efficientnet

EfficientNet

EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks

Introduction

EfficientNets are a family of image classification models, which achieve state-of-the-art accuracy, yet being an order-of-magnitude smaller and faster than previous models.

EfficientNets are based on AutoML and Compound Scaling. In particular, we first use AutoML MNAS Mobile framework to develop a mobile-size baseline network, named as EfficientNet-B0; Then, we use the compound scaling method to scale up this baseline to obtain EfficientNet-B1 to B7.

Abstract

Click to show the detailed Abstract

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.

How to use it?

Predict image

from mmpretrain import inference_model

predict = inference_model('efficientnet-b0_3rdparty_8xb32_in1k', 'demo/bird.JPEG')
print(predict['pred_class'])
print(predict['pred_score'])

Use the model

import torch
from mmpretrain import get_model

model = get_model('efficientnet-b0_3rdparty_8xb32_in1k', pretrained=True)
inputs = torch.rand(1, 3, 224, 224)
out = model(inputs)
print(type(out))
# To extract features.
feats = model.extract_feat(inputs)
print(type(feats))

Test Command

Prepare your dataset according to the docs.

Test:

python tools/test.py configs/efficientnet/efficientnet-b0_8xb32_in1k.py https://download.openmmlab.com/mmclassification/v0/efficientnet/efficientnet-b0_3rdparty_8xb32_in1k_20220119-a7e2a0b1.pth

Models and results

Image Classification on ImageNet-1k

Model	Pretrain	Params (M)	Flops (G)	Top-1 (%)	Top-5 (%)	Config	Download
efficientnet-b0_3rdparty_8xb32_in1k*	From scratch	5.29	0.42	76.74	93.17	config	model
efficientnet-b0_3rdparty_8xb32-aa_in1k*	From scratch	5.29	0.42	77.26	93.41	config	model
efficientnet-b0_3rdparty_8xb32-aa-advprop_in1k*	From scratch	5.29	0.42	77.53	93.61	config	model
efficientnet-b0_3rdparty-ra-noisystudent_in1k*	From scratch	5.29	0.42	77.63	94.00	config	model
efficientnet-b1_3rdparty_8xb32_in1k*	From scratch	7.79	0.74	78.68	94.28	config	model
efficientnet-b1_3rdparty_8xb32-aa_in1k*	From scratch	7.79	0.74	79.20	94.42	config	model
efficientnet-b1_3rdparty_8xb32-aa-advprop_in1k*	From scratch	7.79	0.74	79.52	94.43	config	model
efficientnet-b1_3rdparty-ra-noisystudent_in1k*	From scratch	7.79	0.74	81.44	95.83	config	model
efficientnet-b2_3rdparty_8xb32_in1k*	From scratch	9.11	1.07	79.64	94.80	config	model
efficientnet-b2_3rdparty_8xb32-aa_in1k*	From scratch	9.11	1.07	80.21	94.96	config	model
efficientnet-b2_3rdparty_8xb32-aa-advprop_in1k*	From scratch	9.11	1.07	80.45	95.07	config	model
efficientnet-b2_3rdparty-ra-noisystudent_in1k*	From scratch	9.11	1.07	82.47	96.23	config	model
efficientnet-b3_3rdparty_8xb32_in1k*	From scratch	12.23	1.95	81.01	95.34	config	model
efficientnet-b3_3rdparty_8xb32-aa_in1k*	From scratch	12.23	1.95	81.58	95.67	config	model
efficientnet-b3_3rdparty_8xb32-aa-advprop_in1k*	From scratch	12.23	1.95	81.81	95.69	config	model
efficientnet-b3_3rdparty-ra-noisystudent_in1k*	From scratch	12.23	1.95	84.02	96.89	config	model
efficientnet-b4_3rdparty_8xb32_in1k*	From scratch	19.34	4.66	82.57	96.09	config	model
efficientnet-b4_3rdparty_8xb32-aa_in1k*	From scratch	19.34	4.66	82.95	96.26	config	model
efficientnet-b4_3rdparty_8xb32-aa-advprop_in1k*	From scratch	19.34	4.66	83.25	96.44	config	model
efficientnet-b4_3rdparty-ra-noisystudent_in1k*	From scratch	19.34	4.66	85.25	97.52	config	model
efficientnet-b5_3rdparty_8xb32_in1k*	From scratch	30.39	10.80	83.18	96.47	config	model
efficientnet-b5_3rdparty_8xb32-aa_in1k*	From scratch	30.39	10.80	83.82	96.76	config	model
efficientnet-b5_3rdparty_8xb32-aa-advprop_in1k*	From scratch	30.39	10.80	84.21	96.98	config	model
efficientnet-b5_3rdparty-ra-noisystudent_in1k*	From scratch	30.39	10.80	86.08	97.75	config	model
efficientnet-b6_3rdparty_8xb32-aa_in1k*	From scratch	43.04	19.97	84.05	96.82	config	model
efficientnet-b6_3rdparty_8xb32-aa-advprop_in1k*	From scratch	43.04	19.97	84.74	97.14	config	model
efficientnet-b6_3rdparty-ra-noisystudent_in1k*	From scratch	43.04	19.97	86.47	97.87	config	model
efficientnet-b7_3rdparty_8xb32-aa_in1k*	From scratch	66.35	39.32	84.38	96.88	config	model
efficientnet-b7_3rdparty_8xb32-aa-advprop_in1k*	From scratch	66.35	39.32	85.14	97.23	config	model
efficientnet-b7_3rdparty-ra-noisystudent_in1k*	From scratch	66.35	39.32	86.83	98.08	config	model
efficientnet-b8_3rdparty_8xb32-aa-advprop_in1k*	From scratch	87.41	65.00	85.38	97.28	config	model
efficientnet-l2_3rdparty-ra-noisystudent_in1k-800px*	From scratch	480.31	174.20	88.33	98.65	config	model
efficientnet-l2_3rdparty-ra-noisystudent_in1k-475px*	From scratch	480.31	484.98	88.18	98.55	config	model

Models with * are converted from the official repo. The config files of these models are only for inference. We haven't reproduce the training results.

Citation

@inproceedings{tan2019efficientnet,
  title={Efficientnet: Rethinking model scaling for convolutional neural networks},
  author={Tan, Mingxing and Le, Quoc},
  booktitle={International Conference on Machine Learning},
  pages={6105--6114},
  year={2019},
  organization={PMLR}
}