Image Classification On Imagenet

Metrics

Hardware Burden

Number of params

Operations per network pass

Top 1 Accuracy

Results

Performance results of various models on this benchmark

Comparison Table

Model Name	Hardware Burden	Number of params	Operations per network pass	Top 1 Accuracy
xception-deep-learning-with-depthwise	87G	22.855952M	0.838G	79%
deep-residual-learning-for-image-recognition	-	40M	-	78.25%
cvt-introducing-convolutions-to-vision	-	20M	-	83%
densely-connected-convolutional-networks	-	-	-	77.42%
convit-improving-vision-transformers-with	-	10M	-	76.7%
when-vision-transformers-outperform-resnets	-	87M	-	79.9%
pvtv2-improved-baselines-with-pyramid-vision	-	25.4M	-	82%
metaformer-baselines-for-vision	-	40M	-	85.4%
mambavision-a-hybrid-mamba-transformer-vision	-	241.5M	-	85.3%
fbnetv5-neural-architecture-search-for	-	-	-	78.4%
gated-convolutional-networks-with-hybrid	-	12.9M	-	78.5%
convit-improving-vision-transformers-with	-	48M	-	82.2%
spatial-channel-token-distillation-for-vision	-	122.6M	-	82.4%
rethinking-and-improving-relative-position	-	22M	-	80.9%
bottleneck-transformers-for-visual	-	53.9M	-	84%
augmenting-sub-model-to-improve-main-model	-	86.6M	-	84.2%
maxvit-multi-axis-vision-transformer	-	120M	-	84.94%
biformer-vision-transformer-with-bi-level	-	-	-	84.3%
incorporating-convolution-designs-into-visual	-	6.4M	-	76.4%
automix-unveiling-the-power-of-mixup	-	44.6M	-	80.98%
resnest-split-attention-networks	-	27.5M	-	81.13%
dilated-neighborhood-attention-transformer	-	90M	-	84.4%
torchdistill-a-modular-configuration-driven	-	-	-	71.08%
next-vit-next-generation-vision-transformer	-	44.8M	-	83.2%
improved-multiscale-vision-transformers-for	-	218M	-	86.3%
cyclemlp-a-mlp-like-architecture-for-dense	-	76M	-	83.2%
efficientnetv2-smaller-models-and-faster	-	54M	-	86.2%
sp-vit-learning-2d-spatial-priors-for-vision	-	-	-	83.9%
an-improved-one-millisecond-mobile-backbone	-	2.1M	-	72.5%
high-performance-large-scale-image	-	193.8M	-	85.1%
mixpro-data-augmentation-with-maskmix-and	-	-	-	76.7%
co-training-2-l-submodels-for-visual	-	-	-	87.5%
hvt-a-comprehensive-vision-framework-for	-	-	-	87.4%
basisnet-two-stage-model-synthesis-for-1	-	-	-	80%
vitae-vision-transformer-advanced-by	-	-	-	75.3%
semi-supervised-recognition-under-a-noisy-and	-	25.58M	-	83.0%
Model 37	-	-	-	80.91%
alphanet-improved-training-of-supernet-with	-	-	-	80.0%
which-transformer-to-favor-a-comparative	-	-	-	82.29%
localvit-bringing-locality-to-vision	-	22.4M	-	80.8%
bottleneck-transformers-for-visual	-	-	-	83.5%
a-convnet-for-the-2020s	-	350M	-	87.8%
an-improved-one-millisecond-mobile-backbone	-	4.8M	-	75.9
an-image-is-worth-16x16-words-transformers-1	-	-	-	87.76%
190409925	-	-	-	79.1%
shape-texture-debiased-neural-network-1		-		81.2
gswin-gated-mlp-vision-model-with	-	39.8M	-	83.01%
revbifpn-the-fully-reversible-bidirectional	-	48.7M	-	83%
rethinking-spatial-dimensions-of-vision	-	4.9M	-	74.6%
dilated-neighborhood-attention-transformer	-	-	-	86.5%
an-evolutionary-approach-to-dynamic	-	-	-	86.74%
2103-14899	-	27.4M	-	81.5%
exploring-target-representations-for-masked	-	-	-	85.7%
on-the-adequacy-of-untuned-warmup-for	-	-	-	72.1%
alphanet-improved-training-of-supernet-with	-	-	-	77.8%
multigrain-a-unified-image-embedding-for	-	-	-	83.2%
gswin-gated-mlp-vision-model-with	-	15.5M	-	80.32%
Model 58	-	-	-	80.25%
wavemix-lite-a-resource-efficient-neural-1	-	32.4M	-	67.7%
randaugment-practical-data-augmentation-with	-	-	-	85.4%
maxvit-multi-axis-vision-transformer	-	-	-	88.32%
gtp-vit-efficient-vision-transformers-via	-	-	-	85.4%
an-improved-one-millisecond-mobile-backbone	-	14.8M	-	79.4%
cvt-introducing-convolutions-to-vision	-	32M	-	84.9%
splitnet-divide-and-co-training	-	88.6M	-	82.13%
tinyvit-fast-pretraining-distillation-for	-	21M	-	84.8%
performance-of-gaussian-mixture-model	-	-	-	84.1%
resnest-split-attention-networks	-	70M	-	83.9%
what-do-deep-networks-like-to-see	-	-	-	77.12%
filter-response-normalization-layer	-	-	-	78.95%
do-you-even-need-attention-a-stack-of-feed		-		74.9
repmlp-re-parameterizing-convolutions-into		52.77M		78.60%
scaling-vision-with-sparse-mixture-of-experts	-	7200M	-	88.36%
drop-an-octave-reducing-spatial-redundancy-in	20771G	66.8M	2.22G	82.9%
self-training-with-noisy-student-improves	-	9.2M	-	82.4%
visformer-the-vision-friendly-transformer	-	10.3M	-	78.6%
mixnet-mixed-depthwise-convolutional-kernels	-	4.1M	-	75.8%
three-things-everyone-should-know-about	-	-	-	83.4%
multigrain-a-unified-image-embedding-for	-	-	-	83.6%
dynamic-convolution-attention-over	-	18.6M	-	67.7%
token-labeling-training-a-85-5-top-1-accuracy	-	26M	-	83.3%
metaformer-baselines-for-vision	-	56M	-	85.2%
rethinking-local-perception-in-lightweight	-	4.2M	-	77%
lets-keep-it-simple-using-simple	-	3M	-	75.66
elsa-enhanced-local-self-attention-for-vision	-	298M	-	87.2%
an-improved-one-millisecond-mobile-backbone	-	2.1M	-	71.4%
bottleneck-transformers-for-visual	-	66.6M	-	82.2%
dilated-neighborhood-attention-transformer	-	20M	-	81.8%
revbifpn-the-fully-reversible-bidirectional	-	142.3M	-	84.2%
exploring-target-representations-for-masked	-	-	-	88.2%
fixing-the-train-test-resolution-discrepancy-2	-	9.2M	-	83.6%
kolmogorov-arnold-transformer	-	86.6M	-	81.8
Model 93	-	-	-	70.54
debiformer-vision-transformer-with-deformable	-	-	-	83.9%
mega-moving-average-equipped-gated-attention	-	90M	-	82.4%
espnetv2-a-light-weight-power-efficient-and	-	5.9M	-	74.9%
multigrain-a-unified-image-embedding-for	-	-	-	83.0%
semi-supervised-recognition-under-a-noisy-and	-	25.58M	-	84.0%
when-shift-operation-meets-vision-transformer	-	88M	-	83.3%
polynomial-networks-in-deep-classifiers	-	11.51M	-	71.6%
peco-perceptual-codebook-for-bert-pre	-	-	-	87.5%
puzzle-mix-exploiting-saliency-and-local-1		-		78.76%
token-labeling-training-a-85-5-top-1-accuracy	-	56M	-	84.1%
bottleneck-transformers-for-visual	-	-	-	83.8%
exploring-the-limits-of-weakly-supervised	-	829M	-	85.4%
fastvit-a-fast-hybrid-vision-transformer	-	-	-	82.6%
training-data-efficient-image-transformers	-	86M	-	84.2%
an-image-is-worth-16x16-words-transformers-1	-	-	-	24%
coca-contrastive-captioners-are-image-text		2100M		91.0%
circumventing-outliers-of-autoaugment-with		88M		85.8%
contrastive-learning-rivals-masked-image	-	307M	-	89.0%
transnext-robust-foveal-visual-perception-for	-	12.8M	-	82.5%
hyenapixel-global-image-context-with	-	-	-	84.9%
2103-15358	-	39.7M	-	83.3%
a-dot-product-attention-free-transformer	-	22.6M	-	79.8%
tinyvit-fast-pretraining-distillation-for	-	11M	-	83.2%
torchdistill-a-modular-configuration-driven	-	-	-	70.09%
florence-a-new-foundation-model-for-computer	-	893M	-	90.05%
coatnet-marrying-convolution-and-attention	-	42M	-	83.3%
high-performance-large-scale-image	-	377.2M	-	86.0%
transboost-improving-the-best-imagenet	-	25.56M	-	81.15%
maxvit-multi-axis-vision-transformer	-	69M	-	84.45%
sp-vit-learning-2d-spatial-priors-for-vision	-	-	-	85.5%
transboost-improving-the-best-imagenet	-	60.19M	-	80.64%
efficient-self-supervised-learning-with	-	-	-	87.4%
revbifpn-the-fully-reversible-bidirectional	-	5.11M	-	75.9%
