EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks
Mingxing Tan 1 Quoc V. Le 1
Abstract EfficientNet-B7
84 B6 AmoebaNet-C
Convolutional Neural Networks (ConvNets) are B5 AmoebaNet-A
B4
commonly developed at a fixed resource budget, NASNet-A SENet
arXiv:1905.11946v5 [cs.LG] 11 Sep 2020
82
Imagenet Top-1 Accuracy (%)
and then scaled up for better accuracy if more B3
resources are available. In this paper, we sys- ResNeXt-101
80 Inception-ResNet-v2
tematically study model scaling and identify that
carefully balancing network depth, width, and res- Xception
olution can lead to better performance. Based 78 ResNet-152
Top1 Acc. #Params
on this observation, we propose a new scaling B0 DenseNet-201 ResNet-152 (He et al., 2016) 77.8% 60M
EfficientNet-B1 79.1% 7.8M
ResNeXt-101 (Xie et al., 2017) 80.9% 84M
method that uniformly scales all dimensions of 76 EfficientNet-B3 81.6% 12M
ResNet-50 SENet (Hu et al., 2018) 82.7% 146M
depth/width/resolution using a simple yet highly NASNet-A (Zoph et al., 2018) 82.7% 89M
EfficientNet-B4 82.9% 19M
effective compound coefficient. We demonstrate Inception-v2
GPipe (Huang et al., 2018) † 84.3% 556M
74
NASNet-A EfficientNet-B7 84.3% 66M
the effectiveness of this method on scaling up †
Not plotted
ResNet-34
MobileNets and ResNet. 0 20 40 60 80 100 120 140 160 180
Number of Parameters (Millions)
To go even further, we use neural architec-
Figure 1. Model Size vs. ImageNet Accuracy. All numbers are
ture search to design a new baseline network
for single-crop, single-model. Our EfficientNets significantly out-
and scale it up to obtain a family of models,
perform other ConvNets. In particular, EfficientNet-B7 achieves
called EfficientNets, which achieve much new state-of-the-art 84.3% top-1 accuracy but being 8.4x smaller
better accuracy and efficiency than previous and 6.1x faster than GPipe. EfficientNet-B1 is 7.6x smaller and
ConvNets. In particular, our EfficientNet-B7 5.7x faster than ResNet-152. Details are in Table 2 and 4.
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 time larger. However, the process of scaling up ConvNets
ConvNet. Our EfficientNets also transfer well and has never been well understood and there are currently many
achieve state-of-the-art accuracy on CIFAR-100 ways to do it. The most common way is to scale up Con-
(91.7%), Flowers (98.8%), and 3 other transfer vNets by their depth (He et al., 2016) or width (Zagoruyko &
learning datasets, with an order of magnitude Komodakis, 2016). Another less common, but increasingly
fewer parameters. Source code is at https: popular, method is to scale up models by image resolution
//github.com/tensorflow/tpu/tree/ (Huang et al., 2018). In previous work, it is common to scale
master/models/official/efficientnet. only one of the three dimensions – depth, width, and image
size. Though it is possible to scale two or three dimensions
arbitrarily, arbitrary scaling requires tedious manual tuning
and still often yields sub-optimal accuracy and efficiency.
1. Introduction
In this paper, we want to study and rethink the process
Scaling up ConvNets is widely used to achieve better accu- of scaling up ConvNets. In particular, we investigate the
racy. For example, ResNet (He et al., 2016) can be scaled central question: is there a principled method to scale up
up from ResNet-18 to ResNet-200 by using more layers; ConvNets that can achieve better accuracy and efficiency?
Recently, GPipe (Huang et al., 2018) achieved 84.3% Ima- Our empirical study shows that it is critical to balance all
geNet top-1 accuracy by scaling up a baseline model four dimensions of network width/depth/resolution, and surpris-
1
ingly such balance can be achieved by simply scaling each
Google Research, Brain Team, Mountain View, CA. Corre- of them with constant ratio. Based on this observation, we
spondence to: Mingxing Tan <tanmingxing@google.com>.
propose a simple yet effective compound scaling method.
Proceedings of the 36 th International Conference on Machine Unlike conventional practice that arbitrary scales these fac-
Learning, Long Beach, California, PMLR 97, 2019. tors, our method uniformly scales network width, depth,
EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks
wider
#channels
wider
deeper
deeper
layer_i
higher higher
resolution HxW resolution resolution
(a) baseline (b) width scaling (c) depth scaling (d) resolution scaling (e) compound scaling
Figure 2. Model Scaling. (a) is a baseline network example; (b)-(d) are conventional scaling that only increases one dimension of network
width, depth, or resolution. (e) is our proposed compound scaling method that uniformly scales all three dimensions with a fixed ratio.
and resolution with a set of fixed scaling coefficients. For of-the-art accuracy on 5 out of 8 widely used datasets, while
example, if we want to use 2N times more computational reducing parameters by up to 21x than existing ConvNets.
resources, then we can simply increase the network depth by
αN , width by β N , and image size by γ N , where α, β, γ are 2. Related Work
constant coefficients determined by a small grid search on
the original small model. Figure 2 illustrates the difference ConvNet Accuracy: Since AlexNet (Krizhevsky et al.,
between our scaling method and conventional methods. 2012) won the 2012 ImageNet competition, ConvNets have
become increasingly more accurate by going bigger: while
Intuitively, the compound scaling method makes sense be-
the 2014 ImageNet winner GoogleNet (Szegedy et al., 2015)
cause if the input image is bigger, then the network needs
achieves 74.8% top-1 accuracy with about 6.8M parameters,
more layers to increase the receptive field and more channels
the 2017 ImageNet winner SENet (Hu et al., 2018) achieves
to capture more fine-grained patterns on the bigger image. In
82.7% top-1 accuracy with 145M parameters. Recently,
fact, previous theoretical (Raghu et al., 2017; Lu et al., 2018)
GPipe (Huang et al., 2018) further pushes the state-of-the-art
and empirical results (Zagoruyko & Komodakis, 2016) both
ImageNet top-1 validation accuracy to 84.3% using 557M
show that there exists certain relationship between network
parameters: it is so big that it can only be trained with a
width and depth, but to our best knowledge, we are the
specialized pipeline parallelism library by partitioning the
first to empirically quantify the relationship among all three
network and spreading each part to a different accelera-
dimensions of network width, depth, and resolution.
