MobileNetV2——Inverted Residuals and Linear Bottlenecks论文翻译——中英文对照

博客  |  CSDN  |  简书



MobileNetV2: Inverted Residuals and Linear Bottlenecks


In this paper we describe a new mobile architecture, MobileNetV2, that improves the state of the art performance of mobile models on multiple tasks and benchmarks as well as across a spectrum of different model sizes. We also describe efficient ways of applying these mobile models to object detection in a novel framework we call SSDLite. Additionally, we demonstrate how to build mobile semantic segmentation models through a reduced form of DeepLabv3 which we call Mobile DeepLabv3.


在本文中,我们描述了一种新的移动架构MobileNetV2,该架构提高了移动模型在多个任务和多个基准数据集上以及在不同模型尺寸范围内的最佳性能。我们还描述了在我们称之为SSDLite的新框架中将这些移动模型应用于目标检测的有效方法。此外,我们还演示了如何通过DeepLabv3的简化形式,我们称之为Mobile DeepLabv3来构建移动语义分割模型。

The MobileNetV2 architecture is based on an inverted residual structure where the shortcut connections are between the thin bottle-neck layers. The intermediate expansion layer uses lightweight depthwise convolutions to filter features as a source of non-linearity. Additionally, we find that it is important to remove non-linearities in the narrow layers in order to maintain representational power. We demonstrate that this improves performance and provide an intuition that led to this design.


Finally, our approach allows decoupling of the input/output domains from the expressiveness of the transformation, which provides a convenient framework for further analysis. We measure our performance on ImageNet [1] classification, COCO object detection [2], VOC image segmentation [3]. We evaluate the trade-offs between accuracy, and number of operations measured by multiply-adds (MAdd), as well as actual latency, and the number of parameters.


1. Introduction

Neural networks have revolutionized many areas of machine intelligence, enabling superhuman accuracy for challenging image recognition tasks. However, the drive to improve accuracy often comes at a cost: modern state of the art networks require high computational resources beyond the capabilities of many mobile and embedded applications.

1. 引言


This paper introduces a new neural network architecture that is specifically tailored for mobile and resource constrained environments. Our network pushes the state of the art for mobile tailored computer vision models, by significantly decreasing the number of operations and memory needed while retaining the same accuracy.


Our main contribution is a novel layer module: the inverted residual with linear bottleneck. This module takes as an input a low-dimensional compressed representation which is first expanded to high dimension and filtered with a lightweight depthwise convolution. Features are subsequently projected back to a low-dimensional representation with a linear convolution. The official implementation is available as part of TensorFlow-Slim model library in [4].


This module can be efficiently implemented using standard operations in any modern framework and allows our models to beat state of the art along multiple performance points using standard benchmarks. Furthermore, this convolutional module is particularly suitable for mobile designs, because it allows to significantly reduce the memory footprint needed during inference by never fully materializing large intermediate tensors. This reduces the need for main memory access in many embedded hardware designs, that provide small amounts of very fast software controlled cache memory.


Tuning deep neural architectures to strike an optimal balance between accuracy and performance has been an area of active research for the last several years. Both manual architecture search and improvements in training algorithms, carried out by numerous teams has lead to dramatic improvements over early designs such as AlexNet [5], VGGNet [6], GoogLeNet [7]. , and ResNet [8]. Recently there has been lots of progress in algorithmic architecture exploration included hyper-parameter optimization [9, 10, 11] as well as various methods of network pruning [12, 13, 14, 15, 16, 17] and connectivity learning [18, 19]. A substantial amount of work has also been dedicated to changing the connectivity structure of the internal convolutional blocks such as in ShuffleNet [20] or introducing sparsity [21] and others [22].

2. 相关工作

调整深层神经架构以在精确性和性能之间达到最佳平衡已成为过去几年研究活跃的一个领域。由许多团队进行的手动架构搜索和训练算法的改进,已经比早期的设计(如AlexNet[5],VGGNet [6],GoogLeNet[7]和ResNet[8])有了显著的改进。最近在算法架构探索方面取得了很多进展,包括超参数优化[9,10,11]、各种网络修剪方法[12,13,14,15,16,17]和连接学习[18,19]。 也有大量的工作致力于改变内部卷积块的连接结构如ShuffleNet[20]或引入稀疏性[21]和其他[22]。

Recently, [23, 24, 25, 26], opened up a new direction of bringing optimization methods including genetic algorithms and reinforcement learning to architectural search. However one drawback is that the resulting networks end up very complex. In this paper, we pursue the goal of developing better intuition about how neural networks operate and use that to guide the simplest possible network design. Our approach should be seen as complimentary to the one described in [23] and related work. In this vein our approach is similar to those taken by [20, 22] and allows to further improve the performance, while providing a glimpse on its internal operation. Our network design is based on MobileNetV1 [27]. It retains its simplicity and does not require any special operators while significantly improves its accuracy, achieving state of the art on multiple image classification and detection tasks for mobile applications.


3. Preliminaries, discussion and intuition

3.1. Depthwise Separable Convolutions

Depthwise Separable Convolutions are a key building block for many efficient neural network architectures [27, 28, 20] and we use them in the present work as well. The basic idea is to replace a full convolutional operator with a factorized version that splits convolution into two separate layers. The first layer is called a depthwise convolution, it performs lightweight filtering by applying a single convolutional filter per input channel. The second layer is a 1 × 1 convolution, called a pointwise convolution, which is responsible for building new features through computing linear combinations of the input channels.

3. 准备,讨论和直觉

3.1. 深度可分卷积


Standard convolution takes an $h_i\times w_i\times d_i$ input tensor $L_i$, and applies convolutional kernel $K\in \mathbf{R}^{k\times k \times d_i \times d_j}$ to produce an $h_i\times w_i\times d_j$ output tensor $L_j$. Standard convolutional layers have the computational cost of $h_i \cdot w_i \cdot d_i \cdot d_j \cdot k \cdot k$.