scarletnas-bridging-the-gap-between	-	6.5M	-	76.3%
splitnet-divide-and-co-training		98M		83.34%
fastervit-fast-vision-transformers-with	-	53.4M	-	83.2%
fairnas-rethinking-evaluation-fairness-of	-	4.6M	-	75.34%
dynamic-convolution-attention-over	-	4M	-	69.4%
resmlp-feedforward-networks-for-image	-	17.7M	-	78.6%
generalized-parametric-contrastive-learning	-	-	-	86.01%
co-training-2-l-submodels-for-visual	-	-	-	85.8%
mixpro-data-augmentation-with-maskmix-and	-	-	-	73.8%
spinenet-learning-scale-permuted-backbone-for	-	60.5M	-	79%
unicom-universal-and-compact-representation	-	-	-	88.3
billion-scale-semi-supervised-learning-for	-	193M	-	84.8%
efficient-multi-order-gated-aggregation	-	3M	-	77.2%
pattern-attention-transformer-with-doughnut	-	-	-	83.6%
fastervit-fast-vision-transformers-with	-	159.5M	-	84.9%
reproducible-scaling-laws-for-contrastive	-	-	-	88.5%
which-transformer-to-favor-a-comparative	-	-	-	82.11%
190411491	-	14.7M	-	75.7%
graph-rise-graph-regularized-image-semantic	-	-	-	68.29%
muxconv-information-multiplexing-in	-	2.4M	-	71.6%
mobilenets-efficient-convolutional-neural	-	-	-	70.6%
fractalnet-ultra-deep-neural-networks-without	-	-	-	75.88%
xcit-cross-covariance-image-transformers	-	189M	-	86%
understanding-gaussian-attention-bias-of	-	-	-	81.484%
fixing-the-train-test-resolution-discrepancy	-	-	-	79.1%
scaling-vision-transformers-to-22-billion	-	86M	-	88.6%
ghostnet-more-features-from-cheap-operations	-	13M	-	75%
volo-vision-outlooker-for-visual-recognition	-	296M	-	87.1%
dat-spatially-dynamic-vision-transformer-with	-	53M	-	84.6%
convit-improving-vision-transformers-with	-	86M	-	82.4%
efficientvit-enhanced-linear-attention-for	-	64M	-	85.6%
mambavision-a-hybrid-mamba-transformer-vision	-	227.9M	-	85%
greedynas-towards-fast-one-shot-nas-with		6.5M		77.1%
visual-attention-network	-	200M	-	86.9%
a-dot-product-attention-free-transformer	-	23M	-	80.8%
scalable-visual-transformers-with	-	5.74M	-	69.64%
on-the-performance-analysis-of-momentum	-	-	-	76.91%
levit-a-vision-transformer-in-convnet-s	-	39.4M	-	82.5%
fixing-the-train-test-resolution-discrepancy	-	25.6M	-	82.5%
scaling-up-your-kernels-to-31x31-revisiting	-	335M	-	87.8%
contextual-transformer-networks-for-visual	-	55.8M	-	84.6%
meta-knowledge-distillation	-	-	-	83.1%
involution-inverting-the-inherence-of	-	34M	-	79.3%
tokens-to-token-vit-training-vision	-	-	-	83.3%
rethinking-the-design-principles-of-robust	-	23.3M	-	81.9%
which-transformer-to-favor-a-comparative	-	-	-	83.61%
a-convnet-for-the-2020s	-	1827M	-	88.36%
clcnet-rethinking-of-ensemble-modeling-with	-	-	-	86.42%
going-deeper-with-image-transformers	-	438M	-	86.5%
three-things-everyone-should-know-about	-	-	-	85.5%
fbnetv5-neural-architecture-search-for	-	-	-	84.1%
sp-vit-learning-2d-spatial-priors-for-vision	-	-	-	85.1%
semi-supervised-recognition-under-a-noisy-and	-	5.47M	-	79.0%
augmenting-convolutional-networks-with	-	99.4M	-	83.5%
from-xception-to-nexception-new-design	-	-	-	82%
automix-unveiling-the-power-of-mixup	-	21.8M	-	76.1%
balanced-binary-neural-networks-with-gated	-	-	-	59.4%
co-training-2-l-submodels-for-visual	-	-	-	88.0%
transnext-robust-foveal-visual-perception-for	-	49.7M	-	86.0%
refiner-refining-self-attention-for-vision		81M		86.03
co-training-2-l-submodels-for-visual	-	-	-	86.3%
repvgg-making-vgg-style-convnets-great-again	-	80.31M	-	78.78%
firecaffe-near-linear-acceleration-of-deep	-	-	-	58.9%
convmlp-hierarchical-convolutional-mlps-for	-	42.7M	-	80.2%
Model 191	-	-	-	78.15%
internimage-exploring-large-scale-vision	-	50M	-	84.2%
2103-15358	-	24.6M	-	82%
scaling-up-visual-and-vision-language		480M		88.64%
from-xception-to-nexception-new-design	-	-	-	81.8%
gtp-vit-efficient-vision-transformers-via	-	-	-	83.7%
multigrain-a-unified-image-embedding-for	-	-	-	79.4%
zen-nas-a-zero-shot-nas-for-high-performance	-	183M	-	83.0%
kolmogorov-arnold-transformer	-	86.6M	-	82.8
a-fast-knowledge-distillation-framework-for	-	5M	-	78.7%
hyenapixel-global-image-context-with	-	-	-	83.5%
gtp-vit-efficient-vision-transformers-via	-	-	-	81.9%
rethinking-and-improving-relative-position	-	-	-	81.1%
which-transformer-to-favor-a-comparative	-	-	-	81.33%
one-peace-exploring-one-general	-	1520M	-	-
moat-alternating-mobile-convolution-and	-	483.2M	-	89.1%
sequencer-deep-lstm-for-image-classification	-	28M	-	82.3%
dilated-neighborhood-attention-transformer	-	51M	-	83.8%
tinyvit-fast-pretraining-distillation-for	-	11M	-	81.5%
Model 210	-	-	-	78.75%
incorporating-convolution-designs-into-visual	-	-	-	82%
high-performance-large-scale-image	-	527M	-	89.2%
self-training-with-noisy-student-improves	-	66M	-	86.9%
cswin-transformer-a-general-vision	-	173M	-	87.5%
debiformer-vision-transformer-with-deformable	-	-	-	84.4%
cutmix-regularization-strategy-to-train	-	-	-	78.4%
mambavision-a-hybrid-mamba-transformer-vision	-	97.7M	-	84.2%
expeditious-saliency-guided-mix-up-through	-	-	-	77.39%
nasvit-neural-architecture-search-for	-	-	-	78.2%
volo-vision-outlooker-for-visual-recognition	-	59M	-	86%
wave-vit-unifying-wavelet-and-transformers	-	33.5M	-	84.8%
alphanet-improved-training-of-supernet-with	-	-	-	78.9%
nasvit-neural-architecture-search-for	-	-	-	81.0%
mixpro-data-augmentation-with-maskmix-and	-	-	-	82.9%
omnivec2-a-novel-transformer-based-network	-	-	-	89.3%
transboost-improving-the-best-imagenet	-	71.71M	-	82.16%
rexnet-diminishing-representational	-	2.7M	-	74.6%
next-vit-next-generation-vision-transformer	-	57.8M	-	84.7%
vision-models-are-more-robust-and-fair-when	-	10000M	-	85.8%
nasvit-neural-architecture-search-for	-	-	-	81.4%
masked-autoencoders-are-scalable-vision	-	-	-	85.9%
vitaev2-vision-transformer-advanced-by	-	644M	-	88.5%
neighborhood-attention-transformer	-	20M	-	81.8%
an-image-is-worth-16x16-words-transformers-1	-	-	-	-
revisiting-resnets-improved-training-and		192M		84.4%
efficientnet-rethinking-model-scaling-for	-	66M	-	84.4%
2103-15358	-	55.7M	-	83.2%
lets-keep-it-simple-using-simple	-	1.5M	-	61.52
vision-gnn-an-image-is-worth-graph-of-nodes	-	10.7M	-	78.2%
adaptive-split-fusion-transformer	-	56.7M	-	83.9%
gtp-vit-efficient-vision-transformers-via	-	-	-	85.8%
uninet-unified-architecture-search-with-1	-	11.5M	-	80.8%
augmenting-convolutional-networks-with	-	99.4M	-	86.5%
mixpro-data-augmentation-with-maskmix-and	-	-	-	80.6%
going-deeper-with-image-transformers	-	68.2M	-	85.4%
which-transformer-to-favor-a-comparative	-	-	-	82.54%
Model 247	-	-	-	76.3
fastvit-a-fast-hybrid-vision-transformer	-	-	-	84.9%
davit-dual-attention-vision-transformers	-	87.9M	-	86.9%
co-training-2-l-submodels-for-visual	-	-	-	87.1%
Model 251	-	-	-	81.92%
firecaffe-near-linear-acceleration-of-deep	-	-	-	68.3%
sp-vit-learning-2d-spatial-priors-for-vision	-	-	-	84.9%
efficientnet-rethinking-model-scaling-for	-	9.2M	-	79.8%
efficientnet-rethinking-model-scaling-for	-	12M	-	81.1%
exploring-the-limits-of-weakly-supervised	-	466M	-	85.1%
correlated-input-dependent-label-noise-in		-		68.6%
not-all-images-are-worth-16x16-words-dynamic	-	-	-	78.48%
rethinking-and-improving-relative-position	-	-	-	81.4%
