EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks

ICML 2019 Mingxing TanQuoc V. Le

Convolutional Neural Networks (ConvNets) are commonly developed at a fixed resource budget, and then scaled up for better accuracy if more resources are available. In this paper, we systematically study model scaling and identify that carefully balancing network depth, width, and resolution can lead to better performance... (read more)

PDF Abstract

Evaluation results from the paper


 SOTA for Image Classification on Stanford Cars (using extra training data)

     Get a GitHub badge
Task Dataset Model Metric name Metric value Global rank Uses extra
training data
Compare
Image Classification CIFAR-10 EfficientNet Percentage correct 98.9 # 2
Image Classification CIFAR-100 EfficientNet Percentage correct 91.7 # 1
Image Classification Flowers-102 EfficientNet Accuracy 98.8% # 1
Image Classification ImageNet EfficientNet-B3 Top 1 Accuracy 81.1% # 19
Image Classification ImageNet EfficientNet-B3 Top 5 Accuracy 95.5% # 17
Image Classification ImageNet EfficientNet-B3 Number of params 12M # 1
Image Classification ImageNet EfficientNet-B2 Top 1 Accuracy 79.8% # 26
Image Classification ImageNet EfficientNet-B2 Top 5 Accuracy 94.9% # 21
Image Classification ImageNet EfficientNet-B2 Number of params 9.2M # 1
Image Classification ImageNet EfficientNet-B1 Top 1 Accuracy 78.8% # 35
Image Classification ImageNet EfficientNet-B1 Top 5 Accuracy 94.4% # 31
Image Classification ImageNet EfficientNet-B1 Number of params 7.8M # 1
Image Classification ImageNet EfficientNet-B6 Top 1 Accuracy 84.0% # 7
Image Classification ImageNet EfficientNet-B6 Top 5 Accuracy 96.9% # 7
Image Classification ImageNet EfficientNet-B6 Number of params 43M # 1
Image Classification ImageNet EfficientNet-B7 Top 1 Accuracy 84.4% # 4
Image Classification ImageNet EfficientNet-B7 Top 5 Accuracy 97.1% # 5
Image Classification ImageNet EfficientNet-B7 Number of params 66M # 1
Image Classification ImageNet EfficientNet-B0 Top 1 Accuracy 76.3% # 52
Image Classification ImageNet EfficientNet-B0 Top 5 Accuracy 93.2% # 44
Image Classification ImageNet EfficientNet-B0 Number of params 5.3M # 1
Image Classification ImageNet EfficientNet-B5 Top 1 Accuracy 83.3% # 11
Image Classification ImageNet EfficientNet-B5 Top 5 Accuracy 96.7% # 9
Image Classification ImageNet EfficientNet-B5 Number of params 30M # 1
Image Classification ImageNet EfficientNet-B4 Top 1 Accuracy 82.6% # 15
Image Classification ImageNet EfficientNet-B4 Top 5 Accuracy 96.3% # 12
Image Classification ImageNet EfficientNet-B4 Number of params 19M # 1
Image Classification Stanford Cars EfficientNet Accuracy 94.7% # 1