tor. While these models are mainly designed for ImageNet,
We demonstrate that our scaling method work well on exist- recent studies have shown better ImageNet models also per-
ing MobileNets (Howard et al., 2017; Sandler et al., 2018) form better across a variety of transfer learning datasets
and ResNet (He et al., 2016). Notably, the effectiveness of (Kornblith et al., 2019), and other computer vision tasks
model scaling heavily depends on the baseline network; to such as object detection (He et al., 2016; Tan et al., 2019).
go even further, we use neural architecture search (Zoph Although higher accuracy is critical for many applications,
& Le, 2017; Tan et al., 2019) to develop a new baseline we have already hit the hardware memory limit, and thus
network, and scale it up to obtain a family of models, called further accuracy gain needs better efficiency.
EfficientNets. Figure 1 summarizes the ImageNet perfor-
mance, where our EfficientNets significantly outperform ConvNet Efficiency: Deep ConvNets are often over-
other ConvNets. In particular, our EfficientNet-B7 surpasses parameterized. Model compression (Han et al., 2016; He
the best existing GPipe accuracy (Huang et al., 2018), but et al., 2018; Yang et al., 2018) is a common way to re-
using 8.4x fewer parameters and running 6.1x faster on in- duce model size by trading accuracy for efficiency. As mo-
ference. Compared to the widely used ResNet-50 (He et al., bile phones become ubiquitous, it is also common to hand-
2016), our EfficientNet-B4 improves the top-1 accuracy craft efficient mobile-size ConvNets, such as SqueezeNets
from 76.3% to 83.0% (+6.7%) with similar FLOPS. Besides (Iandola et al., 2016; Gholami et al., 2018), MobileNets
ImageNet, EfficientNets also transfer well and achieve state- (Howard et al., 2017; Sandler et al., 2018), and ShuffleNets
EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks
(Zhang et al., 2018; Ma et al., 2018). Recently, neural archi- i. Figure 2(a) illustrate a representative ConvNet, where
tecture search becomes increasingly popular in designing the spatial dimension is gradually shrunk but the channel
efficient mobile-size ConvNets (Tan et al., 2019; Cai et al., dimension is expanded over layers, for example, from initial
2019), and achieves even better efficiency than hand-crafted input shape h224, 224, 3i to final output shape h7, 7, 512i.
mobile ConvNets by extensively tuning the network width,
Unlike regular ConvNet designs that mostly focus on find-
depth, convolution kernel types and sizes. However, it is
ing the best layer architecture Fi , model scaling tries to ex-
unclear how to apply these techniques for larger models that
pand the network length (Li ), width (Ci ), and/or resolution
have much larger design space and much more expensive
(Hi , Wi ) without changing Fi predefined in the baseline
tuning cost. In this paper, we aim to study model efficiency
network. By fixing Fi , model scaling simplifies the design
for super large ConvNets that surpass state-of-the-art accu-
problem for new resource constraints, but it still remains
racy. To achieve this goal, we resort to model scaling.
a large design space to explore different Li , Ci , Hi , Wi for
Model Scaling: There are many ways to scale a Con- each layer. In order to further reduce the design space, we
vNet for different resource constraints: ResNet (He et al., restrict that all layers must be scaled uniformly with con-
2016) can be scaled down (e.g., ResNet-18) or up (e.g., stant ratio. Our target is to maximize the model accuracy
ResNet-200) by adjusting network depth (#layers), while for any given resource constraints, which can be formulated
WideResNet (Zagoruyko & Komodakis, 2016) and Mo- as an optimization problem:
bileNets (Howard et al., 2017) can be scaled by network
width (#channels). It is also well-recognized that bigger max Accuracy N (d, w, r)
d,w,r
input image size will help accuracy with the overhead of K
F̂id·L̂i Xhr·Ĥi ,r·Ŵi ,w·Ĉi i
more FLOPS. Although prior studies (Raghu et al., 2017; s.t. N (d, w, r) =
i=1...s
Lin & Jegelka, 2018; Sharir & Shashua, 2018; Lu et al.,
2018) have shown that network depth and width are both Memory(N ) ≤ target memory
important for ConvNets’ expressive power, it still remains FLOPS(N ) ≤ target flops
an open question of how to effectively scale a ConvNet to (2)
achieve better efficiency and accuracy. Our work systemati- where w, d, r are coefficients for scaling network width,
cally and empirically studies ConvNet scaling for all three depth, and resolution; F̂i , L̂i , Ĥi , Ŵi , Ĉi are predefined pa-
dimensions of network width, depth, and resolutions. rameters in baseline network (see Table 1 as an example).
3. Compound Model Scaling 3.2. Scaling Dimensions
In this section, we will formulate the scaling problem, study The main difficulty of problem 2 is that the optimal d, w, r
different approaches, and propose our new scaling method. depend on each other and the values change under different
resource constraints. Due to this difficulty, conventional
3.1. Problem Formulation methods mostly scale ConvNets in one of these dimensions:
A ConvNet Layer i can be defined as a function: Yi =
Depth (d ): Scaling network depth is the most common way
Fi (Xi ), where Fi is the operator, Yi is output tensor, Xi is
used by many ConvNets (He et al., 2016; Huang et al., 2017;
input tensor, with tensor shape hHi , Wi , Ci i1 , where Hi and
Szegedy et al., 2015; 2016). The intuition is that deeper
Wi are spatial dimension and Ci is the channel dimension.
