EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks

Mingxing Tan 1 Quoc V. Le 1

Abstract EfficientNet-B7
84
B6
Convolutional Neural Networks (ConvNets) are B5 AmoebaNet-A AmoebaNet-C
commonly developed at a fixed resource budget, 82
B4 NASNet-A SENet

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
ResNet-152 (He et al., 2016) 77.8% 60M
on this observation, we propose a new scaling DenseNet-201 EfficientNet-B1 78.8% 7.8M
B0 ResNeXt-101 (Xie et al., 2017) 80.9% 84M
method that uniformly scales all dimensions of 76
ResNet-50
EfficientNet-B3
SENet (Hu et al., 2018)
81.1%
82.7%
12M
146M
depth/width/resolution using a simple yet highly NASNet-A (Zoph et al., 2018)
EfficientNet-B4
82.7%
82.6%
89M
19M
Inception-v2
effective compound coefficient. We demonstrate 74
GPipe (Huang et al., 2018) †
EfficientNet-B7
84.3%
84.4%
556M
66M
NASNet-A †
the effectiveness of this method on scaling up ResNet-34
Not plotted

MobileNets and ResNet. 0 20 40 60 80 100 120 140 160 180

Number of Parameters (Millions)
To go even further, we use neural architecture Figure 1. Model Size vs. ImageNet Accuracy. All numbers are
search to design a new baseline network and scale for single-crop, single-model. Our EfficientNets significantly out-
it up to obtain a family of models, called Efficient- perform other ConvNets. In particular, EfficientNet-B7 achieves
Nets, which achieve much better accuracy and effi- new state-of-the-art 84.4% top-1 accuracy but being 8.4x smaller
ciency than previous ConvNets. In particular, our and 6.1x faster than GPipe. EfficientNet-B1 is 7.6x smaller and
EfficientNet-B7 achieves state-of-the-art 84.4% 5.7x faster than ResNet-152. Details are in Table 2 and 4.
top-1 / 97.1% top-5 accuracy on ImageNet, while
being 8.4x smaller and 6.1x faster on inference
than the best existing ConvNet. Our EfficientNets ways to do it. The most common way is to scale up Con-
also transfer well and achieve state-of-the-art ac- vNets by their depth (He et al., 2016) or width (Zagoruyko &
curacy on CIFAR-100 (91.7%), Flowers (98.8%), Komodakis, 2016). Another less common, but increasingly
and 3 other transfer learning datasets, with an popular, method is to scale up models by image resolution
order of magnitude fewer parameters. (Huang et al., 2018). In previous work, it is common to scale
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
1. Introduction and still often yields sub-optimal accuracy and efficiency.

Scaling up ConvNets is widely used to achieve better accu- In this paper, we want to study and rethink the process
racy. For example, ResNet (He et al., 2016) can be scaled of scaling up ConvNets. In particular, we investigate the
up from ResNet-18 to ResNet-200 by using more layers; central question: is there a principled method to scale up
Recently, GPipe (Huang et al., 2018) achieved 84.3% Ima- ConvNets that can achieve better accuracy and efficiency?
geNet top-1 accuracy by scaling up a baseline model four Our empirical study shows that it is critical to balance all
time larger. However, the process of scaling up ConvNets dimensions of network width/depth/resolution, and surpris-
has never been well understood and there are currently many ingly such balance can be achieved by simply scaling each
of them with constant ratio. Based on this observation, we
1
Google Research, Brain Team, Mountain View, CA. Corre- propose a simple yet effective compound scaling method.
spondence to: Mingxing Tan <[email protected]>. Unlike conventional practice that arbitrary scales these fac-
Proceedings of the 36 th International Conference on Machine tors, our method uniformly scales network width, depth,
Learning, Long Beach, California, PMLR 97, 2019. Copyright and resolution with a set of fixed scaling coefficients. For
2019 by the author(s). example, if we want to use 2N times more computational
 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.

