EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks

ICML 2019  ·  Mingxing Tan, Quoc 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. Based on this observation, we propose a new scaling method that uniformly scales all dimensions of depth/width/resolution using a simple yet highly effective compound coefficient. We demonstrate the effectiveness of this method on scaling up MobileNets and ResNet. To go even further, we use neural architecture search to design a new baseline network and scale it up to obtain a family of models, called EfficientNets, which achieve much better accuracy and efficiency than previous ConvNets. In particular, our EfficientNet-B7 achieves state-of-the-art 84.3% top-1 accuracy on ImageNet, while being 8.4x smaller and 6.1x faster on inference than the best existing ConvNet. Our EfficientNets also transfer well and achieve state-of-the-art accuracy on CIFAR-100 (91.7%), Flowers (98.8%), and 3 other transfer learning datasets, with an order of magnitude fewer parameters. Source code is at https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet.

PDF Abstract
Task Dataset Model Metric Name Metric Value Global Rank Uses Extra
Training Data
Result Benchmark
Fine-Grained Image Classification Birdsnap EfficientNet-B7 Accuracy 84.3% # 2
Image Classification CIFAR-10 EfficientNet-B7 Percentage correct 98.9 # 23
PARAMS 64M # 230
Image Classification CIFAR-100 EfficientNet-B7 Percentage correct 91.7 # 19
PARAMS 64M # 196
Fine-Grained Image Classification FGVC Aircraft EfficientNet-B7 Accuracy 92.9 # 29
Image Classification Flowers-102 EfficientNet-B7 Accuracy 98.8% # 17
Fine-Grained Image Classification Food-101 EfficientNet-B7 Accuracy 93.0 # 8
Image Classification GasHisSDB EfficientNet-b0 Accuracy 98.11 # 6
Precision 99.94 # 3
F1-Score 99.01 # 6
Image Classification ImageNet EfficientNet-B4 Top 1 Accuracy 82.6% # 474
Number of params 19M # 528
GFLOPs 4.2 # 202
Image Classification ImageNet EfficientNet-B3 Top 1 Accuracy 81.1% # 607
Number of params 12M # 496
Image Classification ImageNet EfficientNet-B5 Top 1 Accuracy 83.3% # 403
Number of params 30M # 646
GFLOPs 9.9 # 299
Image Classification ImageNet EfficientNet-B7 Top 1 Accuracy 84.4% # 299
Number of params 66M # 777
GFLOPs 37 # 408
Image Classification ImageNet EfficientNet-B0 Top 1 Accuracy 76.3% # 847
Number of params 5.3M # 415
GFLOPs 0.39 # 42
Image Classification ImageNet EfficientNet-B1 Top 1 Accuracy 78.8% # 738
Number of params 7.8M # 459
GFLOPs 0.7 # 83
Image Classification ImageNet EfficientNet-B2 Top 1 Accuracy 79.8% # 676
Number of params 9.2M # 468
GFLOPs 1 # 103
Image Classification ImageNet EfficientNet-B6 Top 1 Accuracy 84% # 336
Number of params 43M # 694
GFLOPs 19 # 366
Medical Image Classification NCT-CRC-HE-100K Efficientnet-b0 Accuracy (%) 95.59 # 1
F1-Score 97.48 # 1
Precision 99.89 # 6
Specificity 99.45 # 1
Image Classification OmniBenchmark EfficientNetB4 Average Top-1 Accuracy 35.8 # 13
Fine-Grained Image Classification Oxford-IIIT Pet Dataset EfficientNet-B7 Accuracy 95.4% # 5
Fine-Grained Image Classification Stanford Cars EfficientNet-B7 Accuracy 94.7% # 30
Domain Generalization VizWiz-Classification EfficientNet-B5 Accuracy - All Images 42.8 # 21
Accuracy - Corrupted Images 37 # 20
Accuracy - Clean Images 47.3 # 21
Domain Generalization VizWiz-Classification EfficientNet-B1 Accuracy - All Images 36.7 # 57
Accuracy - Corrupted Images 30.9 # 53
Accuracy - Clean Images 41.5 # 56
Domain Generalization VizWiz-Classification EfficientNet-B0 Accuracy - All Images 34.2 # 77
Accuracy - Corrupted Images 27.4 # 77
Accuracy - Clean Images 38.4 # 78
Domain Generalization VizWiz-Classification EfficientNet-B2 Accuracy - All Images 38.1 # 51
Accuracy - Corrupted Images 31.4 # 49
Accuracy - Clean Images 42.8 # 45
Domain Generalization VizWiz-Classification EfficientNet-B4 Accuracy - All Images 41.7 # 26
Accuracy - Corrupted Images 35.6 # 25
Accuracy - Clean Images 46.4 # 24
Domain Generalization VizWiz-Classification EfficientNet-B3 Accuracy - All Images 40.7 # 33
Accuracy - Corrupted Images 34.2 # 34
Accuracy - Clean Images 45.3 # 30

Methods