ConvNet can capture richer and more complex features, and
A ConvNet N can be represented by aJ list of composed lay-
generalize well on new tasks. However, deeper networks
ers: N = Fk ... F2 F1 (X1 ) = j=1...k Fj (X1 ). In
are also more difficult to train due to the vanishing gradient
practice, ConvNet layers are often partitioned into multiple
problem (Zagoruyko & Komodakis, 2016). Although sev-
stages and all layers in each stage share the same architec-
eral techniques, such as skip connections (He et al., 2016)
ture: for example, ResNet (He et al., 2016) has five stages,
and batch normalization (Ioffe & Szegedy, 2015), alleviate
and all layers in each stage has the same convolutional type
the training problem, the accuracy gain of very deep network
except the first layer performs down-sampling. Therefore,
diminishes: for example, ResNet-1000 has similar accuracy
we can define a ConvNet as:
as ResNet-101 even though it has much more layers. Figure
3 (middle) shows our empirical study on scaling a baseline
K
FiLi XhHi ,Wi ,Ci i
N = (1)
i=1...s
model with different depth coefficient d, further suggesting
the diminishing accuracy return for very deep ConvNets.
where FiLi denotes layer Fi is repeated Li times in stage i,
hHi , Wi , Ci i denotes the shape of input tensor X of layer
Width (w ): Scaling network width is commonly used for
1
For the sake of simplicity, we omit batch dimension. small size models (Howard et al., 2017; Sandler et al., 2018;
EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks
81 81 81
ImageNet Top-1 Accuracy(%)
80 w=5.0 80 80
w=3.8 d=6.0 d=8.0
w=2.6 r=2.2 r=2.5
79 79 79 r=1.9
d=3.0d=4.0 r=1.7
w=1.8 d=2.0
78 78 78 r=1.5
r=1.3
w=1.4
77 77 77
76 w=1.0 76 d=1.0 76 r=1.0
75 75 75
0 2 4 6 8 0 1 2 3 4 0 1 2 3
FLOPS (Billions) FLOPS (Billions) FLOPS (Billions)
Figure 3. Scaling Up a Baseline Model with Different Network Width (w), Depth (d), and Resolution (r) Coefficients. Bigger
networks with larger width, depth, or resolution tend to achieve higher accuracy, but the accuracy gain quickly saturate after reaching
80%, demonstrating the limitation of single dimension scaling. Baseline network is described in Table 1.
Tan et al., 2019)2 . As discussed in (Zagoruyko & Ko-
82
modakis, 2016), wider networks tend to be able to capture
more fine-grained features and are easier to train. However, 81
ImageNet Top1 Accuracy (%)
extremely wide but shallow networks tend to have difficul-
ties in capturing higher level features. Our empirical results 80
in Figure 3 (left) show that the accuracy quickly saturates
when networks become much wider with larger w. 79
Resolution (r ): With higher resolution input images, Con- 78
d=1.0, r=1.0
vNets can potentially capture more fine-grained patterns.
d=1.0, r=1.3
Starting from 224x224 in early ConvNets, modern Con- 77
d=2.0, r=1.0
vNets tend to use 299x299 (Szegedy et al., 2016) or 331x331 d=2.0, r=1.3
76
(Zoph et al., 2018) for better accuracy. Recently, GPipe
0 5 10 15 20 25
(Huang et al., 2018) achieves state-of-the-art ImageNet ac- FLOPS (billions)
curacy with 480x480 resolution. Higher resolutions, such as
600x600, are also widely used in object detection ConvNets Figure 4. Scaling Network Width for Different Baseline Net-
(He et al., 2017; Lin et al., 2017). Figure 3 (right) shows the works. Each dot in a line denotes a model with different width
results of scaling network resolutions, where indeed higher coefficient (w). All baseline networks are from Table 1. The first
baseline network (d=1.0, r=1.0) has 18 convolutional layers with
resolutions improve accuracy, but the accuracy gain dimin-
resolution 224x224, while the last baseline (d=2.0, r=1.3) has 36
ishes for very high resolutions (r = 1.0 denotes resolution
layers with resolution 299x299.
224x224 and r = 2.5 denotes resolution 560x560).
The above analyses lead us to the first observation:
order to capture more fine-grained patterns with more pixels
Observation 1 – Scaling up any dimension of network in high resolution images. These intuitions suggest that we
width, depth, or resolution improves accuracy, but the accu- need to coordinate and balance different scaling dimensions
racy gain diminishes for bigger models. rather than conventional single-dimension scaling.
To validate our intuitions, we compare width scaling under
3.3. Compound Scaling different network depths and resolutions, as shown in Figure
4. If we only scale network width w without changing
We empirically observe that different scaling dimensions are
depth (d=1.0) and resolution (r=1.0), the accuracy saturates
not independent. Intuitively, for higher resolution images,
quickly. With deeper (d=2.0) and higher resolution (r=2.0),
we should increase network depth, such that the larger re-
width scaling achieves much better accuracy under the same
ceptive fields can help capture similar features that include
FLOPS cost. These results lead us to the second observation:
more pixels in bigger images. Correspondingly, we should
also increase network width when resolution is higher, in
Observation 2 – In order to pursue better accuracy and
2
In some literature, scaling number of channels is called “depth efficiency, it is critical to balance all dimensions of network
multiplier”, which means the same as our width coefficient w. width, depth, and resolution during ConvNet scaling.
EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks
In fact, a few prior work (Zoph et al., 2018; Real et al., 2019) Table 1. EfficientNet-B0 baseline network – Each row describes
have already tried to arbitrarily balance network width and a stage i with L̂i layers, with input resolution hĤi , Ŵi i and output
depth, but they all require tedious manual tuning. channels Ĉi . Notations are adopted from equation 2.