container-context-aggregation-network	-	20M	-	82%
self-training-with-noisy-student-improves	-	30M	-	86.1%
srm-a-style-based-recalibration-module-for	-	-	-	78.47%
densely-connected-convolutional-networks	-	-	-	74.98%
self-training-with-noisy-student-improves	-	19M	-	85.3%
Model 265	-	-	-	78.36%
exploring-randomly-wired-neural-networks-for	-	61.5M	-	80.1%
mambavision-a-hybrid-mamba-transformer-vision	-	31.8M	-	82.3%
fast-vision-transformers-with-hilo-attention	-	28M	-	82%
your-diffusion-model-is-secretly-a-zero-shot	-	-	-	79.1%
xcit-cross-covariance-image-transformers	-	84M	-	85.8%
semi-supervised-learning-of-visual-features	-	-	-	75.5%
cutmix-regularization-strategy-to-train	-	-	-	80.53%
convit-improving-vision-transformers-with		152M		82.5%
res2net-a-new-multi-scale-backbone	-	-	-	78.59%
exploring-target-representations-for-masked	-	-	-	87.8%
going-deeper-with-image-transformers	-	271M	-	86.3%
rest-an-efficient-transformer-for-visual	-	13.66M	-	79.6%
resnest-split-attention-networks	-	27.5M	-	80.64%
mobilevitv3-mobile-friendly-vision	-	5.8M	-	79.3%
high-performance-large-scale-image	-	254.9M	-	85.7%
visual-attention-network	-	60M	-	86.6%
tokenmixup-efficient-attention-guided-token	-	-	-	82.37%
wide-residual-networks	-	-	-	78.1%
unsupervised-data-augmentation-1	-	-	-	79.04%
mobilenetv4-universal-models-for-the-mobile	-	-	-	79.9%
parametric-contrastive-learning	-	-	-	80.9%
designing-network-design-spaces	-	6.3M	-	76.3%
adversarial-autoaugment-1	-	-	-	79.4%
mixmim-mixed-and-masked-image-modeling-for	-	88M	-	85.1%
neighborhood-attention-transformer	-	28M	-	83.2%
mixpro-data-augmentation-with-maskmix-and	-	-	-	82.7%
hornet-efficient-high-order-spatial	-	-	-	87.7%
vitae-vision-transformer-advanced-by	-	13.2M	-	81%
going-deeper-with-image-transformers	-	46.9M	-	85.1%
unireplknet-a-universal-perception-large	-	-	-	87.4%
designing-bert-for-convolutional-networks	-	198M	-	86.0%
swin-transformer-hierarchical-vision	-	88M	-	86.4%
Model 298	-	-	-	77.71%
quantnet-learning-to-quantize-by-learning	-	-	-	71.97%
adversarial-examples-improve-image	-	66M	-	85.2%
whats-hidden-in-a-randomly-weighted-neural	-	20.6M	-	73.3%
involution-inverting-the-inherence-of	-	15.5M	-	78.4%
elsa-enhanced-local-self-attention-for-vision	-	28M	-	82.7%
next-vit-next-generation-vision-transformer	-	31.7M	-	82.5%
muxconv-information-multiplexing-in	-	4.0M	-	76.6%
alphanet-improved-training-of-supernet-with	-	-	-	80.3%
2103-14899	-	44.3M	-	82.8%
sliced-recursive-transformer-1	-	21.3M	-	84.3%
resmlp-feedforward-networks-for-image	-	-	-	79.4%
mobilevit-light-weight-general-purpose-and	-	5.6M	-	78.4%
masked-image-residual-learning-for-scaling-1	-	96M	-	84.8%
neighborhood-attention-transformer	-	90M	-	84.3%
bottleneck-transformers-for-visual	-	49.2M	-	81.4%
pvtv2-improved-baselines-with-pyramid-vision	-	45.2M	-	83.2%
scalable-pre-training-of-large-autoregressive	-	-	-	84.0
distilled-gradual-pruning-with-pruned-fine	-	2.56M	-	73.66%
uninet-unified-architecture-search-with-1	-	72.9M	-	87%
fastvit-a-fast-hybrid-vision-transformer	-	-	-	80.6%
gated-convolutional-networks-with-hybrid	-	42.2M	-	80.5%
augmenting-sub-model-to-improve-main-model	-	632M	-	85.7%
fbnetv5-neural-architecture-search-for	-	-	-	77.2%
tiny-models-are-the-computational-saver-for	-	-	-	85.75
mish-a-self-regularized-non-monotonic-neural	-	-	-	79.8%
differentiable-model-compression-via-pseudo		-		82.0
graph-convolutions-enrich-the-self-attention	-	-	-	82.8%
model-soups-averaging-weights-of-multiple	-	1843M	-	90.94%
torchdistill-a-modular-configuration-driven	-	-	-	70.52%
filter-response-normalization-layer	-	-	-	77.21%
scarletnas-bridging-the-gap-between	-	6M	-	75.6%
cas-vit-convolutional-additive-self-attention	-	21.76	-	84.1%
lets-keep-it-simple-using-simple	-	9.5M	-	74.17
fast-autoaugment	-	-	-	80.6%
self-training-with-noisy-student-improves	-	12M	-	84.1%
uninet-unified-architecture-search-with	-	11.9M	-	79.1%
lets-keep-it-simple-using-simple	-	1.5M	-	69.11
involution-inverting-the-inherence-of	-	9.2	-	75.9%
not-all-images-are-worth-16x16-words-dynamic		-		80.43%
uninet-unified-architecture-search-with	-	73.5M	-	85.2%
omnivore-a-single-model-for-many-visual	-	-	-	86.0%
densely-connected-convolutional-networks	-	-	-	76.2%
efficientnet-rethinking-model-scaling-for	-	5.3M	-	76.3%
rethinking-local-perception-in-lightweight	-	12.3M	-	81.6%
compact-global-descriptor-for-neural-networks	-	4.26M	-	72.56%
improved-multiscale-vision-transformers-for	-	667M	-	88%
revisiting-a-knn-based-image-classification	-	-	-	79.8%
visual-representation-learning-from-unlabeled	-	-	-	85%
efficientvit-enhanced-linear-attention-for	-	64M	-	86%
billion-scale-semi-supervised-learning-for	-	42M	-	83.4%
incepformer-efficient-inception-transformer	-	14.0M	-	80.5%
fast-vision-transformers-with-hilo-attention	-	87M	-	84.7%
rexnet-diminishing-representational	-	34.8M	-	84.5%
identity-mappings-in-deep-residual-networks	-	-	-	79.9%
mixpro-data-augmentation-with-maskmix-and	-	-	-	81.2%
maxvit-multi-axis-vision-transformer	-	-	-	86.7%
asymmnet-towards-ultralight-convolution	-	3.1M	-	68.4%
Model 356	-	62M	-	63.3
three-things-everyone-should-know-about	-	-	-	84.1%
mobilevitv3-mobile-friendly-vision	-	1.2M	-	70.98%
edgenext-efficiently-amalgamated-cnn	-	1.3M	-	71.2%
distilled-gradual-pruning-with-pruned-fine	-	1.03M	-	65.59%
sparse-mlp-for-image-recognition-is-self	-	65.9M	-	83.4%
tokens-to-token-vit-training-vision	-	64.4M	-	82.6%
bias-loss-for-mobile-neural-networks	-	5.5M	-	76.2%
metaformer-baselines-for-vision	-	27M	-	85.0%
transboost-improving-the-best-imagenet	-	-	-	79.03%
graph-convolutions-enrich-the-self-attention	-	-	-	83%
an-improved-one-millisecond-mobile-backbone	-	10.1M	-	80.0%
torchdistill-a-modular-configuration-driven	-	-	-	71.56%
on-the-performance-analysis-of-momentum	-	-	-	67.74%
hrformer-high-resolution-transformer-for	-	50.3M	-	82.8%
collaboration-of-experts-achieving-80-top-1	-	-	-	81.5%
metaformer-baselines-for-vision	-	40M	-	86.4%
incorporating-convolution-designs-into-visual	-	24.2M	-	83.3%
bossnas-exploring-hybrid-cnn-transformers	-	-	-	82.2%
efficientnetv2-smaller-models-and-faster	-	22M	-	84.9%
unireplknet-a-universal-perception-large	-	-	-	86.4%
torchdistill-a-modular-configuration-driven	-	-	-	70.93%
densenets-reloaded-paradigm-shift-beyond	-	50M	-	83.7%
deepmad-mathematical-architecture-design-for	-	89M	-	84%
involution-inverting-the-inherence-of	-	25.6M	-	79.1%
uninet-unified-architecture-search-with	-	73.5M	-	84.2%
automix-unveiling-the-power-of-mixup	-	25.6M	-	79.25%
fixing-the-train-test-resolution-discrepancy	62G	829M	-	86.4%
wavemix-lite-a-resource-efficient-neural	-	-	-	74.93%
fast-autoaugment	-	-	-	77.6%
activemlp-an-mlp-like-architecture-with	-	27.2M	-	82%
a-dot-product-attention-free-transformer	-	20.3M	-	80.2%
uninet-unified-architecture-search-with	-	14M	-	80.4%
going-deeper-with-image-transformers	-	185.9M	-	85.8%
maxvit-multi-axis-vision-transformer	-	-	-	89.53%
maxup-a-simple-way-to-improve-generalization		87.42M		85.8%
scaling-local-self-attention-for-parameter	-	87M	-	85.5%
global-context-vision-transformers	-	20M	-	82.0%
