# EfficientNetV2: Smaller Models and Faster Training

**Authors:** Mingxing Tan, Quoc V. Le  
**arXiv:** [2104.00298](https://arxiv.org/abs/2104.00298)  
**Date:** 2021-04-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2104.00298v3/

## Key points

- EfficientNetV2 is a new family of convolutional networks
- Aims to improve training speed and parameter efficiency compared to previous models
- Combines training-aware neural architecture search and scaling techniques
- Models were searched from a search space enriched with new operations such as Fused-MBConv
- EfficientNetV2 models can train significantly faster than state-of-the-art models
- Models are up to 6.8 times smaller in size
- Adaptive adjustment of regularization techniques like dropout and data augmentation addresses accuracy drop when increasing image size during training
- Outperforms previous models on various datasets including ImageNet, CIFAR, Cars, and Flowers
- Achieves an impressive top-1 accuracy of 87.3% on ImageNet ILSVRC2012 by pretraining on the ImageNet21k dataset
- Surpasses the recent ViT model by 2.0% accuracy while utilizing computing resources that are 5 to 11 times faster
- Code for EfficientNetV2 is available at https://github.com/google/automl/tree/master/efficientnetv2

## Summary

EfficientNetV2 is a new family of convolutional networks that aims to improve training speed and parameter efficiency compared to previous models. It combines training-aware neural architecture search and scaling techniques to optimize both training speed and parameter efficiency. The models in this family were searched from a search space enriched with new operations such as Fused-MBConv. Experiments show that EfficientNetV2 models can train significantly faster than state-of-the-art models while also being up to 6.8 times smaller in size. To address the issue of accuracy drop when increasing image size during training, the authors propose adaptively adjusting regularization techniques like dropout and data augmentation. Through progressive learning, EfficientNetV2 outperforms previous models on various datasets including ImageNet, CIFAR, Cars, and Flowers. By pretraining on the ImageNet21k dataset, it achieves an impressive top-1 accuracy of 87.3% on ImageNet ILSVRC2012 which surpasses the recent ViT model by 2.0% accuracy while utilizing computing resources that are 5 to 11 times faster. The code for EfficientNetV2 is available at https://github.com/google/automl/tree/master/efficientnetv2. In conclusion, EfficientNetV2 presents a significant advancement in convolutional network architecture by offering faster training speed and better parameter efficiency without compromising on accuracy.