
摘要
卷积神经网络(ConvNets)通常在固定的资源预算下开发,然后在有更多资源可用时进行扩展以提高精度。本文系统地研究了模型扩展,并发现仔细平衡网络深度、宽度和分辨率可以带来更好的性能。基于这一观察,我们提出了一种新的扩展方法,该方法使用一个简单但非常有效的复合系数来均匀地扩展深度、宽度和分辨率的所有维度。我们在扩大MobileNets和ResNet方面展示了这种方法的有效性。为了进一步提升性能,我们利用神经架构搜索设计了一个新的基线网络,并将其扩展以获得一系列模型,称为EfficientNets,这些模型在准确性和效率上均显著优于以往的ConvNets。特别是,我们的EfficientNet-B7在ImageNet数据集上达到了最先进的84.3%的Top-1精度,同时其规模比现有的最佳ConvNet小8.4倍,推理速度也快6.1倍。此外,我们的EfficientNets在网络迁移学习中表现优异,在CIFAR-100(91.7%)、Flowers(98.8%)和其他三个迁移学习数据集上均达到了最先进的精度,参数量却减少了数量级。源代码位于https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet。
代码仓库
darya-baranovskaya/keyword_spotting
pytorch
GitHub 中提及
rwightman/efficientnet-jax
jax
GitHub 中提及
filipmu/Kaggle-APTOS-2019-Blindness
pytorch
GitHub 中提及
houstonsantos/CassavaLeafDisease
GitHub 中提及
JacobM184/EfficientNet-for-Gun-detection
pytorch
GitHub 中提及
amirdy/dog-breed-classification
pytorch
GitHub 中提及
lukemelas/EfficientNet-PyTorch
pytorch
GitHub 中提及
prakhargoyal106/MelanomaClassification
pytorch
GitHub 中提及
rmwkwok/product_visual_search
tf
GitHub 中提及
facebookresearch/pycls
pytorch
GitHub 中提及
clovaai/rexnet
pytorch
GitHub 中提及
james77777778/keras-image-models
pytorch
GitHub 中提及
konstantinos-p/image_classification_SOTA
pytorch
GitHub 中提及
toanbkmt/EfficientnetFruitDetect
pytorch
GitHub 中提及
mnikitin/EfficientNet
mxnet
GitHub 中提及
pikkaay/efficientnet_gpu
tf
GitHub 中提及
triple7/Keras-WGAN-RGB-128x128
tf
GitHub 中提及
lpirola13/flower_recognizer
tf
GitHub 中提及
maxwelltsai/DeepGalaxy
tf
GitHub 中提及
pytorch/vision
pytorch
rajneeshaggarwal/google-efficientnet
tf
GitHub 中提及
titu1994/keras-efficientnets
tf
GitHub 中提及
seunghwan1228/CNN_EfficientNet
GitHub 中提及
rohitgr7/tvmodels
pytorch
rwightman/pytorch-image-models
pytorch
GitHub 中提及
reyvaz/steel-defect-segmentation
tf
GitHub 中提及
wusaifei/HWCC_image_classification
pytorch
GitHub 中提及
buiquangmanhhp1999/age_gender_estimation
GitHub 中提及
lvweiwolf/efficientdet
tf
GitHub 中提及
reyvaz/pneumothorax_detection
tf
GitHub 中提及
denizyuret/playground
pytorch
GitHub 中提及
Mind23-2/MindCode-37
mindspore
hyang0129/foodclassapp
tf
GitHub 中提及
PaddlePaddle/PaddleClas
paddle
jaketae/mlp-mixer
pytorch
GitHub 中提及
ZackPashkin/YOLOv3-EfficientNet-EffYolo
tf
GitHub 中提及
filaPro/visda2019
tf
GitHub 中提及
miramind/efficientnets_pytorch
pytorch
GitHub 中提及
federicopozzi33/MobileOne-PyTorch
pytorch
GitHub 中提及
Cyprien0105/DataScience
GitHub 中提及
lpirola13/flower-recognizer
tf
GitHub 中提及
open-mmlab/mmdetection
pytorch
nimiew/Grab-Computer-Vision
GitHub 中提及
AmirmohammadRostami/KeywordsSpotting-EfficientNet-A0
pytorch
GitHub 中提及
christiansafka/img2vec
pytorch
GitHub 中提及
isaachaw/GrabCarRecognition
pytorch
GitHub 中提及
kairess/efficientnet_example
GitHub 中提及
DeepBrainsMe/FSnet
pytorch
GitHub 中提及
facebookresearch/ClassyVision
pytorch
GitHub 中提及
kdha0727/lung-opacity-and-covid-chest-x-ray-detection
pytorch
GitHub 中提及
danielpatrickhug/Research_Paper_Parser
GitHub 中提及