meal-v2-boosting-vanilla-resnet-50-to-80-top	-	-	-	80.67%
ghostnet-more-features-from-cheap-operations	-	2.6M	-	66.2%
self-training-with-noisy-student-improves	-	5.3M	-	78.8%
convmlp-hierarchical-convolutional-mlps-for	-	9M	-	76.8
Model 398	-	-	-	66.04
efficientnet-rethinking-model-scaling-for	-	43M	-	84%
semi-supervised-learning-of-visual-features	-	-	-	66.5%
multigrain-a-unified-image-embedding-for	-	-	-	83.1%
metaformer-baselines-for-vision	-	39M	-	84.5%
rethinking-and-improving-relative-position	-	6M	-	73.7%
perceiver-general-perception-with-iterative	-	44.9M	-	78%
rexnet-diminishing-representational	-	34.7M	-	82.8%
maxvit-multi-axis-vision-transformer	-	212M	-	85.17%
eca-net-efficient-channel-attention-for-deep	-	57.40M	-	78.92%
vision-transformer-with-deformable-attention	-	50M	-	83.7%
neighborhood-attention-transformer	-	51M	-	83.7%
unconstrained-open-vocabulary-image	-	-	-	88.21%
bottleneck-transformers-for-visual	-	44.4M	-	80%
parametric-contrastive-learning	-	-	-	81.8%
metaformer-baselines-for-vision	-	39M	-	85.8%
which-transformer-to-favor-a-comparative	-	-	-	81.96%
an-algorithm-for-routing-vectors-in-sequences	-	312.8M	-	86.7%
global-context-vision-transformers	-	28M	-	83.4%
torchdistill-a-modular-configuration-driven		-		71.71%
maxvit-multi-axis-vision-transformer	-	-	-	86.4%
deformable-kernels-adapting-effective	-	-	-	78.5%
internimage-exploring-large-scale-vision	-	97M	-	84.9%
resnest-split-attention-networks	-	48M	-	83.0%
sparse-mlp-for-image-recognition-is-self	-	24.1M	-	81.9%
autodropout-learning-dropout-patterns-to		-		80.3%
coatnet-marrying-convolution-and-attention	-	75M	-	84.1%
adaptively-connected-neural-networks	-	29.38M	-	77.5%
mixpro-data-augmentation-with-maskmix-and	-	-	-	83.7%
rexnet-diminishing-representational	-	4.8M	-	77.9%
boosting-discriminative-visual-representation	-	11.7M	-	72.33%
cas-vit-convolutional-additive-self-attention	-	3.2	-	78.7%
an-improved-one-millisecond-mobile-backbone	-	14.8M	-	81.4%
the-information-pathways-hypothesis	-	-	-	81.89%
enhance-the-visual-representation-via	-	-	-	87.02%
polyloss-a-polynomial-expansion-perspective-1	-	-	-	87.2%
wave-vit-unifying-wavelet-and-transformers	-	22.7M	-	83.9%
global-context-vision-transformers	-	90M	-	84.5%
mixpro-data-augmentation-with-maskmix-and	-	-	-	82.8%
efficient-multi-order-gated-aggregation	-	83M	-	84.7%
multimodal-autoregressive-pre-training-of	-	-	-	88.5%
aggregating-nested-transformers	-	17M	-	81.5%
fairnas-rethinking-evaluation-fairness-of	-	4.4M	-	74.69%
scarletnas-bridging-the-gap-between	12G	27.8M	0.42G	82.3%
coatnet-marrying-convolution-and-attention	-	-	-	87.6%
self-training-with-noisy-student-improves	-	7.8M	-	81.5%
gtp-vit-efficient-vision-transformers-via	-	-	-	82.8%
Model 445	-	-	-	81.12%
rethinking-local-perception-in-lightweight	-	7.2M	-	79.8%
metaformer-baselines-for-vision	-	26M	-	83.6%
maxvit-multi-axis-vision-transformer	-	-	-	88.69%
metaformer-baselines-for-vision	-	100M	-	85.7%
bias-loss-for-mobile-neural-networks	-	7.1M	-	77.1%
high-performance-large-scale-image	-	71.5M	-	83.6%
metaformer-baselines-for-vision	-	57M	-	84.5%
augmenting-convolutional-networks-with	-	188.6M	-	84.1%
revbifpn-the-fully-reversible-bidirectional	-	3.42M	-	72.8%
metaformer-baselines-for-vision	-	40M	-	84.1%
cvt-introducing-convolutions-to-vision	-	-	-	81.6%
reversible-column-networks	-	2158M	-	90.0%
wave-vit-unifying-wavelet-and-transformers	-	57.5M	-	85.5%
tokens-to-token-vit-training-vision	-	21.5M	-	81.5%
metaformer-baselines-for-vision	-	27M	-	83.0%
attentional-feature-fusion		34.7M		80.22%
three-things-everyone-should-know-about	-	-	-	84.1%
dynamic-convolution-attention-over	-	11.1M	-	74.4%
pvtv2-improved-baselines-with-pyramid-vision	-	3.4M	-	70.5%
2103-15358	-	39.8M	-	82.9%
adaptive-split-fusion-transformer	-	19.3M	-	82.7%
multiscale-vision-transformers	-	72.9M	-	84.8%
boosting-discriminative-visual-representation	-	44.6M	-	81.08%
Model 469	-	-	-	78.15%
generalized-parametric-contrastive-learning	-	-	-	84.0%
davit-dual-attention-vision-transformers	-	28.3M	-	82.8%
contextual-classification-using-self		-		77.0%
aggregating-nested-transformers	-	38M	-	83.3%
learned-queries-for-efficient-local-attention	-	25M	-	83.2%
generalized-parametric-contrastive-learning	-	-	-	79.7%
gated-attention-coding-for-training-high	-	-	-	70.42
maxvit-multi-axis-vision-transformer	-	-	-	88.46%
dilated-neighborhood-attention-transformer	-	-	-	87.4%
model-soups-averaging-weights-of-multiple	-	2440M	-	90.98%
uniformer-unifying-convolution-and-self	-	100M	-	86.3%
2103-14899	-	43.3M	-	82.5%
dat-spatially-dynamic-vision-transformer-with	-	24M	-	83.9%
global-filter-networks-for-image		54M		82.9%
nasvit-neural-architecture-search-for	-	-	-	79.7%
unireplknet-a-universal-perception-large	-	-	-	83.2%
sparse-mlp-for-image-recognition-is-self	-	48.6M	-	83.1%
colornet-investigating-the-importance-of	-	-	-	82.35%
volo-vision-outlooker-for-visual-recognition	-	27M	-	85.2%
fastervit-fast-vision-transformers-with	-	75.9M	-	84.2%
exploring-the-limits-of-weakly-supervised	-	194M	-	84.2%
nasvit-neural-architecture-search-for	-	-	-	80.5%
deit-iii-revenge-of-the-vit	-	-	-	85.7%
compress-image-to-patches-for-vision	-	-	-	77%
discrete-representations-strengthen-vision-1	-	-	-	85.07%
which-transformer-to-favor-a-comparative	-	-	-	83.65%
mamba2d-a-natively-multi-dimensional-state	-	-	-	82.4%
lip-local-importance-based-pooling	-	42.9M	-	79.33%
coca-contrastive-captioners-are-image-text	-	2100M	-	91.0%
masked-image-residual-learning-for-scaling-1	-	341M	-	86.2%
rethinking-the-design-principles-of-robust	-	10.9M	-	79.2%
large-scale-learning-of-general-visual	-	928M	-	85.39%
co-training-2-l-submodels-for-visual	-	-	-	86.2%
efficient-multi-order-gated-aggregation	-	25M	-	83.4%
when-vision-transformers-outperform-resnets	-	236M	-	81.1%
unireplknet-a-universal-perception-large	-	-	-	88%
rexnet-diminishing-representational	-	16.5M	-	83.2%
dat-spatially-dynamic-vision-transformer-with	-	93M	-	84.9%
spatial-group-wise-enhance-improving-semantic	-	25.56M	-	77.584%
uniformer-unifying-convolution-and-self	-	22M	-	83.4%
augmenting-sub-model-to-improve-main-model	-	304M	-	85.3%
averaging-weights-leads-to-wider-optima-and	-	-	-	78.44%
metaformer-baselines-for-vision	-	100M	-	87.6%
deit-iii-revenge-of-the-vit	-	-	-	81.4%
pay-attention-to-mlps	-	73M	-	81.6%
the-effectiveness-of-mae-pre-pretraining-for	-	6500M	-	90.1%
maxvit-multi-axis-vision-transformer	-	-	-	88.7%
xcit-cross-covariance-image-transformers	-	48M	-	85.6%
visual-attention-network	-	-	-	87%
efficientnet-rethinking-model-scaling-for	-	19M	-	82.6%
revisiting-resnets-improved-training-and	-	-	-	83.8%
multigrain-a-unified-image-embedding-for	-	-	-	82.6%
understanding-the-robustness-in-vision	-	76.8M	-	87.1%
zen-nas-a-zero-shot-nas-for-high-performance	-	5.7M	-	78%
learning-transferable-architectures-for	1648G	88.9M	2.38G	82.7%
rexnet-diminishing-representational	-	19M	-	81.6%
vision-gnn-an-image-is-worth-graph-of-nodes	-	27.3M	-	82.1%
dat-spatially-dynamic-vision-transformer-with	-	94M	-	85.9%