Stage Operator Resolution #Channels #Layers
In this paper, we propose a new compound scaling method,
i F̂i Ĥi × Ŵi Ĉi L̂i
which use a compound coefficient φ to uniformly scales
1 Conv3x3 224 × 224 32 1
network width, depth, and resolution in a principled way: 2 MBConv1, k3x3 112 × 112 16 1
3 MBConv6, k3x3 112 × 112 24 2
depth: d = αφ 4 MBConv6, k5x5 56 × 56 40 2
5 MBConv6, k3x3 28 × 28 80 3
width: w = β φ 6 MBConv6, k5x5 14 × 14 112 3
7 MBConv6, k5x5 14 × 14 192 4
resolution: r = γ φ (3) 8 MBConv6, k3x3 7×7 320 1
2 2 9 Conv1x1 & Pooling & FC 7×7 1280 1
s.t. α · β · γ ≈ 2
α ≥ 1, β ≥ 1, γ ≥ 1
Net, except our EfficientNet-B0 is slightly bigger due to
where α, β, γ are constants that can be determined by a the larger FLOPS target (our FLOPS target is 400M). Ta-
small grid search. Intuitively, φ is a user-specified coeffi- ble 1 shows the architecture of EfficientNet-B0. Its main
cient that controls how many more resources are available building block is mobile inverted bottleneck MBConv (San-
for model scaling, while α, β, γ specify how to assign these dler et al., 2018; Tan et al., 2019), to which we also add
extra resources to network width, depth, and resolution re- squeeze-and-excitation optimization (Hu et al., 2018).
spectively. Notably, the FLOPS of a regular convolution op Starting from the baseline EfficientNet-B0, we apply our
is proportional to d, w2 , r2 , i.e., doubling network depth compound scaling method to scale it up with two steps:
will double FLOPS, but doubling network width or resolu-
tion will increase FLOPS by four times. Since convolution • STEP 1: we first fix φ = 1, assuming twice more re-
ops usually dominate the computation cost in ConvNets, sources available, and do a small grid search of α, β, γ
scaling a ConvNet with equation 3 will approximately in- based on Equation 2 and 3. In particular, we find
φ
crease total FLOPS by α · β 2 · γ 2 . In this paper, we the best values for EfficientNet-B0 are α = 1.2, β =
constraint α · β 2 · γ 2 ≈ 2 such that for any new φ, the total 1.1, γ = 1.15, under constraint of α · β 2 · γ 2 ≈ 2.
FLOPS will approximately3 increase by 2φ . • STEP 2: we then fix α, β, γ as constants and scale up
baseline network with different φ using Equation 3, to
4. EfficientNet Architecture obtain EfficientNet-B1 to B7 (Details in Table 2).
Since model scaling does not change layer operators F̂i Notably, it is possible to achieve even better performance by
in baseline network, having a good baseline network is searching for α, β, γ directly around a large model, but the
also critical. We will evaluate our scaling method using search cost becomes prohibitively more expensive on larger
existing ConvNets, but in order to better demonstrate the models. Our method solves this issue by only doing search
effectiveness of our scaling method, we have also developed once on the small baseline network (step 1), and then use
a new mobile-size baseline, called EfficientNet. the same scaling coefficients for all other models (step 2).
Inspired by (Tan et al., 2019), we develop our baseline net-
work by leveraging a multi-objective neural architecture 5. Experiments
search that optimizes both accuracy and FLOPS. Specifi-
cally, we use the same search space as (Tan et al., 2019), In this section, we will first evaluate our scaling method on
and use ACC(m)×[F LOP S(m)/T ]w as the optimization existing ConvNets and the new proposed EfficientNets.
goal, where ACC(m) and F LOP S(m) denote the accu-
racy and FLOPS of model m, T is the target FLOPS and 5.1. Scaling Up MobileNets and ResNets
w=-0.07 is a hyperparameter for controlling the trade-off
As a proof of concept, we first apply our scaling method
between accuracy and FLOPS. Unlike (Tan et al., 2019;
to the widely-used MobileNets (Howard et al., 2017; San-
Cai et al., 2019), here we optimize FLOPS rather than la-
dler et al., 2018) and ResNet (He et al., 2016). Table 3
tency since we are not targeting any specific hardware de-
shows the ImageNet results of scaling them in different
vice. Our search produces an efficient network, which we
ways. Compared to other single-dimension scaling methods,
name EfficientNet-B0. Since we use the same search space
our compound scaling method improves the accuracy on all
as (Tan et al., 2019), the architecture is similar to Mnas-
these models, suggesting the effectiveness of our proposed
3
FLOPS may differ from theoretical value due to rounding. scaling method for general existing ConvNets.
EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks
Table 2. EfficientNet Performance Results on ImageNet (Russakovsky et al., 2015). All EfficientNet models are scaled from our
baseline EfficientNet-B0 using different compound coefficient φ in Equation 3. ConvNets with similar top-1/top-5 accuracy are grouped
together for efficiency comparison. Our scaled EfficientNet models consistently reduce parameters and FLOPS by an order of magnitude
(up to 8.4x parameter reduction and up to 16x FLOPS reduction) than existing ConvNets.
Model Top-1 Acc. Top-5 Acc. #Params Ratio-to-EfficientNet #FLOPs Ratio-to-EfficientNet
EfficientNet-B0 77.1% 93.3% 5.3M 1x 0.39B 1x
ResNet-50 (He et al., 2016) 76.0% 93.0% 26M 4.9x 4.1B 11x
DenseNet-169 (Huang et al., 2017) 76.2% 93.2% 14M 2.6x 3.5B 8.9x
EfficientNet-B1 79.1% 94.4% 7.8M 1x 0.70B 1x
ResNet-152 (He et al., 2016) 77.8% 93.8% 60M 7.6x 11B 16x
DenseNet-264 (Huang et al., 2017) 77.9% 93.9% 34M 4.3x 6.0B 8.6x
Inception-v3 (Szegedy et al., 2016) 78.8% 94.4% 24M 3.0x 5.7B 8.1x
Xception (Chollet, 2017) 79.0% 94.5% 23M 3.0x 8.4B 12x
EfficientNet-B2 80.1% 94.9% 9.2M 1x 1.0B 1x
Inception-v4 (Szegedy et al., 2017) 80.0% 95.0% 48M 5.2x 13B 13x
Inception-resnet-v2 (Szegedy et al., 2017) 80.1% 95.1% 56M 6.1x 13B 13x
EfficientNet-B3 81.6% 95.7% 12M 1x 1.8B 1x
ResNeXt-101 (Xie et al., 2017) 80.9% 95.6% 84M 7.0x 32B 18x
PolyNet (Zhang et al., 2017) 81.3% 95.8% 92M 7.7x 35B 19x
EfficientNet-B4 82.9% 96.4% 19M 1x 4.2B 1x
SENet (Hu et al., 2018) 82.7% 96.2% 146M 7.7x 42B 10x
NASNet-A (Zoph et al., 2018) 82.7% 96.2% 89M 4.7x 24B 5.7x
AmoebaNet-A (Real et al., 2019) 82.8% 96.1% 87M 4.6x 23B 5.5x
PNASNet (Liu et al., 2018) 82.9% 96.2% 86M 4.5x 23B 6.0x
EfficientNet-B5 83.6% 96.7% 30M 1x 9.9B 1x
AmoebaNet-C (Cubuk et al., 2019) 83.5% 96.5% 155M 5.2x 41B 4.1x
EfficientNet-B6 84.0% 96.8% 43M 1x 19B 1x
EfficientNet-B7 84.3% 97.0% 66M 1x 37B 1x
GPipe (Huang et al., 2018) 84.3% 97.0% 557M 8.4x - -
We omit ensemble and multi-crop models (Hu et al., 2018), or models pretrained on 3.5B Instagram images (Mahajan et al., 2018).
