List your company
ResearchFormerly WizWand

Image Classification on ImageNet (val) (Top-1 Accuracy)

89.6Top-1 AccuracyBEIT-3
64.1270.73577.3583.965Apr 1, 2019May 26, 2020Jul 22, 2021Sep 17, 2022Nov 12, 2023Jan 7, 2025Mar 5, 2026
Updated 7mo ago

Evaluation results

MethodLinks
BEIT-3publicly accessible resources only=true2022.0889.6
FD-CLIPpublicly accessible resources only=true2022.0889
ResNet-RS-152Resolution=224, Training Protocol=Semi-supervised, V100 latency (s)=1.48, TPUv3 latency (ms)=3202021.0386.2
EfficientNet-B5Training Protocol=Semi-supervised (Noisy Student), V100 latency (s)=8.16, TPUv3 latency (ms)=15102021.0386.1
BASICZero-shot=true2021.1185.7
EfficientNet-B7Resolution=600, Params (M)=66, FLOPs (B)=74, TPU-v3 Latency (s)=6.0, TPU-v3 Memory (GB)=28.3, V100 Latency (s)=29.92021.0384.7
ResNet-RS-420Resolution=320, Params (M)=192, FLOPs (B)=128, TPU-v3 Latency (s)=2.1, TPU-v3 Memory (GB)=15.5, V100 Latency (s)=10.22021.0384.4
ResNet-RS-350Resolution=256, Params (M)=164, FLOPs (B)=69, TPU-v3 Latency (s)=1.1, TPU-v3 Memory (GB)=7.3, V100 Latency (s)=4.72021.0384
EfficientNet-B6Resolution=528, Params (M)=43, FLOPs (B)=38, TPU-v3 Latency (s)=3.0, TPU-v3 Memory (GB)=16.6, V100 Latency (s)=15.72021.0384
ViT-B/162023.0381.2
ViT-L/162023.0380.4
Diffusion ClassifierResolution=512x5122023.0379.1
ReLabelBackbone=ResNet-502026.0378.9
BaselineArchitecture=InceptionV3, W-bit=32, A-bit=32, Size (MB)=90.922020.0178.88
ZEROQArchitecture=InceptionV3, No Data=true, No Fine-tuning=true, W-bit=8, A-bit=8, Size (MB)=22.472020.0178.81
ReLabel w/ Our MaskBackbone=ResNet-502026.0378.8
ZEROQArchitecture=InceptionV3, No Data=true, No Fine-tuning=true, W-bit=MP, A-bit=6, Size (MB)=17.022020.0178.76
Multi-labelBackbone=ResNet-502026.0378.7
Original + Label SmoothBackbone=ResNet-502026.0378.2
ViT-L/322023.0377.9
LLBackbone=ResNet-502026.0377.8
ResNet-1012023.0377.7
Original LabelBackbone=ResNet-502026.0377.6
ZEROQArchitecture=InceptionV3, No Data=true, No Fine-tuning=true, W-bit=MP, A-bit=8, Size (MB)=11.352020.0177.57
Diffusion ClassifierResolution=256x2562023.0377.5
SCLBackbone=ResNet-502026.0376.9
ResNet-502023.0376.7
LAMBoptimizer=LAMB, Batch Size=16K, Training Epochs=90, Backbone=ResNet-502019.0476.66
ALIGNZero-shot=true2021.1176.4
CLIPZero-shot=true2021.1176.2
DRConv-MobileNetV2Setting=8 regions, MADDs=328M2020.0375.7
AdCoEpochs=800, Evaluation protocol=linear, Multi-crop=true2022.1275.7
NNCLREpochs=1000, Evaluation protocol=linear, Multi-crop=true2022.1275.6
DRConv-MobileNetV1Setting=8 regions, MADDs=610M2020.0375.5
OBQBackbone=ResNet50, Bitwidth=4bit, Quantization Scheme=Uniform symmetric per-channel weight quantization2022.0875.5
UniGradEpochs=800, Evaluation protocol=linear, Multi-crop=true2022.1275.5
SCEEpochs=200, Evaluation protocol=linear, Multi-crop=true2022.1275.4
SwaVEpochs=800, Evaluation protocol=linear, Multi-crop=true2022.1275.3
DINOEpochs=800, Evaluation protocol=linear, Multi-crop=true2022.1275.3