deit-iii-revenge-of-the-vit	-	-	-	83.8%
Model 529	-	-	-	67.63
bnn-bn-training-binary-neural-networks		-		68.0%
fixing-the-train-test-resolution-discrepancy-2	-	-	-	85.7%
mambavision-a-hybrid-mamba-transformer-vision	-	35.1M	-	82.7%
attentive-normalization	-	-	-	81.87%
alphanet-improved-training-of-supernet-with	-	-	-	80.8%
metaformer-baselines-for-vision	-	27M	-	83.7%
semi-supervised-recognition-under-a-noisy-and		76M		85.1%
hrformer-high-resolution-transformer-for	-	8.0M	-	78.5%
which-transformer-to-favor-a-comparative	-	-	-	81.09%
transboost-improving-the-best-imagenet	-	11.69M	-	73.36%
metaformer-baselines-for-vision	-	26M	-	85.0%
a-large-batch-optimizer-reality-check		-		75.92%
lets-keep-it-simple-using-simple	-	3M	-	68.15
deep-residual-learning-for-image-recognition	-	25M	-	75.3%
centroid-transformers-learning-to-abstract	-	22.3M	-	80.9%
transnext-robust-foveal-visual-perception-for	-	89.7M	-	86.2%
going-deeper-with-image-transformers	-	89.5M	-	85.3%
dilated-neighborhood-attention-transformer	-	200M	-	87.5%
supervised-contrastive-learning		-		80.8%
metaformer-baselines-for-vision	-	57M	-	86.1%
asymmetric-masked-distillation-for-pre	-	87M	-	84.6%
davit-dual-attention-vision-transformers	-	87.9M	-	84.6%
tiny-models-are-the-computational-saver-for	-	-	-	85.24
mobilevitv3-mobile-friendly-vision	-	-	-	78.64%
internimage-exploring-large-scale-vision	-	3000M	-	90.1%
deit-iii-revenge-of-the-vit	-	-	-	83.1%
when-shift-operation-meets-vision-transformer	-	28M	-	81.7%
grafit-learning-fine-grained-image		-		79.6%
mnasnet-platform-aware-neural-architecture	-	3.9M	-	75.2%
tokenlearner-what-can-8-learned-tokens-do-for	-	-	-	87.07%
neural-architecture-transfer	-	9.1M	-	80.5%
distilling-out-of-distribution-robustness-1	-	-	-	81.9%
dynamicvit-efficient-vision-transformers-with		57.1M		83.9
rethinking-spatial-dimensions-of-vision	-	10.6M	-	79.1%
bag-of-tricks-for-image-classification-with	-	25M	-	77.16%
deit-iii-revenge-of-the-vit	-	-	-	84.9%
mnasnet-platform-aware-neural-architecture		5.2M	0.0403G	76.7%
fixing-the-train-test-resolution-discrepancy-2	-	19M	-	85.9%
efficientnet-rethinking-model-scaling-for	-	7.8M	-	78.8%
Model 569	-	-	-	81.97%
localvit-bringing-locality-to-vision	-	4.3M	-	72.5%
tinyvit-fast-pretraining-distillation-for	-	21M	-	86.2%
single-path-nas-designing-hardware-efficient	-	-	-	74.96%
volo-vision-outlooker-for-visual-recognition	-	86M	-	86.3%
levit-a-vision-transformer-in-convnet-s	-	17.8M	-	81.6%
multigrain-a-unified-image-embedding-for	-	-	-	75.1%
fixing-the-train-test-resolution-discrepancy-2		480M		88.5%
dynamic-convolution-attention-over	-	2.8M	-	64.9%
deit-iii-revenge-of-the-vit	-	304.8M	-	85.8%
high-performance-large-scale-image	-	438.4M	-	86.5%
deepvit-towards-deeper-vision-transformer	-	55M	-	82.2%
biformer-vision-transformer-with-bi-level	-	-	-	81.4%
metaformer-baselines-for-vision	-	26M	-	85.4%
selective-kernel-networks	-	48.9M	-	79.81%
going-deeper-with-image-transformers	-	17.3M	-	82.2%
unireplknet-a-universal-perception-large	-	-	-	87.9%
efficientnetv2-smaller-models-and-faster	-	-	-	85.7%
visformer-the-vision-friendly-transformer	-	40.2M	-	82.2%
boosting-discriminative-visual-representation	-	21.8M	-	76.35%
online-training-through-time-for-spiking	-	-	-	65.15%
pattern-attention-transformer-with-doughnut	-	-	-	83.1%
unconstrained-open-vocabulary-image	-	-	-	83.46%
improved-multiscale-vision-transformers-for	-	667M	-	88.8%
spatial-channel-token-distillation-for-vision	-	22.2M	-	75.7%
transboost-improving-the-best-imagenet	-	21.8M	-	76.70%
tresnet-high-performance-gpu-dedicated		77M		84.3%
metaformer-baselines-for-vision	-	99M	-	86.4%
vision-gnn-an-image-is-worth-graph-of-nodes	-	51.7M	-	83.1%
maxvit-multi-axis-vision-transformer	-	-	-	86.34%
mobilevit-light-weight-general-purpose-and	-	2.3M	-	74.8%
mixpro-data-augmentation-with-maskmix-and	-	-	-	84.1%
visual-attention-network	-	26.6M	-	82.8%
visual-attention-network	-	90M	-	86.3%
unireplknet-a-universal-perception-large	-	-	-	83.9%
levit-a-vision-transformer-in-convnet-s	-	4.7M	-	75.7%
lambdanetworks-modeling-long-range-1	-	35M	-	84.0%
rckd-response-based-cross-task-knowledge	-	3M	-	78.6
incorporating-convolution-designs-into-visual	-	-	-	78.8%
greedynas-towards-fast-one-shot-nas-with	-	4.7M	-	76.2%
deeper-vs-wider-a-revisit-of-transformer	-	-	-	84.2
2103-15358	-	6.7M	-	76.7%
fbnetv5-neural-architecture-search-for	-	-	-	81.8%
x-volution-on-the-unification-of-convolution	-	-	-	75%
metaformer-baselines-for-vision	-	56M	-	86.6%
parametric-contrastive-learning	-	-	-	81.3%
re-labeling-imagenet-from-single-to-multi	-	4.8M	-	78.4%
self-training-with-noisy-student-improves	-	43M	-	86.4%
transboost-improving-the-best-imagenet	-	44.55M	-	79.86%
swin-transformer-v2-scaling-up-capacity-and	-	3000M	-	90.17%
deit-iii-revenge-of-the-vit	-	87M	-	85.0%
mlp-mixer-an-all-mlp-architecture-for-vision	-	-	-	85.3%
container-context-aggregation-network	-	22.1M	-	82.7%
graph-convolutions-enrich-the-self-attention	-	-	-	81.1%
contextual-convolutional-neural-networks	-	60M	-	79.03%
tokens-to-token-vit-training-vision	-	-	-	82.3%
conformer-local-features-coupling-global		83.3M		84.1%
differentiable-spike-rethinking-gradient	-	-	-	71.24
which-transformer-to-favor-a-comparative	-	-	-	78.42%
sequencer-deep-lstm-for-image-classification	-	54M	-	83.4%
mobilenetv4-universal-models-for-the-mobile	-	-	-	80.7%
which-transformer-to-favor-a-comparative	-	-	-	84.91%
collaboration-of-experts-achieving-80-top-1	-	-	-	80%
fixing-the-train-test-resolution-discrepancy-2	-	19M	-	84.0%
co-training-2-l-submodels-for-visual	-	-	-	85.8%
efficientvit-enhanced-linear-attention-for	-	49M	-	84.2%
designing-network-design-spaces	-	6.1M	-	75.5%
beit-bert-pre-training-of-image-transformers	-	331M	-	88.60%
lets-keep-it-simple-using-simple	-	9.5M	-	81.24
deeper-vs-wider-a-revisit-of-transformer	-	-	-	87.1
fastervit-fast-vision-transformers-with	-	424.6M	-	85.4%
tiny-models-are-the-computational-saver-for	-	-	-	86.24
cas-vit-convolutional-additive-self-attention	-	12.42	-	83.0%
compounding-the-performance-improvements-of	-	-	-	84.2%
swin-transformer-hierarchical-vision	-	197M	-	87.3%
towards-all-in-one-pre-training-via	-	-	-	89.6%
colornet-investigating-the-importance-of	-	-	-	84.32%
nasvit-neural-architecture-search-for	-	-	-	81.8%
boosting-discriminative-visual-representation	-	25.6M	-	79.41%
muxconv-information-multiplexing-in	-	3.4M	-	75.3%
sp-vit-learning-2d-spatial-priors-for-vision	-	-	-	86.3%
localvit-bringing-locality-to-vision	-	13.5M	-	78.2%
efficientnetv2-smaller-models-and-faster	-	-	-	85.1%
rethinking-spatial-dimensions-of-vision	-	23.5M	-	81.9%
eca-net-efficient-channel-attention-for-deep	-	24.37M	-	77.48%
progressive-neural-architecture-search	-	86.1M	2.5G	82.9%
deit-iii-revenge-of-the-vit	-	22M	-	83.4%
levit-a-vision-transformer-in-convnet-s	-	8.8M	-	79.6%
transformer-in-transformer		65.6M		83.9%
when-vision-transformers-outperform-resnets	-	64M	-	79%
resnest-split-attention-networks	-	111M	-	84.5%
deit-iii-revenge-of-the-vit	-	-	-	85.2%
bottleneck-transformers-for-visual	-	28.02M	-	79.4%