Table 3. Scaling Up MobileNets and ResNet.
EfficientNet-B6
84 AmoebaNet-C
Model FLOPS Top-1 Acc. B5
AmeobaNet-A
Baseline MobileNetV1 (Howard et al., 2017) 0.6B 70.6% B4
NASNet-A SENet
82
Imagenet Top-1 Accuracy (%)
Scale MobileNetV1 by width (w=2) 2.2B 74.2% B3
Scale MobileNetV1 by resolution (r=2) 2.2B 72.7% ResNeXt-101
compound scale (d =1.4, w =1.2, r =1.3) 2.3B 75.6% 80 Inception-ResNet-v2
Baseline MobileNetV2 (Sandler et al., 2018) 0.3B 72.0% Xception
Scale MobileNetV2 by depth (d=4) 1.2B 76.8% 78 ResNet-152
Top1 Acc. FLOPS
Scale MobileNetV2 by width (w=2) 1.1B 76.4% DenseNet-201
B0 ResNet-152 (Xie et al., 2017) 77.8% 11B
Scale MobileNetV2 by resolution (r=2) 1.2B 74.8% EfficientNet-B1 79.1% 0.7B
76 ResNeXt-101 (Xie et al., 2017) 80.9% 32B
MobileNetV2 compound scale 1.3B 77.4% ResNet-50 EfficientNet-B3 81.6% 1.8B
SENet (Hu et al., 2018) 82.7% 42B
Baseline ResNet-50 (He et al., 2016) 4.1B 76.0% NASNet-A (Zoph et al., 2018) 80.7% 24B
Inception-v2 EfficientNet-B4 82.9% 4.2B
Scale ResNet-50 by depth (d=4) 16.2B 78.1% 74 AmeobaNet-C (Cubuk et al., 2019) 83.5% 41B
NASNet-A EfficientNet-B5 83.6% 9.9B
Scale ResNet-50 by width (w=2) 14.7B 77.7% ResNet-34
Scale ResNet-50 by resolution (r=2) 16.4B 77.5% 0 5 10 15 20 25 30 35 40 45
ResNet-50 compound scale 16.7B 78.8% FLOPS (Billions)
Figure 5. FLOPS vs. ImageNet Accuracy – Similar to Figure 1
except it compares FLOPS rather than model size.
Table 4. Inference Latency Comparison – Latency is measured
with batch size 1 on a single core of Intel Xeon CPU E5-2690.
5.2. ImageNet Results for EfficientNet
Acc. @ Latency Acc. @ Latency
ResNet-152 77.8% @ 0.554s GPipe 84.3% @ 19.0s
We train our EfficientNet models on ImageNet using simi-
EfficientNet-B1 78.8% @ 0.098s EfficientNet-B7 84.4% @ 3.1s lar settings as (Tan et al., 2019): RMSProp optimizer with
Speedup 5.7x Speedup 6.1x decay 0.9 and momentum 0.9; batch norm momentum 0.99;
EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks
Table 5. EfficientNet Performance Results on Transfer Learning Datasets. Our scaled EfficientNet models achieve new state-of-the-
art accuracy for 5 out of 8 datasets, with 9.6x fewer parameters on average.
Comparison to best public-available results Comparison to best reported results
Model Acc. #Param Our Model Acc. #Param(ratio) Model Acc. #Param Our Model Acc. #Param(ratio)
†
CIFAR-10 NASNet-A 98.0% 85M EfficientNet-B0 98.1% 4M (21x) Gpipe 99.0% 556M EfficientNet-B7 98.9% 64M (8.7x)
CIFAR-100 NASNet-A 87.5% 85M EfficientNet-B0 88.1% 4M (21x) Gpipe 91.3% 556M EfficientNet-B7 91.7% 64M (8.7x)
Birdsnap Inception-v4 81.8% 41M EfficientNet-B5 82.0% 28M (1.5x) GPipe 83.6% 556M EfficientNet-B7 84.3% 64M (8.7x)
‡
Stanford Cars Inception-v4 93.4% 41M EfficientNet-B3 93.6% 10M (4.1x) DAT 94.8% - EfficientNet-B7 94.7% -
Flowers Inception-v4 98.5% 41M EfficientNet-B5 98.5% 28M (1.5x) DAT 97.7% - EfficientNet-B7 98.8% -
FGVC Aircraft Inception-v4 90.9% 41M EfficientNet-B3 90.7% 10M (4.1x) DAT 92.9% - EfficientNet-B7 92.9% -
Oxford-IIIT Pets ResNet-152 94.5% 58M EfficientNet-B4 94.8% 17M (5.6x) GPipe 95.9% 556M EfficientNet-B6 95.4% 41M (14x)
Food-101 Inception-v4 90.8% 41M EfficientNet-B4 91.5% 17M (2.4x) GPipe 93.0% 556M EfficientNet-B7 93.0% 64M (8.7x)
Geo-Mean (4.7x) (9.6x)
†
GPipe (Huang et al., 2018) trains giant models with specialized pipeline parallelism library.