qubvel/efficientnet
tf
GitHub 中提及
JoegameZhou/efficientnet-b0
mindspore
GitHub 中提及
IMvision12/keras-vision-models
pytorch
GitHub 中提及
mvenouziou/Project-Attention-Is-What-You-Get
tf
GitHub 中提及
BobMcDear/pytorch-efficientnet
pytorch
GitHub 中提及
wangyi111/international-archaeology-ai-challenge
pytorch
GitHub 中提及
abhuse/pytorch-efficientnet
pytorch
GitHub 中提及
tsing-cv/EfficientNet-tensorflow-eager
tf
GitHub 中提及
xslidi/EfficientNets_ddl_apex
pytorch
GitHub 中提及
iamilyasedunov/key_word_spotting
pytorch
GitHub 中提及
Jmak12/Iris1
pytorch
GitHub 中提及
mingxingtan/efficientnet
tf
GitHub 中提及
asad-62/IVP-DNN
tf
GitHub 中提及
osmr/imgclsmob
mxnet
GitHub 中提及
rwightman/gen-efficientnet-pytorch
pytorch
GitHub 中提及
jason90330/EdgeFinal
pytorch
GitHub 中提及
tarikdgny/Optical_Coherence_Tomography
GitHub 中提及
BenjiKCF/EfficientNet
GitHub 中提及
luuchung/cifar-100
GitHub 中提及
canturan10/satellighte
pytorch
GitHub 中提及
epoc88/PFLD_68pts_Pytorch_2020
mxnet
GitHub 中提及
zsef123/EfficientNets-PyTorch
pytorch
GitHub 中提及
cgebbe/kaggle_pku-autonomous-driving
pytorch
GitHub 中提及
lyqcom/efficientnet
mindspore
GitHub 中提及
WonJunPark/Efficientnet
GitHub 中提及
captaindario/dakanji-single-kanji-recognition
tf
GitHub 中提及
setharram/facenet
tf
GitHub 中提及
semskurto/APTOS
GitHub 中提及
marcointrovigne/WeatherDetection
tf
GitHub 中提及
PotatoSpudowski/CactiNet
pytorch
GitHub 中提及
ravi02512/efficientdet-keras
tf
GitHub 中提及
tensorflow/tpu/tree/master/models/official/efficientnet
官方
tf
GitHub 中提及
maragori/DeepfakeForensics-v1
pytorch
GitHub 中提及
HyeonhoonLee/MAIC2021_Sleep
pytorch
GitHub 中提及
najlaeLemrabet/FacialKeypointsDetection
pytorch
GitHub 中提及
TravisLeeTS/grabcvchallenge
tf
GitHub 中提及
2023-MindSpore-1/ms-code-195
mindspore
ultralytics/yolov5
pytorch
GitHub 中提及
Legoons/Melanoma_classification
pytorch
GitHub 中提及
Mayurji/Image-Classification-PyTorch
pytorch
GitHub 中提及
2023-MindSpore-1/ms-code-151
mindspore
northeastsquare/effficientnet
tf
GitHub 中提及
chrisqqq123/FA-Dist-EfficientNet
pytorch
GitHub 中提及
gomezzz/distmsmatch
pytorch
GitHub 中提及
Deci-AI/super-gradients
pytorch
GitHub 中提及
HO4X/TSR_JetsonTX2
GitHub 中提及
huynhtuan17ti/AI4VN-Hackathon2020
tf
GitHub 中提及
linhduongtuan/Fruits_Vegetables_Classifier_WebApp
pytorch
GitHub 中提及
ckyrkou/EmergencyNet
tf
GitHub 中提及
SunDoge/efficientnet-pytorch
pytorch
GitHub 中提及
DableUTeeF/keras-efficientnet
tf
GitHub 中提及
vladthesav/MoldAI
pytorch
GitHub 中提及
shijianjian/efficientnet-pytorch-3d
pytorch
GitHub 中提及
open-edge-platform/geti
pytorch
GitHub 中提及
armin-azh/3DefficientNet
tf
GitHub 中提及
github-luffy/PFLD_68points_Pytorch
mxnet
GitHub 中提及
VinayBhupalam/melanoma-detection
tf
GitHub 中提及
Jintao-Huang/EfficientNet_PyTorch
pytorch
GitHub 中提及
gomezzz/MSMatch
pytorch
GitHub 中提及
js-aguiar/wheat-object-detection
pytorch
GitHub 中提及
SifatMd/Research-Papers
GitHub 中提及
6210612757/facerecognition
tf
GitHub 中提及
narumiruna/efficientnet-pytorch