标准卷积使用$K\in \mathbf{R}^{k\times k \times d_i \times d_j}$维的输入张量$L_i$,并对其应用卷积核$K\in \mathbf{R}^{k\times k \times d_i \times d_j}$来产生$h_i\times w_i\times d_j$维的输出张量$L_j$。标准卷积层的计算代价为$h_i \cdot w_i \cdot d_i \cdot d_j \cdot k \cdot k$。

Depthwise separable convolutions are a drop-in replacement for standard convolutional layers. Empirically they work almost as well as regular convolutions but only cost: $$\begin{equation}h_i \cdot w_i \cdot d_i (k^2 + d_j) \tag{1}\end{equation}$$ which is the sum of the depthwise and $1 \times 1$ pointwise convolutions. Effectively depthwise separable convolution reduces computation compared to traditional layers by almost a factor of $k^2$. MobileNetV2 uses $k=3$ ($3 \times 3$ depthwise separable convolutions) so the computational cost is $8$ to $9$ times smaller than that of standard convolutions at only a small reduction in accuracy [27].

深度可分卷积是标准卷积层的直接替换。经验上,它们几乎与常规卷积一样工作,但其成本为:$$\begin{equation}h_i \cdot w_i \cdot d_i (k^2 + d_j) \tag{1}\end{equation}$$它是深度方向和$1\times 1$逐点卷积的总和。深度可分卷积与传统卷积层相比有效地减少了几乎$k^2$倍的计算量。MobileNetV2使用$k=3$($3\times 3$的深度可分卷积),因此计算成本比标准卷积小$8$到$9$倍,但精度只有很小的降低[27]。

3.2. Linear Bottlenecks

Consider a deep neural network consisting of $n$ layers $L_i$ each of which has an activation tensor of dimensions $h_i \times w_i \times d_i$. Throughout this section we will be discussing the basic properties of these activation tensors, which we will treat as containers of $h_i \times
w_i$ “pixels” with $d_i$ dimensions. Informally, for an input set of real images, we say that the set of layer activations (for any layer $L_i$) forms a “manifold of interest”. It has been long assumed that manifolds of interest in neural networks could be embedded in low-dimensional subspaces. In other words, when we look at all individual $d$-channel pixels of a deep convolutional layer, the information encoded in those values actually lie in some manifold, which in turn is embeddable into a low-dimensional subspace.

3.2. 线性瓶颈

考虑一个由$n$层$L_i$组成的深度神经网络,每层都有一个$h_i \times w_i \times d_i$维的激活张量。在本节中,我们将讨论这些激活张量的基本属性,我们将把它们看作$h_i \times

At a first glance, such a fact could then be captured and exploited by simply reducing the dimensionality of a layer thus reducing the dimensionality of the operating space. This has been successfully exploited by MobileNetV1 [27] to effectively trade off between computation and accuracy via a width multiplier parameter, and has been incorporated into efficient model designs of other networks as well [20]. Following that intuition, the width multiplier approach allows one to reduce the dimensionality of the activation space until the manifold of interest spans this entire space. However, this intuition breaks down when we recall that deep convolutional neural networks actually have non-linear per coordinate transformations, such as ReLU. For example, ReLU applied to a line in 1D space produces a ray, where as in $\mathbf{R}^n$ space, it generally results in a piece-wise linear curve with $n$-joints.

乍一看,这样的实例可以通过简单地减少层的维度来捕获和利用,从而降低操作空间的维度。这已经被MobileNetV1[27]成功利用,通过宽度乘数参数在计算量和精度之间进行有效折衷,并且已经被合并到其他网络的高效模型设计中[20]。遵循这种直觉,宽度乘数方法允许降低激活空间的维度,直到感兴趣的流形横跨整个空间为止。然而,当我们回想到深度卷积神经网络实际上具有非线性的每个坐标变换(例如ReLU)时,这种直觉就会失败。 例如,在1维空间中的一行应用ReLU会产生一个ray,在$\mathbf {R}^n$空间中,它通常会产生一个具有$n$个连接的分段线性曲线。

It is easy to see that in general if a result of a layer transformation ReLU(Bx) has a non-zero volume $S$, the points mapped to interior $S$ are obtained via a linear transformation $B$ of the input, thus indicating that the part of the input space corresponding to the full dimensional output, is limited to a linear transformation. In other words, deep networks only have the power of a linear classifier on the non-zero volume part of the output domain. We refer to supplemental material for a more formal statement.


On the other hand, when ReLU collapses the channel, it inevitably loses information in that channel. However if we have lots of channels, and there is a a structure in the activation manifold that information might still be preserved in the other channels. In supplemental materials, we show that if the input manifold can be embedded into a significantly lower-dimensional subspace of the activation space then the ReLU transformation preserves the information while introducing the needed complexity into the set of expressible functions.


To summarize, we have highlighted two properties that are indicative of the requirement that the manifold of interest should lie in a low-dimensional subspace of the higher-dimensional activation space:

  1. If the manifold of interest remains non-zero volume after ReLU transformation, it corresponds to a linear transformation.

  2. ReLU is capable of preserving complete information about the input manifold, but only if the input manifold lies in a low-dimensional subspace of the input space.




These two insights provide us with an empirical hint for optimizing existing neural architectures: assuming the manifold of interest is low-dimensional we can capture this by inserting linear bottleneck layers into the convolutional blocks. Experimental evidence suggests that using linear layers is crucial as it prevents non-linearities from destroying too much information. In Section 6, we show empirically that using non-linear layers in bottlenecks indeed hurts the performance by several percent, further validating our hypothesis. We note that similar reports where non-linearity was helped were reported in [29] where non-linearity was removed from the input of the traditional residual block and that lead to improved performance on CIFAR dataset.


For the remainder of this paper we will be utilizing bottleneck convolutions. We will refer to the ratio between the size of the input bottleneck and the inner size as the expansion ratio.