‡
DAT denotes domain adaptive transfer learning (Ngiam et al., 2018). Here we only compare ImageNet-based transfer learning results.
Transfer accuracy and #params for NASNet (Zoph et al., 2018), Inception-v4 (Szegedy et al., 2017), ResNet-152 (He et al., 2016) are from (Kornblith et al., 2019).
CIFAR10 CIFAR100 Birdsnap Stanford Cars
1.0
99 92 85
90 94
Accuracy(%)
80
98
0.8 88 93
86 75
97 92
84 91
0.6 70
96
101 102 103 101 102 103 101 102 103 101 102 103
Flowers FGVC Aircraft Oxford-IIIT Pets Food-101
96
0.4 92.5
98.5 92
Accuracy(%)
90.0
94 90
98.0
0.2 87.5
97.5 88
85.0 92
97.0 86
0.0 82.5
0.0 101 102 0.2103 101 2
100.4 103 101 0.6 102 103 0.8101 102 1.03
10
Number of Parameters (Millions, log-scale)
DenseNet-201 ResNet-50 Inception-v1 ResNet-152 NASNet-A
GPIPE ResNet-101 Inception-v3 DenseNet-121 EfficientNet
Inception-ResNet-v2 DenseNet-169 Inception-v4
Figure 6. Model Parameters vs. Transfer Learning Accuracy – All models are pretrained on ImageNet and finetuned on new datasets.
weight decay 1e-5; initial learning rate 0.256 that decays being more accurate but 8.4x smaller than the previous
by 0.97 every 2.4 epochs. We also use SiLU (Swish-1) ac- best GPipe (Huang et al., 2018). These gains come from
tivation (Ramachandran et al., 2018; Elfwing et al., 2018; both better architectures, better scaling, and better training
Hendrycks & Gimpel, 2016), AutoAugment (Cubuk et al., settings that are customized for EfficientNet.
2019), and stochastic depth (Huang et al., 2016) with sur-
Figure 1 and Figure 5 illustrates the parameters-accuracy
vival probability 0.8. As commonly known that bigger mod-
and FLOPS-accuracy curve for representative ConvNets,
els need more regularization, we linearly increase dropout
where our scaled EfficientNet models achieve better accu-
(Srivastava et al., 2014) ratio from 0.2 for EfficientNet-B0 to
racy with much fewer parameters and FLOPS than other
0.5 for B7. We reserve 25K randomly picked images from
ConvNets. Notably, our EfficientNet models are not only
the training set as a minival set, and perform early
small, but also computational cheaper. For example, our
stopping on this minival; we then evaluate the early-
EfficientNet-B3 achieves higher accuracy than ResNeXt-
stopped checkpoint on the original validation set to
101 (Xie et al., 2017) using 18x fewer FLOPS.
report the final validation accuracy.
To validate the latency, we have also measured the inference
Table 2 shows the performance of all EfficientNet models
latency for a few representative CovNets on a real CPU as
that are scaled from the same baseline EfficientNet-B0. Our
shown in Table 4, where we report average latency over
EfficientNet models generally use an order of magnitude
20 runs. Our EfficientNet-B1 runs 5.7x faster than the
fewer parameters and FLOPS than other ConvNets with
widely used ResNet-152, while EfficientNet-B7 runs about
similar accuracy. In particular, our EfficientNet-B7 achieves
6.1x faster than GPipe (Huang et al., 2018), suggesting our
84.3% top1 accuracy with 66M parameters and 37B FLOPS,
EfficientNets are indeed fast on real hardware.
EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks
original image baseline model deeper (d=4) wider (w=2) higher resolution (r=2) compound scaling
bakeshop
maze
Figure 7. Class Activation Map (CAM) (Zhou et al., 2016) for Models with different scaling methods- Our compound scaling method
allows the scaled model (last column) to focus on more relevant regions with more object details. Model details are in Table 7.
83
Table 6. Transfer Learning Datasets.
ImageNet Top-1 Accuracy(%)
82
Dataset Train Size Test Size #Classes 81
CIFAR-10 (Krizhevsky & Hinton, 2009) 50,000 10,000 10
CIFAR-100 (Krizhevsky & Hinton, 2009) 50,000 10,000 100 80
Birdsnap (Berg et al., 2014) 47,386 2,443 500 79
Stanford Cars (Krause et al., 2013) 8,144 8,041 196
Flowers (Nilsback & Zisserman, 2008) 2,040 6,149 102 78
FGVC Aircraft (Maji et al., 2013) 6,667 3,333 100
77 scale by width
Oxford-IIIT Pets (Parkhi et al., 2012) 3,680 3,369 37
scale by depth
Food-101 (Bossard et al., 2014) 75,750 25,250 101 scale by resolution
76
compound scaling
75
5.3. Transfer Learning Results for EfficientNet 0 1 2 3 4 5
FLOPS (Billions)
We have also evaluated our EfficientNet on a list of com- Figure 8. Scaling Up EfficientNet-B0 with Different Methods.
monly used transfer learning datasets, as shown in Table
6. We borrow the same training settings from (Kornblith
Table 7. Scaled Models Used in Figure 7.
et al., 2019) and (Huang et al., 2018), which take ImageNet
pretrained checkpoints and finetune on new datasets. Model FLOPS Top-1 Acc.