resources, then we can simply increase the network depth by 2. Related Work
αN , width by β N , and image size by γ N , where α, β, γ are
constant coefficients determined by a small grid search on ConvNet Accuracy: Since AlexNet (Krizhevsky et al.,
the original small model. Figure 2 illustrates the difference 2012) won the 2012 ImageNet competition, ConvNets have
between our scaling method and conventional methods. become increasingly more accurate by going bigger: while
the 2014 ImageNet winner GoogleNet (Szegedy et al., 2015)
Intuitively, the compound scaling method makes sense be- achieves 74.8% top-1 accuracy with about 6.8M parameters,
cause if the input image is bigger, then the network needs the 2017 ImageNet winner SENet (Hu et al., 2018) achieves
more layers to increase the receptive field and more channels 82.7% top-1 accuracy with 145M parameters. Recently,
to capture more fine-grained patterns on the bigger image. In GPipe (Huang et al., 2018) further pushes the state-of-the-art
fact, previous theoretical (Raghu et al., 2017; Lu et al., 2018) ImageNet top-1 validation accuracy to 84.3% using 557M
and empirical results (Zagoruyko & Komodakis, 2016) both parameters: it is so big that it can only be trained with a
show that there exists certain relationship between network specialized pipeline parallelism library by partitioning the
width and depth, but to our best knowledge, we are the network and spreading each part to a different accelera-
first to empirically quantify the relationship among all three tor. While these models are mainly designed for ImageNet,
dimensions of network width, depth, and resolution. recent studies have shown better ImageNet models also per-
We demonstrate that our scaling method work well on exist- form better across a variety of transfer learning datasets
ing MobileNets (Howard et al., 2017; Sandler et al., 2018) (Kornblith et al., 2019), and other computer vision tasks
and ResNet (He et al., 2016). Notably, the effectiveness of such as object detection (He et al., 2016; Tan et al., 2019).
model scaling heavily depends on the baseline network; to Although higher accuracy is critical for many applications,
go even further, we use neural architecture search (Zoph & we have already hit the hardware memory limit, and thus
Le, 2017; Tan et al., 2019) to develop a new baseline net- further accuracy gain needs better efficiency.
work, and scale it up to obtain a family of models, called Effi-
cientNets. Figure 1 summarizes the ImageNet performance, ConvNet Efficiency: Deep ConvNets are often over-
where our EfficientNets significantly outperform other Con- parameterized. Model compression (Han et al., 2016; He
vNets. In particular, our EfficientNet-B7 surpasses the best et al., 2018; Yang et al., 2018) is a common way to re-
existing GPipe accuracy (Huang et al., 2018), but using duce model size by trading accuracy for efficiency. As mo-
8.4x fewer parameters and running 6.1x faster on inference. bile phones become ubiquitous, it is also common to hand-
Compared to the widely used ResNet (He et al., 2016), our craft efficient mobile-size ConvNets, such as SqueezeNets
EfficientNet-B4 improves the top-1 accuracy from 76.3% (Iandola et al., 2016; Gholami et al., 2018), MobileNets
of ResNet-50 to 82.6% with similar FLOPS. Besides Ima- (Howard et al., 2017; Sandler et al., 2018), and ShuffleNets
geNet, EfficientNets also transfer well and achieve state-of- (Zhang et al., 2018; Ma et al., 2018). Recently, neural archi-
the-art accuracy on 5 out of 8 widely used datasets, while tecture search becomes increasingly popular in designing
reducing parameters by up to 21x than existing ConvNets. efficient mobile-size ConvNets (Tan et al., 2019; Cai et al.,
 EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks

2019), and achieves even better efficiency than hand-crafted dimension is expanded over layers, for example, from initial
mobile ConvNets by extensively tuning the network width, input shape h224, 224, 3i to final output shape h7, 7, 512i.
depth, convolution kernel types and sizes. However, it is
Unlike regular ConvNet designs that mostly focus on find-
unclear how to apply these techniques for larger models that
ing the best layer architecture Fi , model scaling tries to ex-
have much larger design space and much more expensive
pand the network length (Li ), width (Ci ), and/or resolution
tuning cost. In this paper, we aim to study model efficiency
(Hi , Wi ) without changing Fi predefined in the baseline
for super large ConvNets that surpass state-of-the-art accu-
network. By fixing Fi , model scaling simplifies the design
racy. To achieve this goal, we resort to model scaling.
problem for new resource constraints, but it still remains
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

input image size will help accuracy with the overhead of max Accuracy N (d, w, r)
d,w,r
more FLOPS. Although prior studies (Raghu et al., 2017; K
F̂id·L̂i Xhr·Ĥi ,r·Ŵi ,w·Ĉi i


Lin & Jegelka, 2018; Sharir & Shashua, 2018; Lu et al., s.t. N (d, w, r) =
i=1...s
2018) have shown that network deep and width are both
important for ConvNets’ expressive power, it still remains Memory(N ) ≤ target memory
an open question of how to effectively scale a ConvNet to FLOPS(N ) ≤ target flops
achieve better efficiency and accuracy. Our work systemati- (2)
cally and empirically studies ConvNet scaling for all three where w, d, r are coefficients for scaling network width,
dimensions of network width, depth, and resolutions. depth, and resolution; F̂i , L̂i , Ĥi , Ŵi , Ĉi are predefined pa-
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
different approaches, and propose our new scaling method. The main difficulty of problem 2 is that the optimal d, w, r
depend on each other and the values change under different
3.1. Problem Formulation resource constraints. Due to this difficulty, conventional
methods mostly scale ConvNets in one of these dimensions:
A ConvNet Layer i can be defined as a function: Yi =
Fi (Xi ), where Fi is the operator, Yi is output tensor, Xi is
input tensor, with tensor shape hHi , Wi , Ci i1 , where Hi and Depth (dd): Scaling network depth is the most common way
Wi are spatial dimension and Ci is the channel dimension. used by many ConvNets (He et al., 2016; Huang et al., 2017;
A ConvNet N can be represented by aJ list of composed lay- Szegedy et al., 2015; 2016). The intuition is that deeper
ers: N = Fk  ...  F1  F1 (X1 ) = j=1...k Fj (X1 ). In ConvNet can capture richer and more complex features, and
practice, ConvNet layers are often partitioned into multiple generalize well on new tasks. However, deeper networks
stages and all layers in each stage share the same architec- are also more difficult to train due to the vanishing gradient
ture: for example, ResNet (He et al., 2016) has five stages, problem (Zagoruyko & Komodakis, 2016). Although sev-
and all layers in each stage has the same convolutional type eral techniques, such as skip connections (He et al., 2016)
except the first layer performs down-sampling. Therefore, and batch normalization (Ioffe & Szegedy, 2015), alleviate
we can define a ConvNet as: the training problem, the accuracy gain of very deep network
K diminishes: for example, ResNet-1000 has similar accuracy
FiLi XhHi ,Wi ,Ci i


N = (1) as ResNet-101 even though it has much more layers. Figure
i=1...s 3 (middle) shows our empirical study on scaling a baseline
model with different depth coefficient d, further suggesting
where FiLi denotes layer Fi is repeated Li times in stage i,
the diminishing accuracy return for very deep ConvNets.
hHi , Wi , Ci i denotes the shape of input tensor X of layer
i. Figure 2(a) illustrate a representative ConvNet, where
the spatial dimension is gradually shrunk but the channel
Width (w 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 (rr ): 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 depth 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 28 × 28 112 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 theocratic 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 76.3% 93.2% 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 78.8% 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 79.8% 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 81.1% 95.5% 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 82.6% 96.3% 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 83.3% 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% 96.9% 43M 1x 19B 1x
EfficientNet-B7 84.4% 97.1% 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.