fbnet-hardware-aware-efficient-convnet-design	-	5.5M	-	74.9%
dilated-neighborhood-attention-transformer	-	197M	-	87.4%
maxvit-multi-axis-vision-transformer	-	-	-	89.41%
convmlp-hierarchical-convolutional-mlps-for	-	17.4M	-	79%
asymmetric-masked-distillation-for-pre	-	22M	-	82.1%
masked-autoencoders-are-scalable-vision	-	656M	-	87.8%
mobilenetv4-universal-models-for-the-mobile	-	-	-	73.8%
circumventing-outliers-of-autoaugment-with	-	66M	-	85.5%
asymmnet-towards-ultralight-convolution	-	5.99M	-	75.4%
hvt-a-comprehensive-vision-framework-for	-	-	-	80.1%
190409460	-	-	-	77.8%
mobilevitv3-mobile-friendly-vision	-	1.4M	-	72.33%
shufflenet-an-extremely-efficient	-	-	-	70.9%
autoformer-searching-transformers-for-visual	-	22.9M	-	81.7%
going-deeper-with-image-transformers	-	270.9M	-	86.1%
transboost-improving-the-best-imagenet	-	5.29M	-	78.60%
learning-visual-representations-for-transfer-1		-		76.71%
which-transformer-to-favor-a-comparative	-	-	-	78.34%
meta-pseudo-labels	-	390M	-	90%
unireplknet-a-universal-perception-large	-	-	-	81.6%
vitae-vision-transformer-advanced-by	-	4.8M	-	76.8%
pvtv2-improved-baselines-with-pyramid-vision	-	13.1M	-	78.7%
asymmnet-towards-ultralight-convolution	-	2.8M	-	69.2%
scaling-vision-with-sparse-mixture-of-experts	-	656M	-	88.08%
efficient-multi-order-gated-aggregation	-	5.2M	-	80%
tinyvit-fast-pretraining-distillation-for	-	5.4M	-	80.7%
aggregating-nested-transformers	-	68M	-	83.8%
exploring-the-limits-of-weakly-supervised	-	88M	-	82.2%
co-training-2-l-submodels-for-visual	-	-	-	83.1%
cvt-introducing-convolutions-to-vision	-	18M	-	82.2%
uniformer-unifying-convolution-and-self	-	100M	-	85.6%
swin-transformer-hierarchical-vision	-	29M	-	81.3%
resmlp-feedforward-networks-for-image	-	15.4M	-	77.8%
metaformer-baselines-for-vision	-	40M	-	85.4%
a-simple-episodic-linear-probe-improves	-	-	-	76.13
vicinity-vision-transformer	-	61.8M	-	84.1%
patches-are-all-you-need-1	-	51.6M	-	82.20
dicenet-dimension-wise-convolutions-for	-	-	-	75.1%
perceiver-general-perception-with-iterative	-	-	-	76.4%
davit-dual-attention-vision-transformers	-	196.8M	-	87.5%
omnivore-a-single-model-for-many-visual	-	-	-	85.3%
autodropout-learning-dropout-patterns-to	-	-	-	78.7%
a-convnet-for-the-2020s	-	198M	-	85.5%
efficientvit-enhanced-linear-attention-for	-	-	-	83.5%
resnet-strikes-back-an-improved-training	-	22M	-	80.4%
eva-exploring-the-limits-of-masked-visual	-	1000M	-	89.7%
multiscale-deep-equilibrium-models		81M		79.2%
when-shift-operation-meets-vision-transformer	-	50M	-	82.8%
exploring-randomly-wired-neural-networks-for	-	5.6M	-	74.7%
splitnet-divide-and-co-training	-	98M	-	83.6%
three-things-everyone-should-know-about	-	-	-	82.6%
alphanet-improved-training-of-supernet-with	-	-	-	79.1%
dynamic-convolution-attention-over	-	7M	-	72.8%
x-volution-on-the-unification-of-convolution		-		76.6%
global-context-vision-transformers	-	12M	-	79.8%
uninet-unified-architecture-search-with-1	-	117M	-	87.4%
shufflenet-v2-practical-guidelines-for	-	-	-	75.4%
dropblock-a-regularization-method-for	-	-	-	78.35%
greedynas-towards-fast-one-shot-nas-with	-	5.2M	-	76.8%
alphanet-improved-training-of-supernet-with	-	-	-	79.4%
differentially-private-image-classification	-	-	-	88.9%
escaping-the-big-data-paradigm-with-compact	-	22.36M	-	-
visual-attention-network	-	13.9M	-	81.1%
fbnetv5-neural-architecture-search-for	-	-	-	81.7%
multigrain-a-unified-image-embedding-for	-	-	-	78.2%
a-fast-knowledge-distillation-framework-for	-	-	-	81.9%
fast-vision-transformers-with-hilo-attention	-	49M	-	83.3%
resnet-strikes-back-an-improved-training	-	60.2M	-	82.4%
scarletnas-bridging-the-gap-between	-	6.7M	-	76.9%
fastvit-a-fast-hybrid-vision-transformer	-	-	-	79.8%
maxvit-multi-axis-vision-transformer	-	31M	-	83.62%
an-improved-one-millisecond-mobile-backbone	-	10.1M	-	78.1%
2103-14899	-	28.2M	-	82.3%
randaugment-practical-data-augmentation-with	-	66M	-	85%
mixnet-mixed-depthwise-convolutional-kernels	-	7.3M	-	78.9%
sliced-recursive-transformer-1	-	4.8M	-	77.6%
muxconv-information-multiplexing-in	-	1.8M	-	66.7%
clcnet-rethinking-of-ensemble-modeling-with	-	-	-	83.88%
high-performance-large-scale-image	-	132.6M	-	84.7%
model-rubik-s-cube-twisting-resolution-depth	-	11.9M	-	79.4%
tokens-to-token-vit-training-vision	-	-	-	81.9%
dynamic-convolution-attention-over	-	4.8M	-	69.7%
moat-alternating-mobile-convolution-and	-	190M	-	86.7%
go-wider-instead-of-deeper	-	40M	-	79.49%
repvgg-making-vgg-style-convnets-great-again	-	55.77M	-	78.5%
metaformer-baselines-for-vision	-	39M	-	86.9%
co-training-2-l-submodels-for-visual	-	-	-	84.2%
metaformer-baselines-for-vision	-	26M	-	84.1%
davit-dual-attention-vision-transformers	-	362M	-	90.2%
transboost-improving-the-best-imagenet	-	5.48M	-	76.81%
glit-neural-architecture-search-for-global	-	96.1M	-	82.3%
pyramidal-convolution-rethinking		42.3M		81.49%
spatial-channel-token-distillation-for-vision	-	30.1M	-	82.1%
high-performance-large-scale-image	-	316.1M	-	85.9%
multimodal-autoregressive-pre-training-of	-	2700M	-	-
vision-transformer-with-deformable-attention	-	29M	-	82.0%
eca-net-efficient-channel-attention-for-deep	-	3.34M	-	72.56%
maxvit-multi-axis-vision-transformer	-	-	-	89.12%
three-things-everyone-should-know-about	-	-	-	84.3%
twins-revisiting-spatial-attention-design-in	-	99.2M	-	83.7%
visual-attention-network	-	-	-	85.7%
autoformer-searching-transformers-for-visual	-	5.7M	-	74.7%
fixing-the-train-test-resolution-discrepancy	-	-	-	79.8%
moat-alternating-mobile-convolution-and	-	27.8M	-	83.3%
revbifpn-the-fully-reversible-bidirectional	-	10.6M	-	79%
designing-network-design-spaces	-	11.2M	-	78%
tiny-models-are-the-computational-saver-for	-	-	-	83.52
efficientvit-enhanced-linear-attention-for	-	53M	-	84.5%
rethinking-spatial-dimensions-of-vision	-	73.8M	-	84%
visual-attention-network	-	200M	-	87.8%
ghostnetv3-exploring-the-training-strategies	-	-	-	69.4%
edgenext-efficiently-amalgamated-cnn	-	5.6M	-	79.4%
edgeformer-improving-light-weight-convnets-by	-	5M	-	78.63%
billion-scale-semi-supervised-learning-for	-	88M	-	84.3%
vision-gnn-an-image-is-worth-graph-of-nodes	-	92.6M	-	83.7%
global-context-vision-transformers	-	51M	-	84.0%
cvt-introducing-convolutions-to-vision	-	-	-	87.7%
fixing-the-train-test-resolution-discrepancy-2	-	66M	-	87.1%
autodropout-learning-dropout-patterns-to	-	-	-	77.5%
maxvit-multi-axis-vision-transformer	-	-	-	85.72%
designing-network-design-spaces	-	39.2M	-	79.9%
ghostnet-more-features-from-cheap-operations	-	6.5M	-	74.1%
involution-inverting-the-inherence-of	-	12.4M	-	77.6%
rexnet-diminishing-representational	-	9.7M	-	80.3%
adversarial-examples-improve-image	-	88M	-	85.5%
gtp-vit-efficient-vision-transformers-via	-	-	-	81.5%
Model 788	-	-	-	81.16%
resmlp-feedforward-networks-for-image	-	45M	-	79.7%
unireplknet-a-universal-perception-large	-	-	-	80.2%
biformer-vision-transformer-with-bi-level	-	-	-	85.4%
tokenlearner-what-can-8-learned-tokens-do-for	-	460M	-	88.87%
lets-keep-it-simple-using-simple	-	5.7M	-	71.94
rest-an-efficient-transformer-for-visual	-	51.63M	-	83.6%
ghostnet-more-features-from-cheap-operations	-	7.3M	-	75.7%
