EfficientNetV2: Smaller Models and Faster Training

1 Apr 2021 Mingxing Tan Quoc V. Le

This paper introduces EfficientNetV2, a new family of convolutional networks that have faster training speed and better parameter efficiency than previous models. To develop this family of models, we use a combination of training-aware neural architecture search and scaling, to jointly optimize training speed and parameter efficiency... (read more)

PDF Abstract
TASK DATASET MODEL METRIC NAME METRIC VALUE GLOBAL RANK USES EXTRA
TRAINING DATA
RESULT BENCHMARK
Image Classification CIFAR-10 EfficientNetV2-L Percentage correct 99.1 # 7
PARAMS 121M # 149
Image Classification CIFAR-10 EfficientNetV2-M Percentage correct 99.0 # 12
PARAMS 55M # 145
Image Classification CIFAR-10 EfficientNetV2-S Percentage correct 98.7 # 19
PARAMS 24M # 144
Image Classification CIFAR-100 EfficientNetV2-M Percentage correct 92.2 # 10
Image Classification CIFAR-100 EfficientNetV2-L Percentage correct 92.3 # 9
Image Classification CIFAR-100 EfficientNetV2-S Percentage correct 91.5 # 15
Image Classification Flowers-102 EfficientNetV2-M Accuracy 98.5 # 18
Image Classification Flowers-102 EfficientNetV2-S Accuracy 97.9 # 23
Image Classification Flowers-102 EfficientNetV2-L Accuracy 98.8 # 14
Image Classification ImageNet EfficientNetV2-L Top 1 Accuracy 85.7% # 36
Number of params 121M # 36
Image Classification ImageNet EfficientNetV2-S Top 1 Accuracy 83.9% # 87
Number of params 24M # 138
Image Classification ImageNet EfficientNetV2-M Top 1 Accuracy 85.1% # 51
Number of params 55M # 85
Image Classification ImageNet EfficientNetV2-L (21k) Top 1 Accuracy 86.8% # 16
Number of params 121M # 36
Image Classification ImageNet EfficientNetV2-M (21k) Top 1 Accuracy 86.1% # 26
Number of params 55M # 85
Image Classification ImageNet EfficientNetV2-S (21k) Top 1 Accuracy 85.0% # 56
Number of params 24M # 138
Image Classification Stanford Cars EfficientNetV2-S Accuracy 93.8 # 7
Image Classification Stanford Cars EfficientNetV2-L Accuracy 95.1 # 2
Image Classification Stanford Cars EfficientNetV2-M Accuracy 94.6 # 3

Methods used in the Paper


METHOD TYPE
Dropout
Regularization