResNet-50#Param (Million)=25.2, GPU (GB)=3.32022.03 | 86.21 | |
| WaveMix-64/5#Param (Million)=2.88, GPU (GB)=0.32022.03 | 86.16 | |
| ConvMixer-256/8#Param (Million)=0.67, GPU (GB)=3.72022.03 | 85.41 | |
| WaveMix-5# Params=1.4 M, GPU (GB)=0.4, MACs=2.59 G, Batch Size=322022.01 | 83.71 | |
| CCT-6/4# Params=1.3 M, GPU (GB)=13.6, MACs=1.32 G, Batch Size=322022.01 | 82.66 | |
| CCT-128/4 × 4#Param (Million)=0.90, GPU (GB)=15.82022.03 | 82.23 | |
| WaveMix-32/5#Param (Million)=0.72, GPU (GB)=0.22022.03 | 81.47 | |
| OursDistortion=Random Missing Pixels (RM), Evaluation Protocol=Linear Probing2021.10 | 80.93 | |
| CVT-128/4 × 4#Param (Million)=1.10, GPU (GB)=15.42022.03 | 79.93 | |
| BaselineDistortion=Random Missing Pixels (RM), Evaluation Protocol=Linear Probing2021.10 | 79.83 | |
| ZeroQModel=ResNet-20, FP32 Acc.=93.89, Bits=4w4a2021.11 | 79.3 | |
| Hybrid ViLT-6/8# Params=1.3 M, GPU (GB)=2.6, MACs=0.33 G, Batch Size=322022.01 | 78.34 | |
| WaveMix-16/5#Param (Million)=0.18, GPU (GB)=0.22022.03 | 78.04 | |
| Hybrid ViN-6/8# Params=1.3 M, GPU (GB)=5.3, MACs=1.41 G, Batch Size=322022.01 | 77.96 | |
| CvT-5/4# Params=1.3 M, GPU (GB)=9.4, MACs=1.32 G, Batch Size=322022.01 | 77.9 | |
| Hybrid ViP-6/8# Params=1.3 M, GPU (GB)=5.9, MACs=1.31 G, Batch Size=322022.01 | 77.54 | |
| CV-DDBackbone=ResNet18, IPC (Images Per Class)=50, Ratio=1.0%2025.01 | 76.9 | |
| BaselineDistortion=Additive Gaussian Noise (GN), Evaluation Protocol=Linear Probing2021.10 | 76.52 | |
| OursDistortion=Additive Gaussian Noise (GN), Evaluation Protocol=Linear Probing2021.10 | 76.19 | |
| SRe2L++Backbone=ResNet18, IPC (Images Per Class)=50, Ratio=1.0%2025.01 | 75.8 | |
| Hybrid ViN-128/4 × 4#Param (Million)=0.62, GPU (GB)=4.82022.03 | 75.26 | |
| CV-DDBackbone=ResNet101, IPC (Images Per Class)=50, Ratio=1.0%2025.01 | 74.4 | |
| SRe2L++Backbone=ResNet101, IPC (Images Per Class)=50, Ratio=1.0%2025.01 | 73.6 | |
| CV-DDBackbone=ResNet50, IPC (Images Per Class)=50, Ratio=1.0%2025.01 | 72.1 | |
| SRe2L++Backbone=ResNet50, IPC (Images Per Class)=50, Ratio=1.0%2025.01 | 71.4 | |
| 2D FNet-128/5#Param (Million)=0.41, GPU (GB)=1.62022.03 | 70.52 | |
| 2D FNet-64/5#Param (Million)=0.10, GPU (GB)=1.02022.03 | 64.6 | |
| RDEDBackbone=ResNet18, IPC (Images Per Class)=50, Ratio=1.0%2025.01 | 62.1 | |
| MLP Mixer-5# Params=21.3 M, GPU (GB)=1.5, MACs=3.02 G, Batch Size=322022.01 | 60.26 | |
| ViT-10/4# Params=1.3 M, GPU (GB)=14.7, MACs=1.34 G, Batch Size=322022.01 | 57.53 |