3.3. Inverted residuals

The bottleneck blocks appear similar to residual block where each block contains an input followed by several bottlenecks then followed by expansion [8]. However, inspired by the intuition that the bottlenecks actually contain all the necessary information, while an expansion layer acts merely as an implementation detail that accompanies a non-linear transformation of the tensor, we use shortcuts directly between the bottlenecks. Figure 3 provides a schematic visualization of the difference in the designs. The motivation for inserting shortcuts is similar to that of classical residual connections: we want to improve the ability of a gradient to propagate across multiplier layers. However, the inverted design is considerably more memory efficient (see Section 5 for details), as well as works slightly better in our experiments.

Figure 3

Figure 3: The difference between residual block [8, 30] and inverted residual. Diagonally hatched layers do not use non-linearities. We use thickness of each block to indicate its relative number of channels. Note how classical residuals connects the layers with high number of channels, whereas the inverted residuals connect the bottlenecks. Best viewed in color.

3.3. 倒置残差


Figure 3


Running time and parameter count for bottleneck convolution The basic implementation structure is illustrated in Table 1. For a block of size $h\times w$, expansion factor $t$ and kernel size $k$ with $d’$ input channels and $d’’$ output channels, the total number of multiply add required is $h \cdot w \cdot d’ \cdot t(d’ + k^2 + d’’)$. Compared with (1) this expression has an extra term, as indeed we have an extra 1 × 1 convolution, however the nature of our networks allows us to utilize much smaller input and output dimensions. In Table 3 we compare the needed sizes for each resolution between MobileNetV1, MobileNetV2 and ShuffleNet.

Table 1

Table 1: Bottleneck residual block transforming from $k$ to $k’$ channels, with stride $s$, and expansion factor $t$.

Table 3

Table 3: The max number of channels/memory (in Kb) that needs to be materialized at each spatial resolution for different architectures. We assume 16-bit floats for activations. For ShuffleNet, we use $2x, g = 3$ that matches the performance of MobileNetV1 and MobileNetV2. For the first layer of MobileNetV2 and ShuffleNet we can employ the trick described in Section 5 to reduce memory requirement. Even though ShuffleNet employs bottlenecks elsewhere, the non-bottleneck tensors still need to be materialized due to the presence of shortcuts between non-bottleneck tensors.

瓶颈卷积的运行时间和参数计数基本实现结构如表1所示。对于大小为$h\times w$的块,扩展因子为$t$,内核大小为$k$,具有$d’$维输入通道和$d’’$维输出通道,所需的乘法加法总数为$h \cdot w \cdot d’ \cdot t(d’ + k^2 + d’’)$。与(1)相比,这个表达式有一个额外项,因为实际上我们有一个额外的1×1卷积,但是我们的网络性质使我们能够利用更小的输入和输出维度。在表3中,我们比较了MobileNetV1,MobileNetV2和ShuffleNet之间每种分辨率所需的尺寸。

Table 1


Table 3

表3:不同架构中需要在每个空间分辨率上实现的最大通道数/内存(以Kb为单位)。我们假设激活使用16位浮点数。对于ShuffleNet,我们使用与MobileNetV1和MobileNetV2的性能相匹配的$2x,g = 3 $。对于MobileNetV2和ShuffleNet的第一层,我们可以采用第5节中描述的技巧来降低内存需求。尽管ShuffleNet在其它地方使用了瓶颈,但由于存在非瓶颈张量之间的快捷连接,因此非瓶颈张量仍然需要实现。

3.4. Information flow interpretation

One interesting property of our architecture is that it provides a natural separation between the input/output domains of the building blocks (bottleneck layers), and the layer transformation – that is a non-linear function that converts input to the output. The former can be seen as the capacity of the network at each layer, whereas the latter as the expressiveness. This is in contrast with traditional convolutional blocks, both regular and separable, where both expressiveness and capacity are tangled together and are functions of the output layer depth.

3.4. 信息流解释


In particular, in our case, when inner layer depth is 0 the underlying convolution is the identity function thanks to the shortcut connection. When the expansion ratio is smaller than 1, this is a classical residual convolutional block [8, 30]. However, for our purposes we show that expansion ratio greater than 1 is the most useful.


This interpretation allows us to study the expressiveness of the network separately from its capacity and we believe that further exploration of this separation is warranted to provide a better understanding of the network properties.


4. Model Architecture

Now we describe our architecture in detail. As discussed in the previous section the basic building block is a bottleneck depth-separable convolution with residuals. The detailed structure of this block is shown in Table 1. The architecture of MobileNetV2 contains the initial fully convolution layer with 32 filters, followed by 19 residual bottleneck layers described in the Table 2. We use ReLU6 as the non-linearity because of its robustness when used with low-precision computation [27]. We always use kernel size 3 × 3 as is standard for modern networks, and utilize dropout and batch normalization during training.

Table 2

Table 2: MobileNetV2 : Each line describes a sequence of 1 or more identical (modulo stride) layers, repeated $n$ times. All layers in the same sequence have the same number $c$ of output channels. The first layer of each sequence has a stride $s$ and all others use stride $1$. All spatial convolutions use 3 × 3 kernels. The expansion factor $t$ is always applied to the input size as described in Table 1.

4. 模型架构


Table 2


With the exception of the first layer, we use constant expansion rate throughout the network. In our experiments we find that expansion rates between 5 and 10 result in nearly identical performance curves, with smaller networks being better off with slightly smaller expansion rates and larger networks having slightly better performance with larger expansion rates.


For all our main experiments we use expansion factor of $6$ applied to the size of the input tensor. For example, for a bottleneck layer that takes $64$-channel input tensor and produces a tensor with $128$ channels, the intermediate expansion layer is then $64 · 6 = 384$ channels.

对于我们所有的主要实验,我们使用扩展因子$6$来应用于输入张量的大小。例如,对于瓶颈层采用$64$通道的输入张量并产生具有$128$通道的张量,中间扩展层则具有$64·6 =384$个通道。

As in [27] we tailor our architecture to different performance points, by using the input image resolution and width multiplier as tunable hyperparameters, that can be adjusted depending on desired accuracy/performance trade-offs. Our primary network (width multiplier 1, 224 × 224), has a computational cost of 300 million multiply-adds and uses 3.4 million parameters. We explore the performance trade offs, for input resolutions from 96 to 224, and width multipliers of 0.35 to 1.4. The network computational cost ranges from 7 multiply adds to 585M MAdds, while the model size vary between 1.7M and 6.9M parameters.