84 EfficientNet-B6
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%

Scale MobileNetV1 by resolution (r=2) 2.2B 72.7% B3 ResNeXt-101
compound scale (dd=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 ResNet-152 (Xie et al., 2017) 77.8% 11B
Scale MobileNetV2 by resolution (r=2) 1.2B 74.8% B0
EfficientNet-B1 78.8% 0.7B
76 ResNeXt-101 (Xie et al., 2017) 80.9% 32B
MobileNetV2 compound scale 1.3B 77.4% ResNet-50 EfficientNet-B3 81.1% 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 82.6% 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 83.3% 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.

Table 4. Inference Latency Comparison – Latency is measured 5.2. ImageNet Results for EfficientNet
with batch size 1 on a single core of Intel Xeon CPU E5-2690.
We train our EfficientNet models on ImageNet using simi-
Acc. @ Latency Acc. @ Latency lar settings as (Tan et al., 2019): RMSProp optimizer with
ResNet-152 77.8% @ 0.554s GPipe 84.3% @ 19.0s decay 0.9 and momentum 0.9; batch norm momentum 0.99;
EfficientNet-B1 78.8% @ 0.098s EfficientNet-B7 84.4% @ 3.1s weight decay 1e-5; initial learning rate 0.256 that decays
Speedup 5.7x Speedup 6.1x
by 0.97 every 2.4 epochs. We also use swish activation
 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 70 91
0.6
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 1020.4 103 101 0.6 102 103 101
0.8 102 1.03
10
Number of Parameters (Millions, log-scale)
Inception-v1 Inception-v4 ResNet-50 ResNet-152 DenseNet-201 GPIPE
Inception-v3 Inception-ResNet-v2 ResNet-101 DenseNet-169 NASNet-A EfficientNet

Figure 6. Model Parameters vs. Transfer Learning Accuracy – All models are pretrained on ImageNet and finetuned on new datasets.

(Ramachandran et al., 2018; Elfwing et al., 2018), fixed Au- ConvNets. Notably, our EfficientNet models are not only
toAugment policy (Cubuk et al., 2019), and stochastic depth small, but also computational cheaper. For example, our
(Huang et al., 2016) with drop connect ratio 0.3. As com- EfficientNet-B3 achieves higher accuracy than ResNeXt-
monly known that bigger models need more regularization, 101 (Xie et al., 2017) using 18x fewer FLOPS.
we linearly increase dropout (Srivastava et al., 2014) ratio
To validate the computational cost, we have also mea-
from 0.2 for EfficientNet-B0 to 0.5 for EfficientNet-B7.
sured the inference latency on a real CPU as shown in
Table 2 shows the performance of all EfficientNet models Table 4, where we report average latency over 20 runs.
that are scaled from the same baseline EfficientNet-B0. Our Our EfficientNet-B1 runs 5.7x faster than the widely used
EfficientNet models generally use an order of magnitude ResNet-152 (He et al., 2016), while EfficientNet-B7 runs
fewer parameters and FLOPS than other ConvNets with about 6.1x faster than GPipe (Huang et al., 2018), suggest-
similar accuracy. In particular, our EfficientNet-B7 achieves ing our EfficientNets are indeed fast on real hardware.
84.4% top1 / 97.1% top-5 accuracy with 66M parameters
and 37B FLOPS, being more accurate but 8.4x smaller than 5.3. Transfer Learning Results for EfficientNet
the previous best GPipe (Huang et al., 2018).
We have also evaluated our EfficientNet on a list of com-
Figure 1 and Figure 5 illustrates the parameters-accuracy monly used transfer learning datasets, as shown in Table
and FLOPS-accuracy curve for representative ConvNets, 6. We borrow the same training settings from (Kornblith
where our scaled EfficientNet models achieve better accu- et al., 2019) and (Huang et al., 2018), which take ImageNet
racy with much fewer parameters and FLOPS than other pretrained checkpoints and finetune on new datasets.
 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.

