Image Classification on ImageNet 1% ILSVRC-2012 (val)
91.5Top-5 AccSimCLRv2* (Fine-tune+Distillation)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SimCLRv2* (Fine-tune+Distillation)Self-supervised Pre-training=SimCLRv2*, Evaluation Protocol=Fine-tune+Distillation, #Epochs=>1200, #Parameters (train/test)=829.2M / 29.8M2020.11 | 91.5 | 73.9 | |
| MoCov2 (CoMatch)Self-supervised Pre-training=MoCov2, Evaluation Protocol=CoMatch, #Epochs=~1200, #Parameters (train/test)=30.0M / 25.6M2020.11 | 87.1 | 67.1 | |
| CoMatchSelf-supervised Pre-training=None, #Epochs=~400, #Parameters (train/test)=30.0M / 25.6M2020.11 | 86.4 | 66 | |
| SimCLRv2* (Fine-tune)Self-supervised Pre-training=SimCLRv2*, Evaluation Protocol=Fine-tune, #Epochs=~800, #Parameters (train/test)=34.2M / 29.8M2020.11 | 82.5 | 57.9 | |
| MoCov2 (FixMatch w. DA)Self-supervised Pre-training=MoCov2, Evaluation Protocol=FixMatch w. DA, #Epochs=~1200, #Parameters (train/test)=30.0M / 25.6M2020.11 | 79.8 | 59.9 | |
| BARLOW TWINSBackbone=ResNet-50, Evaluation protocol=Semi-supervised fine-tuning2021.03 | 79.2 | 55 | |
| SWAVBackbone=ResNet-50, Evaluation protocol=Semi-supervised fine-tuning2021.03 | 78.5 | 53.9 | |
| SwAVSelf-supervised Pre-training=SwAV, Evaluation Protocol=Fine-tune, #Epochs=~800, #Parameters (train/test)=30.4M / 25.6M2020.11 | 78.5 | 53.9 | |
| BYOLBackbone=ResNet-50, Evaluation protocol=Semi-supervised fine-tuning2021.03 | 78.4 | 53.2 | |
| BYOLSelf-supervised Pre-training=BYOL, Evaluation Protocol=Fine-tune, #Epochs=~1000, #Parameters (train/test)=37.1M / 25.6M2020.11 | 78.4 | 53.2 | |
| MoCov2 (Fine-tune)Self-supervised Pre-training=MoCov2, Evaluation Protocol=Fine-tune, #Epochs=~800, #Parameters (train/test)=30.0M / 25.6M2020.11 | 77.2 | 49.8 | |
| SIMCLRBackbone=ResNet-50, Evaluation protocol=Semi-supervised fine-tuning2021.03 | 75.5 | 48.3 | |
| SimCLRSelf-supervised Pre-training=SimCLR, Evaluation Protocol=Fine-tune, #Epochs=~1000, #Parameters (train/test)=30.0M / 25.6M2020.11 | 75.5 | 48.3 | |
| PCLSelf-supervised Pre-training=PCL, #Epochs=~200, #Parameters (train/test)=25.8M / 25.6M2020.11 | 75.3 | — | |
| FixMatch w. DASelf-supervised Pre-training=None, #Epochs=~400, #Parameters (train/test)=25.6M / 25.6M2020.11 | 74.4 | 53.4 | |
| PIRLBackbone=ResNet-50, Evaluation protocol=Semi-supervised fine-tuning2021.03 | 57.2 | — | |
| PIRLSelf-supervised Pre-training=PIRL, #Epochs=~800, #Parameters (train/test)=26.1M / 25.6M2020.11 | 57.2 | 30.7 | |
| S4L-RotationBackbone=ResNet50v22019.05 | 53.37 | — | |
| Pseudo-labelSelf-supervised Pre-training=None, #Epochs=~100, #Parameters (train/test)=25.6M / 25.6M2020.11 | 51.6 | — | |
| PseudolabelsBackbone=ResNet50v22019.05 | 51.56 | — | |
| Supervised BaselineBackbone=ResNet50v22019.05 | 48.43 | — | |
| SupervisedBackbone=ResNet-50, Evaluation protocol=Fine-tuning2021.03 | 48.4 | 25.4 | |
| Supervised baselineSelf-supervised Pre-training=None, #Epochs=~20, #Parameters (train/test)=25.6M / 25.6M2020.11 | 48.4 | 25.4 | |
| S4L-ExemplarBackbone=ResNet50v22019.05 | 47.02 | — | |
| VAT + Entropy MinimizationBackbone=ResNet50v22019.05 | 46.96 | — | |
| Self-sup. RotationBackbone=ResNet50v2, Evaluation Protocol=Fine-tune2019.05 | 45.11 | — | |
| Self-sup. ExemplarBackbone=ResNet50v2, Evaluation Protocol=Fine-tune2019.05 | 44.9 | — | |
| VATBackbone=ResNet50v22019.05 | 44.05 | — | |
| Self-sup. RotationBackbone=ResNet50v2, Evaluation Protocol=Linear2019.05 | 25.98 | — | |
| Self-sup. ExemplarBackbone=ResNet50v2, Evaluation Protocol=Linear2019.05 | 21.33 | — |