和[27]一样,我们通过使用输入图像分辨率和宽度倍数作为可调超参数来调整我们的架构以适应不同的性能点,可以根据所需的精度/性能权衡来调整。我们的主要网络(宽度乘数1,224×224)的计算成本为3亿次乘法,并使用了340万个参数。我们研究了性能权衡,输入分辨率从96到224,宽度乘数从0.35到1.4。网络计算成本范围从7次乘法增加到585M MAdds,而模型大小在1.7M个参数和6.9M个参数之间变化。

One minor implementation difference, with [27] is that for multipliers less than one, we apply width multiplier to all layers except the very last convolutional layer. This improves performance for smaller models.


5. Implementation Notes

5.1. Memory efficient inference

The inverted residual bottleneck layers allow a particularly memory efficient implementation which is very important for mobile applications. A standard efficient implementation of inference that uses for instance TensorFlow[31] or Caffe [32], builds a directed acyclic compute hypergraph $G$, consisting of edges representing the operations and nodes representing tensors of intermediate computation. The computation is scheduled in order to minimize the total number of tensors that needs to be stored in memory. In the most general case, it searches over all plausible computation orders $\Sigma (G)$ and picks the one that minimizes $$ M(G) = \min_{\pi\in \Sigma(G)} \max_{i \in 1..n} \left[\sum_{A \in R(i, \pi, G)} |A|\right] + \text{size}(\pi_i). $$ where $R(i, \pi, G)$ is the list of intermediate tensors that are connected to any of $\pi_{i}\dots \pi_{n}$ nodes, $|A|$ represents the size of the tensor $A$ and $size(i)$ is the total amount of memory needed for internal storage during operation $i$.

5. 实现说明

5.1. 内存有效推断

倒置的残差颈层允许特定地内存有效的实现,这对于移动应用非常重要。使用TensorFlow[31]或Caffe[32]等标准高效的推断实现,构建了一个有向无环计算超图$G$,由表示操作的边和代表中间计算张量的节点组成。预定计算是为了最小化需要存储在内存中的张量总数。在最一般的情况下,它会搜索所有合理的计算顺序$\Sigma (G)$,并选择最小化$$ M(G) = \min_{\pi\in \Sigma(G)} \max_{i \in 1..n} \left[\sum_{A \in R(i, \pi, G)} |A|\right] + \text{size}(\pi_i)$$。$$其中$R(i, \pi, G)$是连接到任何$\pi_{i}\dots \pi_{n}$节点的中间张量列表,$|A|$表示张量$A$的大小,$size(i)$是操作$i$期间内部存储所需的总内存量。

For graphs that have only trivial parallel structure (such as residual connection), there is only one non-trivial feasible computation order, and thus the total amount and a bound on the memory needed for inference on compute graph $G$ can be simplified: $$M(G) = \max_{op \in G} \left[\sum_{A \in \text{op}_{inp}} |A| + \sum_{B \in \text{op}_{out}} |B| + |op|\right] \tag {2}$$ Or to restate, the amount of memory is simply the maximum total size of combined inputs and outputs across all operations. In what follows we show that if we treat a bottleneck residual block as a single operation (and treat inner convolution as a disposable tensor), the total amount of memory would be dominated by the size of bottleneck tensors, rather than the size of tensors that are internal to bottleneck (and much larger).

对于仅具有平凡并行结构(例如残差连接)的图,只有一个非平凡的可行计算顺序,因此可以简化计算图$G$推断所需的内存总量和界限:$$M(G) = \max_{op \in G} \left[\sum_{A \in \text{op}_{inp}} |A| + \sum_{B \in \text{op}_{out}} |B| + |op|\right] \tag {2}$$或者重申,内存量只是在所有操作中组合输入和输出的最大总大小。在下文中我们将展示如果我们将瓶颈残差块视为单一操作(并将内部卷积视为一次性张量),则总内存量将由瓶颈张量的大小决定,而不是瓶颈的内部张量的大小(更大)。

Bottleneck Residual Block $\mathcal{F}(x)$ shown in Figure 3b can be expressed as a composition of three operators $\mathcal{F}(x) = [A \circ \mathcal{N} \circ B] x$, where $A$ is a linear transformation $A:\mathcal{R}^{s \times s \times k} \rightarrow \mathcal{R}^{s \times s \times n}$, $\mathcal{N}$ is a non-linear per-channel transformation: $\mathcal{N}: \mathcal{R}^{s \times s \times n} \rightarrow \mathcal{R}^{s’ \times s’ \times n}$, and $B$ is again a linear transformation to the output domain: $B: \mathcal{R}^{s’ \times s’ \times n} \rightarrow \mathcal{R}^{s’ \times s’ \times k’}$.

瓶颈残差块 图3b中所示的$\mathcal{F}(x)$可以表示为三个运算符的组合$\mathcal{F}(x) = [A \circ \mathcal{N} \circ B] x$,其中$A$是线性变换$A:\mathcal{R}^{s \times s \times k} \rightarrow \mathcal{R}^{s \times s \times n}$,$\mathcal{N}$是一个非线性的每个通道的转换:$\mathcal{N}: \mathcal{R}^{s \times s \times n} \rightarrow \mathcal{R}^{s’ \times s’ \times n}$,$B$是输出域的线性转换:$B: \mathcal{R}^{s’ \times s’ \times n} \rightarrow \mathcal{R}^{s’ \times s’ \times k’}$。

For our networks $\mathcal{N} = ReLU6 \circ dwise \circ ReLU6$, but the results apply to any per-channel transformation. Suppose the size of the input domain is $|x|$ and the size of the output domain is $|y|$, then the memory required to compute $F(X)$ can be as low as $|s^2 k| + |s’^2 k’| + O(\max(s^2, s’^2))$.

