List your company
ResearchFormerly WizWand

Image Classification on ImageNet-1K linear evaluation 1.0

84.8Top-1 AccCLIP-VIT-L/14
51.57260.198568.82577.4515May 10, 2021Dec 9, 2021Jul 10, 2022Feb 9, 2023Sep 10, 2023Apr 10, 2024Nov 10, 2024
Updated 8mo ago

Evaluation results

MethodLinks
CLIP-VIT-L/14Backbone=ViT-L/14, Pre-trained=ImageNet-21k, Evaluation Protocol=linear probe, # Train Params (M)=321, # Inference Params (M)=3212024.1184.8—
TinyVit-21MBackbone=TinyVit-21M, Over-parameterization method=OPDF, Pre-trained=ImageNet-21k, Evaluation Protocol=linear probe, # Train Params (M)=38, # Inference Params (M)=212024.118497.5
TinyVit-21MBackbone=TinyVit-21M, Over-parameterization method=SVD, Pre-trained=ImageNet-21k, Evaluation Protocol=linear probe, # Train Params (M)=29, # Inference Params (M)=212024.1182.996.8
TinyVit-11MBackbone=TinyVit-11M, Over-parameterization method=OPDF, Pre-trained=ImageNet-21k, Evaluation Protocol=linear probe, # Train Params (M)=23, # Inference Params (M)=112024.1182.596.9
TinyVit-21MBackbone=TinyVit-21M, Over-parameterization method=None, Pre-trained=ImageNet-21k, Evaluation Protocol=linear probe, # Train Params (M)=21, # Inference Params (M)=212024.1182.396.3
TinyVit-11MBackbone=TinyVit-11M, Over-parameterization method=SVD, Pre-trained=ImageNet-21k, Evaluation Protocol=linear probe, # Train Params (M)=17, # Inference Params (M)=112024.118296.7
SupervisedBackbone=Swin-T, Training Epochs=300, Parameters (M)=29, FLOPs (G)=4.5, Throughput (img/s)=755.2, Evaluation Protocol=Linear evaluation2021.0581.3—
TinyVit-11MBackbone=TinyVit-11M, Over-parameterization method=None, Pre-trained=ImageNet-21k, Evaluation Protocol=linear probe, # Train Params (M)=11, # Inference Params (M)=112024.1180.595.6
TinyVit-5MBackbone=TinyVit-5M, Over-parameterization method=OPDF, Pre-trained=ImageNet-21k, Evaluation Protocol=linear probe, # Train Params (M)=9.9, # Inference Params (M)=5.42024.118096.7
SupervisedBackbone=DeiT-S, Training Epochs=300, Parameters (M)=22, FLOPs (G)=4.6, Throughput (img/s)=940.4, Evaluation Protocol=Linear evaluation2021.0579.8—
TinyVit-5MBackbone=TinyVit-5M, Over-parameterization method=SVD, Pre-trained=ImageNet-21k, Evaluation Protocol=linear probe, # Train Params (M)=7.6, # Inference Params (M)=5.42024.1177.995.1
TinyVit-5MBackbone=TinyVit-5M, Over-parameterization method=None, Pre-trained=ImageNet-21k, Evaluation Protocol=linear probe, # Train Params (M)=5.4, # Inference Params (M)=5.42024.1177.494.1
DINOBackbone=DeiT-S, Training Epochs=300, Parameters (M)=22, FLOPs (G)=4.6, Throughput (img/s)=940.4, multi-crop scheme=true, Evaluation Protocol=Linear evaluation2021.0575.9—