Momentumoptimizer=Momentum, Batch Size=16K, Training Epochs=90, Backbone=ResNet-502019.0475.2
BitSplitBackbone=ResNet50, Bitwidth=4bit, Quantization Scheme=Uniform symmetric per-channel weight quantization2022.0874.94
DK-MobileNetV2Setting=global 2x2 & local 4x4, MADDs=760M2020.0374.8
DRConv-MobileNetV2*Setting=8 regions, MADDs=201M, fewer computational cost=true2020.0374.7
ReSSLEpochs=200, Evaluation protocol=linear, Multi-crop=true2022.1274.7
WCLEpochs=800, Evaluation protocol=linear, Multi-crop=true2022.1274.7
CondConv-MobileNetV2Setting=8 weights, MADDs=329M2020.0374.6
DRConv-MobileNetV1*Setting=8 regions, MADDs=344M, fewer computational cost=true2020.0374.4
RVQuantArchitecture=InceptionV3, No Data=false, No Fine-tuning=false, W-bit=8, A-bit=8, Size (MB)=22.472020.0174.22
TripletEpochs=200, Evaluation protocol=linear, Multi-crop=true2022.1274.1
ResNet-342023.0373.8
CondConv-MobileNetV1Setting=8 weights, MADDs=600M2020.0373.7
OBQBackbone=ResNet50, Bitwidth=3bit, Quantization Scheme=Uniform symmetric per-channel weight quantization2022.0873.61
Adamoptimizer=Adam, Batch Size=16K, Training Epochs=90, Backbone=ResNet-50, Learning Rate Scheme=Goyal et al. (2017)2019.0473.48
WCLEpochs=200, Evaluation protocol=linear, Multi-crop=true2022.1273.3
AdCoEpochs=200, Evaluation protocol=linear, Multi-crop=true2022.1273.2
DRConv-ShuffleNetV2 1xSetting=8 regions, MADDs=157M2020.0373.1
AdamWoptimizer=AdamW, Batch Size=16K, Training Epochs=90, Backbone=ResNet-50, Learning Rate Scheme=Goyal et al. (2017)2019.0473.07
OBQBackbone=ResNet34, Bitwidth=4bit, Quantization Scheme=Uniform symmetric per-channel weight quantization2022.0872.85
AdaQuantBackbone=ResNet50, Bitwidth=4bit, Quantization Scheme=Uniform symmetric per-channel weight quantization2022.0872.79
SwaVEpochs=200, Evaluation protocol=linear, Multi-crop=true2022.1272.7
UniGradEpochs=100, Evaluation protocol=linear, Multi-crop=true2022.1272.3
DRConv-ShuffleNetV2 1x*Setting=8 regions, MADDs=109M, fewer computational cost=true2020.0372.2
Adagradoptimizer=Adagrad, Batch Size=16K, Training Epochs=90, Backbone=ResNet-50, Learning Rate Scheme=Goyal et al. (2017)2019.0472.01
CondConv-ShuffleNetV2 1xSetting=8 weights, MADDs=152M2020.0372
MobileNetV2Setting=baseline, MADDs=300M2020.0372
MobileNetV1 (Dense baseline)Backbone=MobileNetV1, Sparsity Ratio=0%2023.1071.95
BitSplitBackbone=ResNet50, Bitwidth=3bit, Quantization Scheme=Uniform symmetric per-channel weight quantization2022.0871.76
BitSplitBackbone=ResNet34, Bitwidth=4bit, Quantization Scheme=Uniform symmetric per-channel weight quantization2022.0871.63
RELICv2Epochs=5000, Pre-training dataset=JFT-300M, Evaluation protocol=linear evaluation, Backbone=ResNet-502022.0171.4
OCSArchitecture=InceptionV3, No Data=false, No Fine-tuning=true, W-bit=6, A-bit=6, Size (MB)=17.222020.0171.3
RELICv2Epochs=3000, Pre-training dataset=JFT-300M, Evaluation protocol=linear evaluation, Backbone=ResNet-502022.0171.1
OBQBackbone=ResNet34, Bitwidth=3bit, Quantization Scheme=Uniform symmetric per-channel weight quantization2022.0871.01