对于我们的网络$\mathcal{N} = ReLU6 \circ dwise \circ ReLU6$,但结果适用于任何的按通道转换。假设输入域的大小是$|x|$并且输出域的大小是$|y|$,那么计算$F(X)$所需的内存可以低至$|s^2 k| + |s’^2 k’| + O(\max(s^2, s’^2))$。

The algorithm is based on the fact that the inner tensor $\cal I$ can be represented as concatenation of $t$ tensors, of size $n/t$ each and our function can then be represented as $$\mathcal{F}(x) = \sum_{i=1}^t (A_i \circ N \circ B_i)(x)$$ by accumulating the sum, we only require one intermediate block of size $n/t$ to be kept in memory at all times. Using $n=t$ we end up having to keep only a single channel of the intermediate representation at all times. The two constraints that enabled us to use this trick is (a) the fact that the inner transformation (which includes non-linearity and depthwise) is per-channel, and (b) the consecutive non-per-channel operators have significant ratio of the input size to the output. For most of the traditional neural networks, such trick would not produce a significant improvement.

该算法基于以下事实:内部张量$\cal I$可以表示为$t$张量的连接,每个大小为$n/t$,则我们的函数可以表示为$$\mathcal{F}(x) = \sum_{i=1}^t (A_i \circ N \circ B_i)(x)$$通过累加和,我们只需要将一个大小为$n/t$的中间块始终保留在内存中。使用$n=t$,我们最终只需要保留中间表示的单个通道。使我们能够使用这一技巧的两个约束是(a)内部变换(包括非线性和深度)是每个通道的事实,以及(b)连续的非按通道运算符具有显著的输入输出大小比。对于大多数传统的神经网络,这种技巧不会产生显著的改善。

We note that, the number of multiply-adds operators needed to compute $F(X)$ using $t$-way split is independent of $t$, however in existing implementations we find that replacing one matrix multiplication with several smaller ones hurts runtime performance due to increased cache misses. We find that this approach is the most helpful to be used with $t$ being a small constant between $2$ and $5$. It significantly reduces the memory requirement, but still allows one to utilize most of the efficiencies gained by using highly optimized matrix multiplication and convolution operators provided by deep learning frameworks. It remains to be seen if special framework level optimization may lead to further runtime improvements.

我们注意到,使用$t$路分割计算$F(X)$所需的乘加运算符的数目是独立于$t$的,但在现有实现中,我们发现由于增加的缓存未命中,用几个较小的矩阵乘法替换一个矩阵乘法会很损坏运行时的性能 。我们发现这种方法最有用,$t$是$2$和$5$之间的一个小常数。它显著降低了内存需求,但仍然可以利用深度学习框架提供的高度优化的矩阵乘法和卷积算子来获得的大部分效率。如果特殊的框架级优化可能导致进一步的运行时改进,这个方法还有待观察。

6. Experiments

6.1. ImageNet Classification

Training setup We train our models using TensorFlow[31]. We use the standard RMSPropOptimizer with both decay and momentum set to 0.9. We use batch normalization after every layer, and the standard weight decay is set to 0.00004. Following MobileNetV1[27] setup we use initial learning rate of 0.045, and learning rate decay rate of 0.98 per epoch. We use 16 GPU asynchronous workers, and a batch size of 96.

6. 实验

6.1. ImageNet分类

训练设置我们使用TensorFlow[31]训练我们的模型。我们使用标准的RMSPropOptimizer,将衰减和动量都设置为0.9。我们在每层之后使用批标准化,并将标准权重衰减设置为0.00004。遵循MobileNetV1 [27]的设置,我们使用初始学习率为0.045,学习率的衰减比率为每个迭代周期衰减0.98。我们使用16个GPU异步,批大小为96。

Results We compare our networks against MobileNetV1, ShuffleNet and NASNet-A models. The statistics of a few selected models is shown in Table 4 with the full performance graph shown in Figure 5.

Table 4

Table 4: Performance on ImageNet, comparison for different networks. As is common practice for ops, we count the total number of Multiply-Adds. In the last column we report running time in milliseconds (ms) for a single large core of the Google Pixel 1 phone (using TF-Lite). We do not report ShuffleNet numbers as efficient group convolutions and shuffling are not yet supported.

Figure 5

Figure 5: Performance curve of MobileNetV2 vs MobileNetV1, ShuffleNet, NAS. For our networks we use multipliers 0.35, 0.5, 0.75, 1.0 for all resolutions, and additional 1.4 for for 224. Best viewed in color.


Table 4

表4:比较不同网络在ImageNet上的性能。正如ops的常见做法一样,我们计算Multiply-Adds的总数。在最后一列中,我们报告了Google Pixel 1手机上的一个大型核心(使用TF-Lite)的运行时间,以毫秒(ms)为单位。我们不报告ShuffleNet的数字,因为高效的群组卷积和混排尚未支持。

Figure 5


6.2. Object Detection

We evaluate and compare the performance of MobileNetV2 and MobileNetV1 as feature extractors [33] for object detection with a modified version of the Single Shot Detector (SSD) [34] on COCO dataset [2]. We also compare to YOLOv2 [35] and original SSD (with VGG-16 [6] as base network) as baselines. We do not compare performance with other architectures such as Faster-RCNN [36] and RFCN [37] since our focus is on mobile/real-time models.

6.2. 目标检测

我们评估和比较了MobileNetV2和MobileNetV1的性能,MobileNetV1使用COCO数据集[2]上Single Shot Detector(SSD)[34]的修改版本作为目标检测的特征提取器[33]。我们还将YOLOv2[35]和原始SSD(以VGG-16[6]为基础网络)作为基准进行比较。由于我们专注于移动/实时模型,因此我们不会比较Faster-RCNN[36]和RFCN[37]等其它架构的性能。

SSDLite: In this paper, we introduce a mobile friendly variant of regular SSD. We replace all the regular convolutions with separable convolutions (depthwise followed by $1 \times 1$ projection) in SSD prediction layers. This design is in line with the overall design of MobileNets and is seen to be much more computationally efficient. We call this modified version SSDLite. Compared to regular SSD, SSDLite dramatically reduces both parameter count and computational cost as shown in Table 5.