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| domain-generalization-on-vizwiz | EfficientNet-B4 | Accuracy - All Images: 41.7 Accuracy - Clean Images: 46.4 Accuracy - Corrupted Images: 35.6 |
| domain-generalization-on-vizwiz | EfficientNet-B2 | Accuracy - All Images: 38.1 Accuracy - Clean Images: 42.8 Accuracy - Corrupted Images: 31.4 |
| domain-generalization-on-vizwiz | EfficientNet-B1 | Accuracy - All Images: 36.7 Accuracy - Clean Images: 41.5 Accuracy - Corrupted Images: 30.9 |
| domain-generalization-on-vizwiz | EfficientNet-B5 | Accuracy - All Images: 42.8 Accuracy - Clean Images: 47.3 Accuracy - Corrupted Images: 37 |
| domain-generalization-on-vizwiz | EfficientNet-B3 | Accuracy - All Images: 40.7 Accuracy - Clean Images: 45.3 Accuracy - Corrupted Images: 34.2 |
| domain-generalization-on-vizwiz | EfficientNet-B0 | Accuracy - All Images: 34.2 Accuracy - Clean Images: 38.4 Accuracy - Corrupted Images: 27.4 |
| fine-grained-image-classification-on-birdsnap | EfficientNet-B7 | Accuracy: 84.3% |
| fine-grained-image-classification-on-fgvc | EfficientNet-B7 | Accuracy: 92.9 |
| fine-grained-image-classification-on-food-101 | EfficientNet-B7 | Accuracy: 93.0 |
| fine-grained-image-classification-on-oxford-1 | EfficientNet-B7 | Accuracy: 95.4% |
| fine-grained-image-classification-on-stanford | EfficientNet-B7 | Accuracy: 94.7% |
| image-classification-on-cifar-10 | EfficientNet-B7 | Percentage correct: 98.9 |
| image-classification-on-cifar-100 | EfficientNet-B7 | PARAMS: 64M Percentage correct: 91.7 |
| image-classification-on-flowers-102 | EfficientNet-B7 | Accuracy: 98.8% |
| image-classification-on-gashissdb | EfficientNet-b0 | Accuracy: 98.11 F1-Score: 99.01 Precision: 99.94 |
| image-classification-on-imagenet | EfficientNet-B7 | GFLOPs: 37 Number of params: 66M Top 1 Accuracy: 84.4% |
| image-classification-on-imagenet | EfficientNet-B2 | GFLOPs: 1 Number of params: 9.2M Top 1 Accuracy: 79.8% |
| image-classification-on-imagenet | EfficientNet-B3 | Number of params: 12M Top 1 Accuracy: 81.1% |
| image-classification-on-imagenet | EfficientNet-B0 | GFLOPs: 0.39 Number of params: 5.3M Top 1 Accuracy: 76.3% |
| image-classification-on-imagenet | EfficientNet-B6 | GFLOPs: 19 Number of params: 43M Top 1 Accuracy: 84% |
| image-classification-on-imagenet | EfficientNet-B4 | GFLOPs: 4.2 Number of params: 19M Top 1 Accuracy: 82.6% |
| image-classification-on-imagenet | EfficientNet-B1 | GFLOPs: 0.7 Number of params: 7.8M Top 1 Accuracy: 78.8% |
| image-classification-on-imagenet | EfficientNet-B5 | GFLOPs: 9.9 Number of params: 30M Top 1 Accuracy: 83.3% |
| image-classification-on-omnibenchmark | EfficientNetB4 | Average Top-1 Accuracy: 35.8 |
| medical-image-classification-on-nct-crc-he | Efficientnet-b0 | Accuracy (%): 95.59 F1-Score: 97.48 Precision: 99.89 Specificity: 99.45 |