MoBYBackbone=Swin-T, Training Epochs=300, Parameters (M)=29, FLOPs (G)=4.5, Throughput (img/s)=755.2, Evaluation Protocol=Linear evaluation2021.0575—
MoBYBackbone=DeiT-S, Training Epochs=300, Parameters (M)=22, FLOPs (G)=4.6, Throughput (img/s)=940.4, Evaluation Protocol=Linear evaluation2021.0572.8—
MoCo v3Backbone=DeiT-S, Training Epochs=300, Parameters (M)=22, FLOPs (G)=4.6, Throughput (img/s)=940.4, Evaluation Protocol=Linear evaluation2021.0572.5—
DINOBackbone=DeiT-S, Training Epochs=300, Parameters (M)=22, FLOPs (G)=4.6, Throughput (img/s)=940.4, Evaluation Protocol=Linear evaluation2021.0572.5—
MoBYBackbone=Swin-T, Training Epochs=100, Parameters (M)=29, FLOPs (G)=4.5, Throughput (img/s)=755.2, Evaluation Protocol=Linear evaluation2021.0570.9—
MoCo.V2(C) + SAMix-CCL method=MoCo.V2(C), Mixup Variant=SAMix-C, Backbone=ResNet-502021.1168.86—
MoCo.V2 + SAMix-IPCL method=MoCo.V2, Mixup Variant=SAMix-IP, Backbone=ResNet-502021.1168.82—
MoCo.V2 + SAMix-ICL method=MoCo.V2, Mixup Variant=SAMix-I, Backbone=ResNet-502021.1168.76—
UnMix+ + Mixup+latentCL method=UnMix+, Mixup Variant=Mixup+latent, Backbone=ResNet-502021.1168.6—
MoCo.V2(C) + PuzzleMix*CL method=MoCo.V2(C), Mixup Variant=PuzzleMix*, Backbone=ResNet-502021.1168.48—
WBSIM+ + Mixup+CutMixCL method=WBSIM+, Mixup Variant=Mixup+CutMix, Backbone=ResNet-502021.1168.4—
MoCo.V2 + SaliencyMixCL method=MoCo.V2, Mixup Variant=SaliencyMix, Backbone=ResNet-502021.1168.31—
MoCo.V2 + CutMixCL method=MoCo.V2, Mixup Variant=CutMix, Backbone=ResNet-502021.1168.28—
i-Mix+ + Mixup+latentCL method=i-Mix+, Mixup Variant=Mixup+latent, Backbone=ResNet-502021.1168.1—
MoCo.V2 + MixupCL method=MoCo.V2, Mixup Variant=Mixup, Backbone=ResNet-502021.1168.07—
MoCHi + Mixup+latentCL method=MoCHi, Mixup Variant=Mixup+latent, Backbone=ResNet-502021.1168.01—
MoCo.V2CL method=MoCo.V2, Backbone=ResNet-502021.1167.66—
MoCo.V2(C) + SAMix-CCL method=MoCo.V2(C), Mixup Variant=SAMix-C, Backbone=ResNet-182021.1153.93—
MoCo.V2 + SAMix-ICL method=MoCo.V2, Mixup Variant=SAMix-I, Backbone=ResNet-182021.1153.75—
MoCo.V2 + SAMix-IPCL method=MoCo.V2, Mixup Variant=SAMix-IP, Backbone=ResNet-182021.1153.72—
MoCo.V2(C) + PuzzleMix*CL method=MoCo.V2(C), Mixup Variant=PuzzleMix*, Backbone=ResNet-182021.1153.46—
MoCHi + Mixup+latentCL method=MoCHi, Mixup Variant=Mixup+latent, Backbone=ResNet-182021.1153.12—
i-Mix+ + Mixup+latentCL method=i-Mix+, Mixup Variant=Mixup+latent, Backbone=ResNet-182021.1153.09—
MoCo.V2 + SaliencyMixCL method=MoCo.V2, Mixup Variant=SaliencyMix, Backbone=ResNet-182021.1153.06—
MoCo.V2 + MixupCL method=MoCo.V2, Mixup Variant=Mixup, Backbone=ResNet-182021.1153.03—
MoCo.V2 + CutMixCL method=MoCo.V2, Mixup Variant=CutMix, Backbone=ResNet-182021.1152.98—
MoCo.V2CL method=MoCo.V2, Backbone=ResNet-182021.1152.85—