Table 5

Table 5: Comparison of the size and the computational cost between SSD and SSDLite configured with MobileNetV2 and making predictions for 80 classes.

SSDLite 在本文中,我们将介绍常规SSD的移动友好型变种。我们在SSD预测层中用可分离卷积(深度方向后接$1\times 1$投影)替换所有常规卷积。这种设计符合MobileNets的整体设计,并且在计算上效率更高。我们称之为修改版本的SSDLite。与常规SSD相比,SSDLite显著降低了参数计数和计算成本,如表5所示。

Table 5


For MobileNetV1, we follow the setup in [33]. For MobileNetV2, the first layer of SSDLite is attached to the expansion of layer 15 (with output stride of 16). The second and the rest of SSDLite layers are attached on top of the last layer (with output stride of 32). This setup is consistent with MobileNetV1 as all layers are attached to the feature map of the same output strides.


Both MobileNet models are trained and evaluated with Open Source TensorFlow Object Detection API [38]. The input resolution of both models is $320 \times 320$. We benchmark and compare both mAP (COCO challenge metrics), number of parameters and number of Multiply-Adds. The results are shown in Table 6. MobileNetV2 SSDLite is not only the most efficient model, but also the most accurate of the three. Notably, MobileNetV2 SSDLite is $20\times$ more efficient and $10\times$ smaller while still outperforms YOLOv2 on COCO dataset.

Table 6

Table 6: Performance comparison of MobileNetV2 + SSDLite and other realtime detectors on the COCO dataset object detection task. MobileNetV2 + SSDLite achieves competitive accuracy with significantly fewer parameters and smaller computational complexity. All models are trained on trainval35k and evaluated on test-dev. SSD/YOLOv2 numbers are from [35]. The running time is reported for the large core of the Google Pixel 1 phone, using an internal version of the TF-Lite engine.

MobileNet模型都经过了开源TensorFlow目标检测API的训练和评估[38]。 两个模型的输入分辨率为$320 \times 320$。我们进行了基准测试并比较了mAP(COCO挑战度量标准),参数数量和Multiply-Adds数量。结果如表6所示。MobileNetV2 SSDLite不仅是最高效的模型,而且也是三者中最准确的模型。值得注意的是,MobileNetV2 SSDLite效率高20倍,模型要小10倍,但仍优于COCO数据集上的YOLOv2。

Table 6

表6:MobileNetV2+SSDLite和其他实时检测器在COCO数据集目标检测任务中的性能比较。MobileNetV2+SSDLite以更少的参数和更小的计算复杂性实现了具有竞争力的精度。所有模型都在trainval35k上进行训练,并在test-dev上进行评估。SSD/YOLOv2的数字来自于[35]。使用内部版本的TF-Lite引擎,报告了在Google Pixel 1手机的大型核心上的运行时间。

6.3. Semantic Segmentation

In this section, we compare MobileNetV1 and MobileNetV2 models used as feature extractors with DeepLabv3 [39] for the task of mobile semantic segmentation. DeepLabv3 adopts atrous convolution [40, 41, 42], a powerful tool to explicitly control the resolution of computed feature maps, and builds five parallel heads including (a) Atrous Spatial Pyramid Pooling module (ASPP) [43] containing three $3 \times 3$ convolutions with different atrous rates, (b) $1\times 1$ convolution head, and (c) Image-level features [44]. We denote by output stride the ratio of input image spatial resolution to final output resolution, which is controlled by applying the atrous convolution properly. For semantic segmentation, we usually employ output $stride = 16$ or $8$ for denser feature maps. We conduct the experiments on the PASCAL VOC 2012 dataset [3], with extra annotated images from [45] and evaluation metric mIOU.

6.3. 语义分割

在本节中,我们使用MobileNetV1和MobileNetV2模型作为特征提取器与DeepLabv3[39]在移动语义分割任务上进行比较。DeepLabv3采用了空洞卷积[40,41,42],这是一种显式控制计算特征映射分辨率的强大工具,并构建了五个平行头部,包括(a)包含三个具有不同空洞率的$3 \times 3$卷积的Atrous Spatial Pyramid Pooling模块(ASPP)[43],(b)$1 \times 1$卷积头部,以及(c)图像级特征[44]。我们用输出步长来表示输入图像空间分辨率与最终输出分辨率的比值,该分辨率通过适当地应用空洞卷积来控制。对于语义分割,我们通常使用输出$stride = 16$或$8$来获取更密集的特征映射。我们在PASCAL VOC 2012数据集[3]上进行了实验,使用[45]中的额外标注图像和评估指标mIOU。

To build a mobile model, we experimented with three design variations: (1) different feature extractors, (2) simplifying the DeepLabv3 heads for faster computation, and (3) different inference strategies for boosting the performance. Our results are summarized in Table 7. We have observed that: (a) the inference strategies, including multi-scale inputs and adding left-right flipped images, significantly increase the MAdds and thus are not suitable for on-device applications, (b) using output stride = 16 is more efficient than output stride = 8, (c) MobileNetV1 is already a powerful feature extractor and only requires about 4.9 - 5.7 times fewer MAdds than ResNet-101 [8] (e.g., mIOU: 78.56 vs 82.70, and MAdds: 941.9B vs 4870.6B), (d) it is more efficient to build DeepLabv3 heads on top of the second last feature map of MobileNetV2 than on the original last-layer feature map, since the second to last feature map contains 320 channels instead of 1280, and by doing so, we attain similar performance, but require about 2.5 times fewer operations than the MobileNetV1 counterparts, and (e) DeepLabv3 heads are computationally expensive and removing the ASPP module significantly reduces the MAdds with only a slight performance degradation. In the end of the Table 7, we identify a potential candidate for on-device applications (in bold face), which attains $75.32\%$ mIOU and only requires 2.75B MAdds.