Divide and ContrastEpochs=4500, Pre-training dataset=JFT-300M, Evaluation protocol=linear evaluation, Backbone=ResNet-502022.0170.7
MobileNetV1Setting=baseline, MADDs=569M2020.0370.6
RELICv2Epochs=1000, Pre-training dataset=JFT-300M, Evaluation protocol=linear evaluation, Backbone=ResNet-502022.0170.3
ResNet-182023.0370.3
Divide and ContrastEpochs=3000, Pre-training dataset=JFT-300M, Evaluation protocol=linear evaluation, Backbone=ResNet-502022.0169.8
ShuffleNetV2 1xSetting=baseline, MADDs=147M2020.0369.5
AdaQuantBackbone=ResNet34, Bitwidth=4bit, Quantization Scheme=Uniform symmetric per-channel weight quantization2022.0869.49
BaselineW-bit=32, A-bit=32, Size (MB)=9.86, Backbone=SqueezeNext2020.0169.38
OBQBackbone=ResNet18, Bitwidth=4bit, Quantization Scheme=Uniform symmetric per-channel weight quantization2022.0869.18
ZEROQNo Data=true, No Fine-tuning=true, W-bit=8, A-bit=8, Size (MB)=2.47, Backbone=SqueezeNext2020.0169.17
DISSLBackbone=ResNet-50, Training Epochs=100, Batch Size=2560, Evaluation Protocol=Linear Probing, Representation Dimensionality=81922022.0968.9
ZEROQNo Data=true, No Fine-tuning=true, W-bit=MP, A-bit=6, Size (MB)=1.85, Backbone=SqueezeNext2020.0168.17
Feather-GlobalBackbone=MobileNetV1, Sparsity Ratio=89%2023.1068.13
Divide and ContrastEpochs=1000, Pre-training dataset=JFT-300M, Evaluation protocol=linear evaluation, Backbone=ResNet-502022.0167.9
BYOLEpochs=5000, Pre-training dataset=JFT-300M, Evaluation protocol=linear evaluation, Backbone=ResNet-502022.0167.9
CISSLBackbone=ResNet-50, Training Epochs=100, Batch Size=2560, Evaluation Protocol=Linear Probing, Representation Dimensionality=81922022.0967.7
BYOLEpochs=3000, Pre-training dataset=JFT-300M, Evaluation protocol=linear evaluation, Backbone=ResNet-502022.0167.6
BitSplitBackbone=ResNet18, Bitwidth=4bit, Quantization Scheme=Uniform symmetric per-channel weight quantization2022.0867.58
AdamWoptimizer=AdamW, Batch Size=16K, Training Epochs=90, Backbone=ResNet-50, Learning Rate Scheme=Standard2019.0467.27
OBQBackbone=ResNet18, Bitwidth=3bit, Quantization Scheme=Uniform symmetric per-channel weight quantization2022.0867.14
DRConv-ShuffleNetV2 0.5xSetting=8 regions, MADDs=46M2020.0367.1
BYOLEpochs=1000, Pre-training dataset=JFT-300M, Evaluation protocol=linear evaluation, Backbone=ResNet-502022.0167
ST-3Backbone=MobileNetV1, Sparsity Ratio=89%2023.1066.67
Barlow TwinsBackbone=ResNet-50, Training Epochs=100, Batch Size=2560, Evaluation Protocol=Linear Probing, Representation Dimensionality=20482022.0966.1
Adamoptimizer=Adam, Batch Size=16K, Training Epochs=90, Backbone=ResNet-50, Learning Rate Scheme=Standard2019.0466.04
AdaQuantBackbone=ResNet18, Bitwidth=4bit, Quantization Scheme=Uniform symmetric per-channel weight quantization2022.0865.45
ProbMaskBackbone=MobileNetV1, Sparsity Ratio=89%2023.1065.19
SimCLRBackbone=ResNet-50, Training Epochs=100, Batch Size=2560, Evaluation Protocol=Linear Probing, Representation Dimensionality=20482022.0965.1