Table 5 shows the transfer learning performance: (1) Com- Baseline model (EfficientNet-B0) 0.4B 77.3%
pared to public available models, such as NASNet-A (Zoph Scale model by depth (d=4) 1.8B 79.0%
et al., 2018) and Inception-v4 (Szegedy et al., 2017), our Ef- Scale model by width (w=2) 1.8B 78.9%
Scale model by resolution (r=2) 1.9B 79.1%
ficientNet models achieve better accuracy with 4.7x average Compound Scale (d =1.4, w =1.2, r =1.3) 1.8B 81.1%
(up to 21x) parameter reduction. (2) Compared to state-
of-the-art models, including DAT (Ngiam et al., 2018) that
dynamically synthesizes training data and GPipe (Huang ods for the same EfficientNet-B0 baseline network. In gen-
et al., 2018) that is trained with specialized pipeline paral- eral, all scaling methods improve accuracy with the cost
lelism, our EfficientNet models still surpass their accuracy of more FLOPS, but our compound scaling method can
in 5 out of 8 datasets, but using 9.6x fewer parameters further improve accuracy, by up to 2.5%, than other single-
dimension scaling methods, suggesting the importance of
Figure 6 compares the accuracy-parameters curve for a va-
our proposed compound scaling.
riety of models. In general, our EfficientNets consistently
achieve better accuracy with an order of magnitude fewer pa- In order to further understand why our compound scaling
rameters than existing models, including ResNet (He et al., method is better than others, Figure 7 compares the class
2016), DenseNet (Huang et al., 2017), Inception (Szegedy activation map (Zhou et al., 2016) for a few representative
et al., 2017), and NASNet (Zoph et al., 2018). models with different scaling methods. All these models are
scaled from the same baseline, and their statistics are shown
6. Discussion in Table 7. Images are randomly picked from ImageNet
validation set. As shown in the figure, the model with com-
To disentangle the contribution of our proposed scaling pound scaling tends to focus on more relevant regions with
method from the EfficientNet architecture, Figure 8 com- more object details, while other models are either lack of
pares the ImageNet performance of different scaling meth- object details or unable to capture all objects in the images.
EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks
7. Conclusion Cubuk, E. D., Zoph, B., Mane, D., Vasudevan, V., and Le,
Q. V. Autoaugment: Learning augmentation policies
In this paper, we systematically study ConvNet scaling and from data. CVPR, 2019.
identify that carefully balancing network width, depth, and
resolution is an important but missing piece, preventing us Elfwing, S., Uchibe, E., and Doya, K. Sigmoid-weighted
from better accuracy and efficiency. To address this issue, linear units for neural network function approximation
we propose a simple and highly effective compound scaling in reinforcement learning. Neural Networks, 107:3–11,
method, which enables us to easily scale up a baseline Con- 2018.
vNet to any target resource constraints in a more principled
way, while maintaining model efficiency. Powered by this Gholami, A., Kwon, K., Wu, B., Tai, Z., Yue, X., Jin, P.,
compound scaling method, we demonstrate that a mobile- Zhao, S., and Keutzer, K. Squeezenext: Hardware-aware
size EfficientNet model can be scaled up very effectively, neural network design. ECV Workshop at CVPR’18,
surpassing state-of-the-art accuracy with an order of magni- 2018.
tude fewer parameters and FLOPS, on both ImageNet and
five commonly used transfer learning datasets. Han, S., Mao, H., and Dally, W. J. Deep compression:
Compressing deep neural networks with pruning, trained
quantization and huffman coding. ICLR, 2016.
Acknowledgements
He, K., Zhang, X., Ren, S., and Sun, J. Deep residual
We thank Ruoming Pang, Vijay Vasudevan, Alok Aggarwal,
learning for image recognition. CVPR, pp. 770–778,
Barret Zoph, Hongkun Yu, Jonathon Shlens, Raphael Gon-
2016.
tijo Lopes, Yifeng Lu, Daiyi Peng, Xiaodan Song, Samy
Bengio, Jeff Dean, and the Google Brain team for their help. He, K., Gkioxari, G., Dollár, P., and Girshick, R. Mask
r-cnn. ICCV, pp. 2980–2988, 2017.
Appendix
He, Y., Lin, J., Liu, Z., Wang, H., Li, L.-J., and Han, S.
Since 2017, most research papers only report and compare Amc: Automl for model compression and acceleration
ImageNet validation accuracy; this paper also follows this on mobile devices. ECCV, 2018.
convention for better comparison. In addition, we have
also verified the test accuracy by submitting our predictions Hendrycks, D. and Gimpel, K. Gaussian error linear units
on the 100k test set images to http://image-net.org; (gelus). arXiv preprint arXiv:1606.08415, 2016.
results are in Table 8. As expected, the test accuracy is very
close to the validation accuracy. Howard, A. G., Zhu, M., Chen, B., Kalenichenko, D., Wang,
W., Weyand, T., Andreetto, M., and Adam, H. Mobilenets:
Table 8. ImageNet Validation vs. Test Top-1/5 Accuracy. Efficient convolutional neural networks for mobile vision
B0 B1 B2 B3 B4 B5 B6 B7 applications. arXiv preprint arXiv:1704.04861, 2017.
Val top1 77.11 79.13 80.07 81.59 82.89 83.60 83.95 84.26
Test top1 77.23 79.17 80.16 81.72 82.94 83.69 84.04 84.33 Hu, J., Shen, L., and Sun, G. Squeeze-and-excitation net-
Val top5 93.35 94.47 94.90 95.67 96.37 96.71 96.76 96.97 works. CVPR, 2018.
Test top5 93.45 94.43 94.98 95.70 96.27 96.64 96.86 96.94
Huang, G., Sun, Y., Liu, Z., Sedra, D., and Weinberger,
References K. Q. Deep networks with stochastic depth. ECCV, pp.
646–661, 2016.
Berg, T., Liu, J., Woo Lee, S., Alexander, M. L., Jacobs,
D. W., and Belhumeur, P. N. Birdsnap: Large-scale Huang, G., Liu, Z., Van Der Maaten, L., and Weinberger,
fine-grained visual categorization of birds. CVPR, pp. K. Q. Densely connected convolutional networks. CVPR,
2011–2018, 2014. 2017.
Bossard, L., Guillaumin, M., and Van Gool, L. Food-101– Huang, Y., Cheng, Y., Chen, D., Lee, H., Ngiam, J., Le,
mining discriminative components with random forests. Q. V., and Chen, Z. Gpipe: Efficient training of giant
ECCV, pp. 446–461, 2014. neural networks using pipeline parallelism. arXiv preprint
Cai, H., Zhu, L., and Han, S. Proxylessnas: Direct neural arXiv:1808.07233, 2018.
architecture search on target task and hardware. ICLR,
Iandola, F. N., Han, S., Moskewicz, M. W., Ashraf, K.,
2019.