Table 7

Table 7: MobileNet + DeepLabv3 inference strategy on the PASCAL VOC 2012 validation set. MNet V2*: Second last feature map is used for DeepLabv3 heads, which includes (1) Atrous Spatial Pyramid Pooling (ASPP) module, and (2) $1 \times 1$ convolution as well as image-pooling feature. OS: output stride that controls the output resolution of the segmentation map. MF: Multi-scale and left-right flipped inputs during test. All of the models have been pretrained on COCO. The potential candidate for on-device applications is shown in bold face. PASCAL images have dimension $512 \times 512$ and atrous convolution allows us to control output feature resolution without increasing the number of parameters.

为了构建移动模型,我们尝试了三种设计变体:(1)不同的特征提取器,(2)简化DeepLabv3头部以加快计算速度,以及(3)提高性能的不同推断策略。我们的结果总结在表7中。我们已经观察到:(a)包括多尺度输入和添加左右翻转图像的推断策略显著增加了MAdds,因此不适合于在设备上应用,(b)使用输出步长16比使用输出步长8更有效率,(c)MobileNetV1已经是一个强大的特征提取器,并且只需要比ResNet-101少约4.9-5.7倍的MAdd[8](例如,mIOU:78.56与82.70和MAdds:941.9B vs 4870.6B),(d)在MobileNetV2的倒数第二个特征映射的顶部构建DeepLabv3头部比在原始的最后一个特征映射上更高效,因为倒数第二个特征映射包含320个通道而不是1280个通道,这样我们就可以达到类似的性能,但是要比MobileNetV1的通道少2.5倍,(e)DeepLabv3头部的计算成本很高,移除ASPP模块会显著减少MAdd并且只会稍微降低性能。在表7末尾,我们鉴定了一个设备上的潜在候选应用(粗体),该应用可以达到$75.32\%$mIOU并且只需要2.75B MAdds。

Table 7

表7:PASCAL VOC 2012验证集上的MobileNet+DeepLabv3推断策略。MNet V2*:用于DeepLabv3头部的倒数第二个特征映射,其中包括(1)Atrous Spatial Pyramid Pooling(ASPP)模块和(2)$1\times 1$卷积以及图像池化功能。OS:控制分割映射输出分辨率的输出步长。MF:测试期间多尺度和左右翻转输入。所有的模型都在COCO上进行预训练。设备上的潜在候选应用以粗体显示。PASCAL图像的尺寸为$ 512 \ times 512 $,而空洞卷积使得我们可以在不增加参数数量的情况下控制输出特征分辨率。

6.4. Ablation study

Inverted residual connections. The importance of residual connection has been studied extensively [8, 30, 46]. The new result reported in this paper is that the shortcut connecting bottleneck perform better than shortcuts connecting the expanded layers (see Figure 6b for comparison).

Figure 6

Figure 6: The impact of non-linearities and various types of shortcut (residual) connections.

6.4. 消融研究


Figure 6


Importance of linear bottlenecks. The linear bottleneck models are strictly less powerful than models with non-linearities, because the activations can always operate in linear regime with appropriate changes to biases and scaling. However our experiments shown in Figure 6a indicate that linear bottlenecks improve performance, providing support that non-linearity destroys information in low-dimensional space.


7. Conclusions and future work

We described a very simple network architecture that allowed us to build a family of highly efficient mobile models. Our basic building unit, has several properties that make it particularly suitable for mobile applications. It allows very memory-efficient inference and relies utilize standard operations present in all neural frameworks.

7. 总结及将来工作


For the ImageNet dataset, our architecture improves the state of the art for wide range of performance points. For object detection task, our network outperforms state-of-art realtime detectors on COCO dataset both in terms of accuracy and model complexity. Notably, our architecture combined with the SSDLite detection module is 20× less computation and 10× less parameters than YOLOv2.


On the theoretical side: the proposed convolutional block has a unique property that allows to separate the network expressiviness (encoded by expansion layers) from its capacity (encoded by bottleneck inputs). Exploring this is an important direction for future research.



We would like to thank Matt Streeter and Sergey Ioffe for their helpful feedback and discussion.


我们要感谢Matt Streeter和Sergey Ioffe的有益反馈和讨论。


[1] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. Int. J. Comput. Vision, 115(3):211–252, December 2015.

[2] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014.

[3] Mark Everingham, S. M. Ali Eslami, Luc Van Gool, Christopher K. I. Williams, John Winn, and Andrew Zisserma. The pascal visual object classes challenge a retrospective. IJCV, 2014.

[4] Mobilenetv2 source code. Available from

[5] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In Bartlett et al. [48], pages 1106–1114.

[6] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2014.

[7] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015, pages 1–9. IEEE Computer Society, 2015.

[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CoRR, abs/1512.03385, 2015.

[9] James Bergstra and Yoshua Bengio. Random search for hyper-parameter optimization. Journal of Machine Learning Research, 13:281–305, 2012.

[10] Jasper Snoek, Hugo Larochelle, and Ryan P. Adams. Practical bayesian optimization of machine learning algorithms. In Bartlett et al. [48], pages 2960–2968.

[11] Jasper Snoek, Oren Rippel, Kevin Swersky, Ryan Kiros, Nadathur Satish, Narayanan Sundaram, Md. Mostofa Ali Patwary, Prabhat, and Ryan P. Adams. Scalable bayesian optimization using deep neural networks. In Francis R. Bach and David M. Blei, editors, Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015, volume 37 of JMLR Workshop and Conference Proceedings, pages 2171–2180., 2015.

[12] Babak Hassibi and David G. Stork. Second order derivatives for network pruning: Optimal brain surgeon. In Stephen Jose Hanson, Jack D. Cowan, and C. Lee Giles, editors, Advances in Neural Information Processing Systems 5, [NIPS Conference, Denver, Colorado, USA, November 30 - December 3, 1992], pages 164–171. Morgan Kaufmann, 1992.

[13] Yann LeCun, John S. Denker, and Sara A. Solla. Optimal brain damage. In David S. Touretzky, editor, Advances in Neural Information Processing Systems 2, [NIPS Conference, Denver, Colorado, USA, November 27-30, 1989], pages 598–605. Morgan Kaufmann, 1989.