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 80
CIFAR-100 (Krizhevsky & Hinton, 2009) 50,000 10,000 100
Birdsnap (Berg et al., 2014) 47,386 2,443 500 79
Stanford Cars (Krause et al., 2013) 8,144 8,041 196
78
Flowers (Nilsback & Zisserman, 2008) 2,040 6,149 102 scale by width
FGVC Aircraft (Maji et al., 2013) 6,667 3,333 100 77 scale by depth
Oxford-IIIT Pets (Parkhi et al., 2012) 3,680 3,369 37 scale by resolution
Food-101 (Bossard et al., 2014) 75,750 25,250 101 76
compound scaling
75
0 1 2 3 4 5
FLOPS (Billions)
Table 5 shows the transfer learning performance: (1) Com- Figure 8. Scaling Up EfficientNet-B0 with Different Methods.
pared to public available models, such as NASNet-A (Zoph
et al., 2018) and Inception-v4 (Szegedy et al., 2017), our Ef-
In order to further understand why our compound scaling
ficientNet models achieve better accuracy with 4.7x average
method is better than others, Figure 7 compares the class ac-
(up to 21x) parameter reduction. (2) Compared to state-
tivation map for a few representative models with different
of-the-art models, including DAT (Ngiam et al., 2018) that
scaling methods. All these models are scaled from the same
dynamically synthesizes training data and GPipe (Huang
EfficientNet-B0 baseline with about 4x more FLOPS than
et al., 2018) that is trained with specialized pipeline paral-
the baseline. Images are randomly picked from ImageNet
lelism, our EfficientNet models still surpass their accuracy
validation set. As shown in the figure, the model with com-
in 5 out of 8 datasets, but using 9.6x fewer parameters
pound scaling tends to focus on more relevant regions with
Figure 6 compares the accuracy-parameters curve for a va- more object details, while other models are either lack of
riety of models. In general, our EfficientNets consistently object details or unable to capture all objects in the images.
achieve better accuracy with an order of magnitude fewer pa-
rameters than existing models, including ResNet (He et al., 7. Conclusion
2016), DenseNet (Huang et al., 2017), Inception (Szegedy
et al., 2017), and NASNet (Zoph et al., 2018). In this paper, we systematically study ConvNet scaling and
identify that carefully balancing network width, depth, and
6. Discussion resolution is an important but missing piece, preventing us
from better accuracy and efficiency. To address this issue,
To disentangle the contribution of our proposed scaling we propose a simple and highly effective compound scaling
method from the EfficientNet architecture, Figure 8 com- method, which enables us to easily scale up a baseline Con-
pares the ImageNet performance of different scaling meth- vNet to any target resource constraints in a more principled
ods for the same EfficientNet-B0 baseline network. In gen- way, while maintaining model efficiency. Powered by this
eral, all scaling methods improve accuracy with the cost compound scaling method, we demonstrate that a mobile-
of more FLOPS, but our compound scaling method can size EfficientNet model can be scaled up very effectively,
further improve accuracy, by up to 2.5%, than other single- surpassing state-of-the-art accuracy with an order of magni-
dimension scaling methods, suggesting the importance of tude fewer parameters and FLOPS, on both ImageNet and
our proposed compound scaling. five commonly used transfer learning datasets.
 EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks

Acknowledgements Hu, J., Shen, L., and Sun, G. Squeeze-and-excitation net-

works. CVPR, 2018.
We thank Ruoming Pang, Vijay Vasudevan, Alok Aggarwal,
Barret Zoph, Hongkun Yu, Xiaodan Song, Samy Bengio, Huang, G., Sun, Y., Liu, Z., Sedra, D., and Weinberger,
Jeff Dean, and Google Brain team for their help. K. Q. Deep networks with stochastic depth. ECCV, pp.
646–661, 2016.
References Huang, G., Liu, Z., Van Der Maaten, L., and Weinberger,
Berg, T., Liu, J., Woo Lee, S., Alexander, M. L., Jacobs, K. Q. Densely connected convolutional networks. CVPR,
D. W., and Belhumeur, P. N. Birdsnap: Large-scale 2017.
fine-grained visual categorization of birds. CVPR, pp.
2011–2018, 2014. Huang, Y., Cheng, Y., Chen, D., Lee, H., Ngiam, J., Le,
Q. V., and Chen, Z. Gpipe: Efficient training of giant
Bossard, L., Guillaumin, M., and Van Gool, L. Food-101– neural networks using pipeline parallelism. arXiv preprint
mining discriminative components with random forests. arXiv:1808.07233, 2018.
ECCV, pp. 446–461, 2014.
Iandola, F. N., Han, S., Moskewicz, M. W., Ashraf, K.,
Cai, H., Zhu, L., and Han, S. Proxylessnas: Direct neural Dally, W. J., and Keutzer, K. Squeezenet: Alexnet-level
architecture search on target task and hardware. ICLR, accuracy with 50x fewer parameters and <0.5 mb model
2019. size. arXiv preprint arXiv:1602.07360, 2016.