Dally, W. J., and Keutzer, K. Squeezenet: Alexnet-level
Chollet, F. Xception: Deep learning with depthwise separa- accuracy with 50x fewer parameters and <0.5 mb model
ble convolutions. CVPR, pp. 1610–02357, 2017. size. arXiv preprint arXiv:1602.07360, 2016.
EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks
Ioffe, S. and Szegedy, C. Batch normalization: Accelerating Raghu, M., Poole, B., Kleinberg, J., Ganguli, S., and Sohl-
deep network training by reducing internal covariate shift. Dickstein, J. On the expressive power of deep neural
ICML, pp. 448–456, 2015. networks. ICML, 2017.
Kornblith, S., Shlens, J., and Le, Q. V. Do better imagenet Ramachandran, P., Zoph, B., and Le, Q. V. Searching for
models transfer better? CVPR, 2019. activation functions. arXiv preprint arXiv:1710.05941,
2018.
Krause, J., Deng, J., Stark, M., and Fei-Fei, L. Collecting a
large-scale dataset of fine-grained cars. Second Workshop Real, E., Aggarwal, A., Huang, Y., and Le, Q. V. Regu-
on Fine-Grained Visual Categorizatio, 2013. larized evolution for image classifier architecture search.
AAAI, 2019.
Krizhevsky, A. and Hinton, G. Learning multiple layers of
features from tiny images. Technical Report, 2009. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S.,
Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein,
Krizhevsky, A., Sutskever, I., and Hinton, G. E. Imagenet M., et al. Imagenet large scale visual recognition chal-
classification with deep convolutional neural networks. lenge. International Journal of Computer Vision, 115(3):
In NIPS, pp. 1097–1105, 2012. 211–252, 2015.
Lin, H. and Jegelka, S. Resnet with one-neuron hidden Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., and
layers is a universal approximator. NeurIPS, pp. 6172– Chen, L.-C. Mobilenetv2: Inverted residuals and linear
6181, 2018. bottlenecks. CVPR, 2018.
Lin, T.-Y., Dollár, P., Girshick, R., He, K., Hariharan, B., Sharir, O. and Shashua, A. On the expressive power of
and Belongie, S. Feature pyramid networks for object overlapping architectures of deep learning. ICLR, 2018.
detection. CVPR, 2017.
Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I.,
Liu, C., Zoph, B., Shlens, J., Hua, W., Li, L.-J., Fei-Fei, L., and Salakhutdinov, R. Dropout: a simple way to prevent
Yuille, A., Huang, J., and Murphy, K. Progressive neural neural networks from overfitting. The Journal of Machine
architecture search. ECCV, 2018. Learning Research, 15(1):1929–1958, 2014.
Lu, Z., Pu, H., Wang, F., Hu, Z., and Wang, L. The expres- Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S.,
sive power of neural networks: A view from the width. Anguelov, D., Erhan, D., Vanhoucke, V., and Rabinovich,
NeurIPS, 2018. A. Going deeper with convolutions. CVPR, pp. 1–9,
2015.
Ma, N., Zhang, X., Zheng, H.-T., and Sun, J. Shufflenet v2:
Practical guidelines for efficient cnn architecture design. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., and Wojna,
ECCV, 2018. Z. Rethinking the inception architecture for computer
vision. CVPR, pp. 2818–2826, 2016.
Mahajan, D., Girshick, R., Ramanathan, V., He, K., Paluri,
M., Li, Y., Bharambe, A., and van der Maaten, L. Explor- Szegedy, C., Ioffe, S., Vanhoucke, V., and Alemi, A. A.
ing the limits of weakly supervised pretraining. arXiv Inception-v4, inception-resnet and the impact of residual
preprint arXiv:1805.00932, 2018. connections on learning. AAAI, 4:12, 2017.
Maji, S., Rahtu, E., Kannala, J., Blaschko, M., and Vedaldi, Tan, M., Chen, B., Pang, R., Vasudevan, V., Sandler, M.,
A. Fine-grained visual classification of aircraft. arXiv Howard, A., and Le, Q. V. MnasNet: Platform-aware
preprint arXiv:1306.5151, 2013. neural architecture search for mobile. CVPR, 2019.
Ngiam, J., Peng, D., Vasudevan, V., Kornblith, S., Le, Q. V., Xie, S., Girshick, R., Dollár, P., Tu, Z., and He, K. Aggre-
and Pang, R. Domain adaptive transfer learning with spe- gated residual transformations for deep neural networks.
cialist models. arXiv preprint arXiv:1811.07056, 2018. CVPR, pp. 5987–5995, 2017.
Nilsback, M.-E. and Zisserman, A. Automated flower clas- Yang, T.-J., Howard, A., Chen, B., Zhang, X., Go, A., Sze,
sification over a large number of classes. ICVGIP, pp. V., and Adam, H. Netadapt: Platform-aware neural net-
722–729, 2008. work adaptation for mobile applications. ECCV, 2018.
Parkhi, O. M., Vedaldi, A., Zisserman, A., and Jawahar, C. Zagoruyko, S. and Komodakis, N. Wide residual networks.
Cats and dogs. CVPR, pp. 3498–3505, 2012. BMVC, 2016.
EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks
Zhang, X., Li, Z., Loy, C. C., and Lin, D. Polynet: A pursuit
of structural diversity in very deep networks. CVPR, pp.
3900–3908, 2017.
Zhang, X., Zhou, X., Lin, M., and Sun, J. Shufflenet: An ex-
tremely efficient convolutional neural network for mobile
devices. CVPR, 2018.
Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., and Torralba,
A. Learning deep features for discriminative localization.
CVPR, pp. 2921–2929, 2016.
Zoph, B. and Le, Q. V. Neural architecture search with
reinforcement learning. ICLR, 2017.
Zoph, B., Vasudevan, V., Shlens, J., and Le, Q. V. Learning
transferable architectures for scalable image recognition.
CVPR, 2018.
Source notes and reports