[14] Song Han, Jeff Pool, John Tran, and William J. Dally. Learning both weights and connections for efficient neural network. In Corinna Cortes, Neil D. Lawrence, Daniel D. Lee, Masashi Sugiyama, and Roman Garnett, editors, Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 1135–1143, 2015.

[15] Song Han, Jeff Pool, Sharan Narang, Huizi Mao, Shijian Tang, Erich Elsen, Bryan Catanzaro, John Tran, and William J. Dally. DSD: regularizing deep neural networks with dense-sparse-dense training flow. CoRR, abs/1607.04381, 2016.

[16] Yiwen Guo, Anbang Yao, and Yurong Chen. Dynamic network surgery for efficient dnns. In Daniel D. Lee, Masashi Sugiyama, Ulrike von Luxburg, Isabelle Guyon, and Roman Garnett, editors, Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain, pages 1379–1387, 2016.

[17] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. CoRR, abs/1608.08710, 2016.

[18] Karim Ahmed and Lorenzo Torresani. Connectivity learning in multi-branch networks. CoRR, abs/1709.09582, 2017.

[19] Tom Veniat and Ludovic Denoyer. Learning time-efficient deep architectures with budgeted super networks. CoRR, abs/1706.00046, 2017.

[20] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. CoRR, abs/1707.01083, 2017.

[21] Soravit Changpinyo, Mark Sandler, and Andrey Zhmoginov. The power of sparsity in convolutional neural networks. CoRR, abs/1702.06257, 2017.

[22] Min Wang, Baoyuan Liu, and Hassan Foroosh. Design of efficient convolutional layers using single intra-channel convolution, topological subdivisioning and spatial ”bottleneck” structure. CoRR, abs/1608.04337, 2016.

[23] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V. Le. Learning transferable architectures for scalable image recognition. CoRR, abs/1707.07012, 2017.

[24] Lingxi Xie and Alan L. Yuille. Genetic CNN. CoRR, abs/1703.01513, 2017.

[25] Esteban Real, Sherry Moore, Andrew Selle, Saurabh Saxena, Yutaka Leon Suematsu, Jie Tan, Quoc V. Le, and Alexey Kurakin. Large-scale evolution of image classifiers. In Doina Precup and Yee Whye Teh, editors, Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, volume 70 of Proceedings of Machine Learning Research, pages 2902–2911. PMLR, 2017.

[26] Barret Zoph and Quoc V. Le. Neural architecture search with reinforcement learning. CoRR, abs/1611.01578, 2016.

[27] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam.
Mobilenets: Efficient convolutional neural networks for mobile vision applications. CoRR, abs/1704.04861, 2017.

[28] Francois Chollet. Xception: Deep learning with depthwise separable convolutions. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017.

[29] Dongyoon Han, Jiwhan Kim, and Junmo Kim. Deep pyramidal residual networks. CoRR, abs/1610.02915, 2016.

[30] Saining Xie, Ross B. Girshick, Piotr Dolla ́r, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. CoRR, abs/1611.05431, 2016.

[31] Martın Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dan Mané, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Viégas, Oriol Vinyals, Pete Warden, Martin Wattenberg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. Software available from

[32] Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick, Sergio Guadarrama, and Trevor Darrell. Caffe: Convolutional architecture for fast feature embed- ding. arXiv preprint arXiv:1408.5093, 2014.

[33] Jonathan Huang, Vivek Rathod, Chen Sun, Men- glong Zhu, Anoop Korattikara, Alireza Fathi, Ian Fischer, Zbigniew Wojna, Yang Song, Sergio Guadarrama, et al. Speed/accuracy trade-offs for modern convolutional object detectors. In CVPR, 2017.

[34] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, 2016.

[35] Joseph Redmon and Ali Farhadi. Yolo9000: Better, faster, stronger. arXiv preprint arXiv:1612.08242, 2016.

[36] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in neural information processing systems, pages 91–99, 2015.

[37] Jifeng Dai, Yi Li, Kaiming He, and Jian Sun. R-fcn: Object detection via region-based fully convolutional networks. In Advances in neural information processing systems, pages 379–387, 2016.

[38] Jonathan Huang, Vivek Rathod, Derek Chow, Chen Sun, and Menglong Zhu. Tensorflow object detection api, 2017.

[39] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. CoRR, abs/1706.05587, 2017.

[40] Matthias Holschneider, Richard Kronland-Martinet, Jean Morlet, and Ph Tchamitchian. A real-time algorithm for signal analysis with the help of the wavelet transform. In Wavelets: Time-Frequency Methods and Phase Space, pages 289–297. 1989.

[41] Pierre Sermanet, David Eigen, Xiang Zhang, Michaël Mathieu, Rob Fergus, and Yann LeCun. Overfeat: Integrated recognition, localization and detection using convolutional networks. arXiv:1312.6229, 2013.

[42] George Papandreou,Iasonas Kokkinos, and Pierre Andre Savalle. Modeling local and global deformations in deep learning: Epitomic convolution, multiple instance learning, and sliding window detection. In CVPR, 2015.

[43] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. TPAMI, 2017.

[44] Wei Liu, Andrew Rabinovich, and Alexander C. Berg. Parsenet: Looking wider to see better. CoRR, abs/1506.04579, 2015.

[45] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In ICCV, 2011.

[46] Christian Szegedy, Sergey Ioffe, and Vincent Vanhoucke. Inception-v4, inception-resnet and the impact of residual connections on learning. CoRR, abs/1602.07261, 2016.

[47] Guido Montúfar, Razvan Pascanu, Kyunghyun Cho, and Yoshua Bengio. On the number of linear regions of deep neural networks. In Proceedings of the 27th International Conference on Neural Information Processing Systems, NIPS’14, pages 2924–2932, Cambridge, MA, USA, 2014. MIT Press.

[48] Peter L. Bartlett, Fernando C. N. Pereira, Christopher J. C. Burges, Léon Bottou, and Kilian Q. Weinberger, editors. Advances in Neural Information Processing Systems 25: 26th Annual Conference on Neural Information Processing Systems 2012. Proceedings of a meeting held December 3-6, 2012, Lake Tahoe, Nevada, United States, 2012.