co-training-2-l-submodels-for-visual	-	-	-	85.0%
inception-v4-inception-resnet-and-the-impact	-	55.8M	-	80.1%
deep-residual-learning-for-image-recognition	-	-	-	78.57%
distilled-gradual-pruning-with-pruned-fine	-	1.15M	-	65.22
sliced-recursive-transformer-1	-	4M	-	74.0%
beit-bert-pre-training-of-image-transformers	-	86M	-	86.3%
metaformer-baselines-for-vision	-	56M	-	87.5%
designing-network-design-spaces	-	4.3M	-	74.1%
Model 804	-	-	-	78.62%
localvit-bringing-locality-to-vision	-	6.3M	-	75.9%
metaformer-baselines-for-vision	-	56M	-	86.2%
coatnet-marrying-convolution-and-attention	-	-	-	88.52%
tinyvit-fast-pretraining-distillation-for	-	21M	-	83.1%
debiformer-vision-transformer-with-deformable	-	-	-	81.9%
metaformer-baselines-for-vision	-	100M	-	87.0%
going-deeper-with-image-transformers	-	38.6M	-	84.8%
drawing-multiple-augmentation-samples-per		377.2M		86.78%
multimodal-autoregressive-pre-training-of	-	-	-	89.5%
scaling-vision-transformers-to-22-billion	-	307M	-	89.6%
metaformer-is-actually-what-you-need-for	-	73M	-	82.5%
visual-attention-network	-	4.1M	-	75.4%
gtp-vit-efficient-vision-transformers-via	-	-	-	79.5%
sliced-recursive-transformer-1	-	21M	-	83.8%
ghostnet-more-features-from-cheap-operations	-	5.2M	-	73.9%
lip-local-importance-based-pooling	-	25.8M	-	78.15%
elsa-enhanced-local-self-attention-for-vision	-	27M	-	84.7%
mnasnet-platform-aware-neural-architecture	-	4.8M	-	75.6%
glit-neural-architecture-search-for-global	-	7.2M	-	76.3%
autoformer-searching-transformers-for-visual	-	54M	-	82.4%
which-transformer-to-favor-a-comparative	-	-	-	71.53%
fixing-the-train-test-resolution-discrepancy-2	-	7.8M	-	82.6%
automix-unveiling-the-power-of-mixup	-	11.7M	-	72.05%
deepvit-towards-deeper-vision-transformer	-	-	-	83.1%
unireplknet-a-universal-perception-large	-	-	-	77%
knowledge-distillation-a-good-teacher-is		-		82.8%
cas-vit-convolutional-additive-self-attention	-	5.76	-	81.1%
mobilevitv3-mobile-friendly-vision	-	3M	-	76.55%
internimage-exploring-large-scale-vision	-	30M	-	83.5%
meta-pseudo-labels	-	-	-	83.2%
training-data-efficient-image-transformers		87M		85.2%
mobilevitv3-mobile-friendly-vision	-	2.5M	-	76.7%
vicinity-vision-transformer	-	61.8M	-	84.7%
a-dot-product-attention-free-transformer	-	23M	-	80.1%
regularized-evolution-for-image-classifier	-	469M	-	83.9%
an-image-is-worth-16x16-words-transformers-1	-	-	-	88.55%
coatnet-marrying-convolution-and-attention	-	168M	-	84.5%
evo-vit-slow-fast-token-evolution-for-dynamic	-	39.6M	-	82.2%
swin-transformer-v2-scaling-up-capacity-and	-	88M	-	87.1%
sequencer-deep-lstm-for-image-classification	-	38M	-	82.8%
maxvit-multi-axis-vision-transformer	-	-	-	88.51%
cspnet-a-new-backbone-that-can-enhance	-	20.5M	-	79.8%
clcnet-rethinking-of-ensemble-modeling-with	-	-	-	86.46%
spatial-group-wise-enhance-improving-semantic	-	44.55M	-	78.798%
2103-15358	-	79M	-	81.9%
mobilenetv4-universal-models-for-the-mobile	-	-	-	83.4%
efficient-multi-order-gated-aggregation	-	44M	-	84.3%
revbifpn-the-fully-reversible-bidirectional	-	82M	-	83.7%
a-fast-knowledge-distillation-framework-for	-	-	-	80.1%
uninet-unified-architecture-search-with	-	22.5M	-	82.7%
collaboration-of-experts-achieving-80-top-1	-	95.3M	-	80.7%
self-training-with-noisy-student-improves	51800G	480M		88.4%
gpipe-efficient-training-of-giant-neural	-	-	-	84.4%
proxylessnas-direct-neural-architecture	-	4.0M	-	74.6%
densenets-reloaded-paradigm-shift-beyond	-	24M	-	82.8%
efficient-multi-order-gated-aggregation	-	181M	-	87.8%
go-wider-instead-of-deeper	-	63M	-	80.09%
a-convnet-for-the-2020s	-	29M	-	82.1%
rethinking-the-design-principles-of-robust	-	91.8M	-	82.7%
densenets-reloaded-paradigm-shift-beyond	-	186M	-	84.8%
deit-iii-revenge-of-the-vit	-	-	-	86.7%
metaformer-baselines-for-vision	-	100M	-	84.8%
training-data-efficient-image-transformers	-	5M	-	76.6%
vitae-vision-transformer-advanced-by	-	6.5M	-	77.9%
harmonic-convolutional-networks-based-on	-	88.2M	-	82.85%
bottleneck-transformers-for-visual	-	75.1M	-	84.7%
bottleneck-transformers-for-visual	-	-	-	84.2%
the-effectiveness-of-mae-pre-pretraining-for	-	-	-	88.8%
resmlp-feedforward-networks-for-image	-	30M	-	80.8%
hyenapixel-global-image-context-with	-	-	-	83.6%
mambavision-a-hybrid-mamba-transformer-vision	-	50.1M	-	83.3%
resnet-strikes-back-an-improved-training	-	60.2M	-	81.8%
multiscale-vision-transformers	-	37M	-	83.0%
multimodal-autoregressive-pre-training-of	-	300M	-	86.6%
sharpness-aware-minimization-for-efficiently-1	-	480M	-	88.61%
bottleneck-transformers-for-visual	-	54.7M	-	82.8%
internimage-exploring-large-scale-vision	-	223M	-	87.7%
nasvit-neural-architecture-search-for	-	-	-	82.9%
ghostnetv3-exploring-the-training-strategies	-	-	-	77.1%
fbnetv5-neural-architecture-search-for	-	-	-	82.6%
lets-keep-it-simple-using-simple	-	5.7M	-	79.12
levit-a-vision-transformer-in-convnet-s	-	10.4M	-	80%
transnext-robust-foveal-visual-perception-for	-	49.7M	-	84.7%
the-effectiveness-of-mae-pre-pretraining-for	-	2000M	-	89.8%
metaformer-baselines-for-vision	-	57M	-	86.9%
hyenapixel-global-image-context-with	-	-	-	83.2%
meta-pseudo-labels	95040G	480M		90.2%
improved-multiscale-vision-transformers-for	-	24M	-	82.3%
coatnet-marrying-convolution-and-attention	-	25M	-	81.6%
data2vec-a-general-framework-for-self-1	-	656M	-	86.6%
internimage-exploring-large-scale-vision	-	1080M	-	89.6%
colormae-exploring-data-independent-masking	-	-	-	83.8%
maxvit-multi-axis-vision-transformer	-	-	-	86.19%
kolmogorov-arnold-transformer	-	86.6M	-	79.1
cvt-introducing-convolutions-to-vision	-	-	-	83.3%
vision-transformer-with-deformable-attention	-	88M	-	84.8%
dilated-neighborhood-attention-transformer	-	28M	-	82.7%
which-transformer-to-favor-a-comparative	-	-	-	80.66%
hiera-a-hierarchical-vision-transformer	-	-	-	86.9%
an-improved-one-millisecond-mobile-backbone	-	7.8M	-	79.1%
metaformer-baselines-for-vision	-	27M	-	84.4%
multigrain-a-unified-image-embedding-for	-	-	-	81.3%
xcit-cross-covariance-image-transformers	-	26M	-	85.1%
residual-attention-network-for-image	-	-	-	80.5%
fairnas-rethinking-evaluation-fairness-of	-	4.5M	-	75.10%
beyond-self-attention-external-attention		-		81.7%
minivit-compressing-vision-transformers-with	-	47M	-	85.5%
learned-queries-for-efficient-local-attention	-	16M	-	81.7%
token-labeling-training-a-85-5-top-1-accuracy	-	151M	-	86.4%
meal-v2-boosting-vanilla-resnet-50-to-80-top	-	25.6M	-	81.72%
how-to-use-dropout-correctly-on-residual	-	-	-	79.152%
densenets-reloaded-paradigm-shift-beyond	-	186M	-	85.8%
cvt-introducing-convolutions-to-vision	-	-	-	82.5%
activemlp-an-mlp-like-architecture-with	-	76.4M	-	84.8%
moga-searching-beyond-mobilenetv3		5.1M	0.0304G	75.9%
mlp-mixer-an-all-mlp-architecture-for-vision		-		87.94%
fixing-the-train-test-resolution-discrepancy-2	-	12M	-	85%
mixnet-mixed-depthwise-convolutional-kernels	-	5.0M	-	77%
learned-queries-for-efficient-local-attention	-	56M	-	83.7%
ghostnetv3-exploring-the-training-strategies	-	-	-	80.4%
ghostnetv3-exploring-the-training-strategies	-	-	-	79.1%
190409460	-	-	-	79.03%
searching-for-mobilenetv3	-	5.4M	-	75.2%