Chollet, F. Xception: Deep learning with depthwise separa- Ioffe, S. and Szegedy, C. Batch normalization: Accelerating
ble convolutions. CVPR, pp. 1610–02357, 2017. deep network training by reducing internal covariate shift.
ICML, pp. 448–456, 2015.
Cubuk, E. D., Zoph, B., Mane, D., Vasudevan, V., and Le,
Q. V. Autoaugment: Learning augmentation policies Kornblith, S., Shlens, J., and Le, Q. V. Do better imagenet
from data. CVPR, 2019. models transfer better? CVPR, 2019.

Elfwing, S., Uchibe, E., and Doya, K. Sigmoid-weighted Krause, J., Deng, J., Stark, M., and Fei-Fei, L. Collecting a
linear units for neural network function approximation large-scale dataset of fine-grained cars. Second Workshop
in reinforcement learning. Neural Networks, 107:3–11, on Fine-Grained Visual Categorizatio, 2013.
2018. Krizhevsky, A. and Hinton, G. Learning multiple layers of
features from tiny images. Technical Report, 2009.
Gholami, A., Kwon, K., Wu, B., Tai, Z., Yue, X., Jin, P.,
Zhao, S., and Keutzer, K. Squeezenext: Hardware-aware Krizhevsky, A., Sutskever, I., and Hinton, G. E. Imagenet
neural network design. ECV Workshop at CVPR’18, classification with deep convolutional neural networks.
2018. In NIPS, pp. 1097–1105, 2012.
Han, S., Mao, H., and Dally, W. J. Deep compression: Lin, H. and Jegelka, S. Resnet with one-neuron hidden
Compressing deep neural networks with pruning, trained layers is a universal approximator. NeurIPS, pp. 6172–
quantization and huffman coding. ICLR, 2016. 6181, 2018.
He, K., Zhang, X., Ren, S., and Sun, J. Deep residual Lin, T.-Y., Dollár, P., Girshick, R., He, K., Hariharan, B.,
learning for image recognition. CVPR, pp. 770–778, and Belongie, S. Feature pyramid networks for object
2016. detection. CVPR, 2017.
He, K., Gkioxari, G., Dollár, P., and Girshick, R. Mask Liu, C., Zoph, B., Shlens, J., Hua, W., Li, L.-J., Fei-Fei, L.,
r-cnn. ICCV, pp. 2980–2988, 2017. Yuille, A., Huang, J., and Murphy, K. Progressive neural
architecture search. ECCV, 2018.
He, Y., Lin, J., Liu, Z., Wang, H., Li, L.-J., and Han, S.
Amc: Automl for model compression and acceleration Lu, Z., Pu, H., Wang, F., Hu, Z., and Wang, L. The expres-
on mobile devices. ECCV, 2018. sive power of neural networks: A view from the width.
NeurIPS, 2018.
Howard, A. G., Zhu, M., Chen, B., Kalenichenko, D., Wang,
W., Weyand, T., Andreetto, M., and Adam, H. Mobilenets: Ma, N., Zhang, X., Zheng, H.-T., and Sun, J. Shufflenet v2:
Efficient convolutional neural networks for mobile vision Practical guidelines for efficient cnn architecture design.
applications. arXiv preprint arXiv:1704.04861, 2017. ECCV, 2018.
 EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks

Mahajan, D., Girshick, R., Ramanathan, V., He, K., Paluri, Szegedy, C., Ioffe, S., Vanhoucke, V., and Alemi, A. A.
M., Li, Y., Bharambe, A., and van der Maaten, L. Explor- Inception-v4, inception-resnet and the impact of residual
ing the limits of weakly supervised pretraining. arXiv connections on learning. AAAI, 4:12, 2017.
preprint arXiv:1805.00932, 2018.
Tan, M., Chen, B., Pang, R., Vasudevan, V., Sandler, M.,
Maji, S., Rahtu, E., Kannala, J., Blaschko, M., and Vedaldi, Howard, A., and Le, Q. V. MnasNet: Platform-aware
A. Fine-grained visual classification of aircraft. arXiv neural architecture search for mobile. CVPR, 2019.
preprint arXiv:1306.5151, 2013.
Xie, S., Girshick, R., Dollár, P., Tu, Z., and He, K. Aggre-
Ngiam, J., Peng, D., Vasudevan, V., Kornblith, S., Le, Q. V., gated residual transformations for deep neural networks.
and Pang, R. Domain adaptive transfer learning with spe- CVPR, pp. 5987–5995, 2017.
cialist models. arXiv preprint arXiv:1811.07056, 2018.
Yang, T.-J., Howard, A., Chen, B., Zhang, X., Go, A., Sze,
Nilsback, M.-E. and Zisserman, A. Automated flower clas- V., and Adam, H. Netadapt: Platform-aware neural net-
sification over a large number of classes. ICVGIP, pp. work adaptation for mobile applications. ECCV, 2018.
722–729, 2008.
Zagoruyko, S. and Komodakis, N. Wide residual networks.
Parkhi, O. M., Vedaldi, A., Zisserman, A., and Jawahar, C. BMVC, 2016.
Cats and dogs. CVPR, pp. 3498–3505, 2012.
Zhang, X., Li, Z., Loy, C. C., and Lin, D. Polynet: A pursuit
Raghu, M., Poole, B., Kleinberg, J., Ganguli, S., and Sohl- of structural diversity in very deep networks. CVPR, pp.
Dickstein, J. On the expressive power of deep neural 3900–3908, 2017.
networks. ICML, 2017. Zhang, X., Zhou, X., Lin, M., and Sun, J. Shufflenet: An ex-
Ramachandran, P., Zoph, B., and Le, Q. V. Searching for tremely efficient convolutional neural network for mobile
activation functions. arXiv preprint arXiv:1710.05941, devices. CVPR, 2018.
2018. Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., and Torralba,
Real, E., Aggarwal, A., Huang, Y., and Le, Q. V. Regu- A. Learning deep features for discriminative localization.
larized evolution for image classifier architecture search. CVPR, pp. 2921–2929, 2016.
AAAI, 2019. Zoph, B. and Le, Q. V. Neural architecture search with
reinforcement learning. ICLR, 2017.
Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S.,
Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, Zoph, B., Vasudevan, V., Shlens, J., and Le, Q. V. Learning
M., et al. Imagenet large scale visual recognition chal- transferable architectures for scalable image recognition.
lenge. International Journal of Computer Vision, 115(3): CVPR, 2018.
211–252, 2015.

Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., and

Chen, L.-C. Mobilenetv2: Inverted residuals and linear
bottlenecks. CVPR, 2018.

Sharir, O. and Shashua, A. On the expressive power of

overlapping architectures of deep learning. ICLR, 2018.

Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I.,

and Salakhutdinov, R. Dropout: a simple way to prevent
neural networks from overfitting. The Journal of Machine
Learning Research, 15(1):1929–1958, 2014.

Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S.,
Anguelov, D., Erhan, D., Vanhoucke, V., and Rabinovich,
A. Going deeper with convolutions. CVPR, pp. 1–9,
2015.

Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., and Wojna,
Z. Rethinking the inception architecture for computer
vision. CVPR, pp. 2818–2826, 2016.