clcnet-rethinking-of-ensemble-modeling-with	-	-	-	85.28%
convit-improving-vision-transformers-with	-	6M	-	73.1%
eca-net-efficient-channel-attention-for-deep	-	42.49M	-	78.65%
metaformer-baselines-for-vision	-	99M	-	85.5%
fastervit-fast-vision-transformers-with	-	957.5M	-	85.6%
transboost-improving-the-best-imagenet	-	22.05M	-	83.67%
fastervit-fast-vision-transformers-with	-	31.4M	-	82.1%
hyenapixel-global-image-context-with	-	-	-	85.2%
which-transformer-to-favor-a-comparative	-	-	-	82.22%
resnet-strikes-back-an-improved-training	-	25M	-	78.1%
balanced-binary-neural-networks-with-gated	-	-	-	62.6%
from-xception-to-nexception-new-design	-	-	-	81.5%
gswin-gated-mlp-vision-model-with	-	21.8M	-	81.71%
semantic-aware-local-global-vision	-	6.5M	-	75.9%
mobilenetv2-inverted-residuals-and-linear	-	3.4M	-	72%
incepformer-efficient-inception-transformer	-	39.3M	-	83.6%
volo-vision-outlooker-for-visual-recognition	-	193M	-	86.8%
tiny-models-are-the-computational-saver-for	-	-	-	85.74
which-transformer-to-favor-a-comparative	-	-	-	83.09%
metaformer-baselines-for-vision	-	99M	-	88.1%
multimodal-autoregressive-pre-training-of	-	1200M	-	88.1%
revisiting-unreasonable-effectiveness-of-data	-	-	-	79.2%
going-deeper-with-image-transformers	-	12M	-	80.9%
pvtv2-improved-baselines-with-pyramid-vision	-	82M	-	83.8%
hvt-a-comprehensive-vision-framework-for	-	-	-	85%
metaformer-baselines-for-vision	-	99M	-	87.4%
averaging-weights-leads-to-wider-optima-and	-	-	-	78.94%
differentiable-top-k-classification-learning-1	-	-	-	88.37%
fixing-the-train-test-resolution-discrepancy-2	-	43M	-	86.7%
fastvit-a-fast-hybrid-vision-transformer	-	-	-	84.5%
coatnet-marrying-convolution-and-attention	-	-	-	87.1%
sp-vit-learning-2d-spatial-priors-for-vision	-	-	-	86%
mobilenetv2-inverted-residuals-and-linear	-	6.9M	-	74.7%
mobilenetv4-universal-models-for-the-mobile	-	-	-	82.9%
masked-autoencoders-are-scalable-vision	-	-	-	83.6%
vitae-vision-transformer-advanced-by	-	19.2M	-	82.2%
convit-improving-vision-transformers-with	-	27M	-	81.3%
fastervit-fast-vision-transformers-with	-	1360M	-	85.8%
resnet-strikes-back-an-improved-training	-	25M	-	80.4%
training-data-efficient-image-transformers	-	22M	-	82.6%
incepformer-efficient-inception-transformer	-	24.3M	-	82.9%
soft-conditional-computation	-	-	-	78.3%
transnext-robust-foveal-visual-perception-for	-	28.2M	-	84.0%
go-wider-instead-of-deeper	-	29M	-	77.54%
multigrain-a-unified-image-embedding-for	-	-	-	82.7%
an-improved-one-millisecond-mobile-backbone	-	4.8M	-	77.4%
large-scale-learning-of-general-visual	-	-	-	87.54%
augmenting-convolutional-networks-with	-	25.2M	-	82.1%
dynamic-convolution-attention-over	-	42.7M	-	72.7%
efficientnetv2-smaller-models-and-faster	-	-	-	83.9%
fastvit-a-fast-hybrid-vision-transformer	-	-	-	79.1%
revit-enhancing-vision-transformers-with	-	-	-	82.4
bottleneck-transformers-for-visual	-	33.5M	-	81.7%
self-knowledge-distillation-a-simple-way-for	-	-	-	79.24%
resmlp-feedforward-networks-for-image	-	116M	-	83.6%
tinyvit-fast-pretraining-distillation-for	-	5.4M	-	79.1%
190409460	-	-	-	79.38%
densely-connected-convolutional-networks	-	-	-	77.85%
augmenting-convolutional-networks-with	-	47.7M	-	83.2%
fast-vision-transformers-with-hilo-attention	-	-	-	83.6%
glit-neural-architecture-search-for-global	-	24.6M	-	80.5%
fixing-the-train-test-resolution-discrepancy-2	-	5.3M	-	80.2%
sharpness-aware-minimization-for-efficiently-1	-	-	-	81.6%
high-performance-large-scale-image	-	377.2M	-	86.3%
tokens-to-token-vit-training-vision	-	39.2M	-	82.2%
densely-connected-search-space-for-more	-	-	-	75.9%
the-effectiveness-of-mae-pre-pretraining-for	-	-	-	86.8%
davit-dual-attention-vision-transformers	-	1437M	-	90.4%
tinyvit-fast-pretraining-distillation-for	-	21M	-	86.5%
not-all-images-are-worth-16x16-words-dynamic	-	-	-	79.74%
graph-convolutions-enrich-the-self-attention	-	-	-	81.5%
efficientnetv2-smaller-models-and-faster	-	208M	-	87.3%
res2net-a-new-multi-scale-backbone	-	-	-	81.23%
internimage-exploring-large-scale-vision	-	335M	-	88%
fastvit-a-fast-hybrid-vision-transformer	-	-	-	75.6%
torchdistill-a-modular-configuration-driven	-	-	-	71.37%
adversarial-autoaugment-1	-	-	-	81.32%
bottleneck-transformers-for-visual	-	25.5M	-	78.8%
efficientnetv2-smaller-models-and-faster	-	120M	-	86.8%
rexnet-diminishing-representational	-	7.6M	-	79.5%
designing-network-design-spaces	-	20.6M	-	79.4%
maxvit-multi-axis-vision-transformer	-	-	-	88.82%
revisiting-weakly-supervised-pre-training-of	-	633.5M	-	88.6%
clcnet-rethinking-of-ensemble-modeling-with	-	-	-	86.61%
efficientnet-rethinking-model-scaling-for	-	30M	-	83.3%
unireplknet-a-universal-perception-large	-	-	-	78.6%
masked-autoencoders-are-scalable-vision	-	-	-	86.9%
lambdanetworks-modeling-long-range-1		42M		84.3%
three-things-everyone-should-know-about	-	-	-	82.3%
repmlpnet-hierarchical-vision-mlp-with-re	-	-	-	81.8%
densenets-reloaded-paradigm-shift-beyond	-	87M	-	84.4%
rethinking-and-improving-relative-position	-	87M	-	82.4%
metaformer-baselines-for-vision	-	57M	-	85.6%
metaformer-baselines-for-vision	-	39M	-	85.7%
maxvit-multi-axis-vision-transformer	-	-	-	88.38%
meta-knowledge-distillation	-	-	-	77.1%
sequencer-deep-lstm-for-image-classification	-	54M	-	84.6%
Model 1025	-	-	-	78.8
fixing-the-train-test-resolution-discrepancy-2	-	30M	-	86.4%
meal-v2-boosting-vanilla-resnet-50-to-80-top	-	-	-	73.19%
meta-knowledge-distillation	-	-	-	86.5%
transboost-improving-the-best-imagenet	-	28.59M	-	82.46%
improved-multiscale-vision-transformers-for	-	218M	-	88.4%
aggregated-residual-transformations-for-deep	-	83.6M	-	80.9%
deeper-vs-wider-a-revisit-of-transformer	-	-	-	86.3
an-improved-one-millisecond-mobile-backbone	-	7.8	-	77.4%
deep-polynomial-neural-networks	-	-	-	77.17%
improving-vision-transformers-by-revisiting	-	295.5M	-	87.3%
scalable-visual-transformers-with	-	21.74M	-	78.00%
revbifpn-the-fully-reversible-bidirectional	-	19.6M	-	81.1%
localvit-bringing-locality-to-vision	-	5.9M	-	74.8%
contextual-transformer-networks-for-visual	-	23.1M	-	81.6%
maxvit-multi-axis-vision-transformer	-	-	-	89.36%
mlp-mixer-an-all-mlp-architecture-for-vision	-	46M	-	76.44%
rexnet-diminishing-representational	-	4.1M	-	77.2%
an-intriguing-failing-of-convolutional-neural	-	-	-	75.74%
efficientvit-enhanced-linear-attention-for	-	24M	-	82.7%
scaling-vision-with-sparse-mixture-of-experts	-	3400M	-	87.41%
lip-local-importance-based-pooling	-	8.7M	-	76.64%
vitae-vision-transformer-advanced-by	-	48.5M	-	83.6%
the-effectiveness-of-mae-pre-pretraining-for	-	650M	-	89.5%
multimodal-autoregressive-pre-training-of	-	600M	-	87.5%
going-deeper-with-image-transformers	-	26.6M	-	84.1%
visual-parser-representing-part-whole	-	-	-	84.2%
augmenting-convolutional-networks-with	-	334.3M	-	87.1%
learnable-polynomial-trigonometric-and	-	28M	-	82.34
sliced-recursive-transformer-1	-	71.2M	-	84.8%
contextual-transformer-networks-for-visual	-	40.9M	-	83.2%
model-rubik-s-cube-twisting-resolution-depth	-	5.1M	-	77.7%
augmenting-convolutional-networks-with	-	25.2M	-	85.4%
meta-knowledge-distillation	-	-	-	85.1%