Professional Documents
Culture Documents
Deep ResNet
Deep ResNet
Abstract
arXiv:1610.02915v4 [cs.CV] 6 Sep 2017
1
Figure 1. Schematic illustration of (a) basic residual units [7], (b) bottleneck residual units [7], (c) wide residual units [34], (d) our pyramidal
residual units, and (e) our pyramidal bottleneck residual units.
this phenomenon is related to the overall improvement in residual unit that can further improve ResNet. Section 3
the classification performance enabled by stochastic depth. closely analyzes our PyramidNets via several discussions.
Motivated by the ensemble interpretation of residual net- Section 4 presents experimental results and comparisons
works in Veit et al. [33] and the results with stochastic with several state-of-the-art deep network architectures.
depth [10], we devised another method to handle the phe- Section 5 concludes our paper with suggestions for future
nomenon associated with deleting the downsampling unit. works.
In the proposed method, the feature map dimensions are in-
creased at all layers to distribute the burden concentrated at 2. Network Architecture
locations of residual units affected by downsampling, such
that it is equally distributed across all units. It was found In this section, we introduce the network architectures of
that using the proposed new network architecture, deleting our PyramidNets. The major difference between Pyramid-
the units with downsampling does not degrade the perfor- Nets and other network architectures is that the dimension
mance significantly. In our paper, we refer to this network of channels gradually increases, instead of maintaining the
architecture as a deep “pyramidal” network and a “pyrami- dimension until a residual unit with downsampling appears.
dal” residual network with a residual-type network archi- A schematic illustration is shown in Figure 1 (d) to facilitate
tecture. This reflects the fact that the shape of the network understanding of our network architecture.
architecture can be compared to that of a pyramid. That is,
2.1. Feature Map Dimension Configuration
the number of channels gradually increases as a function
of the depth at which the layer occurs, which is similar to a Most deep CNN architectures [7, 8, 13, 25, 31, 35] uti-
pyramid structure of which the shape gradually widens from lize an approach whereby feature map dimensions are in-
the top downwards. This structure is illustrated in compar- creased by a large margin when the size of the feature map
ison to other network architectures in Figure 1. The key decreases, and feature map dimensions are not increased
contributions are summarized as follows: until they encounter a layer with downsampling. In the case
of the original ResNet for CIFAR datasets [12], the number
• A deep pyramidal residual network (PyramidNet) is in- of feature map dimensions Dk of the k-th residual unit that
troduced. The key idea is to concentrate on the feature belongs to the n-th group can be described as follows:
map dimension by increasing it gradually instead of by (
increasing it sharply at each residual unit with down- 16, if n(k) = 1,
sampling. In addition, our network architecture works Dk = (1)
16 · 2n(k)−2 , if n(k) ≥ 2,
as a mixture of both plain and residual networks by
using zero-padded identity-mapping shortcut connec-
in which n(k) ∈ {1, 2, 3, 4} denotes the index of the group
tions when increasing the feature map dimension.
to which the k-th residual unit belongs. The residual units
• A novel residual unit is also proposed, which can fur- that belong to the same group have an equal feature map
ther improve the performance of ResNet-based archi- size, and the n-th group contains Nn residual units. In the
tectures (compared with state-of-the-art network archi- first group, there is only one convolutional layer that con-
tectures). verts an RGB image into multiple feature maps. For the
n-th group, after Nn residual units have passed, the feature
The remainder of this paper is organized as follows. Sec- size is downsampled by half and the number of dimensions
tion 2 presents our PyramidNets and introduces a novel is doubled. We propose a method of increasing the feature
Group Output size Building Block
conv 1 32×32 [3 × 3, 16]
3 × 3, b16 + α(k − 1)/N c
conv 2 32×32 × N2
3 × 3, b16 + α(k − 1)/N c
3 × 3, b16 + α(k − 1)/N c
conv 3 16×16 × N3
3 × 3, b16 + α(k − 1)/N c
3 × 3, b16 + α(k − 1)/N c
conv 4 8×8 × N4
3 × 3, b16 + α(k − 1)/N c
avg pool 1×1 [8 × 8, 16 + α]
(a) (b) (c)
Table 1. Structure of our PyramidNet for benchmarking with
Figure 2. Visual illustrations of (a) additive PyramidNet, (b) mul- CIFAR-10 and CIFAR-100 datasets. α denotes the widening fac-
tiplicative PyramidNet, and (c) a comparison of (a) and (b). tor, and Nn signifies the number of blocks in a group. Downsam-
pling is performed at conv3 1 and conv4 1 with a stride of 2.
map dimension as follows:
( Figure 6, the layers can be stacked in various manners to
16, if k = 1,
Dk = (2) construct a single building block. We found the building
bDk−1 + α/N c, if 2 ≤ k ≤ N + 1, block shown in Figure 6 (d) to be the most promising, and
in which N denotes therefore we included this structure as building block in our
P4 the total number of residual units, de- PyramidNets. The discussion of this matter is continued in
fined as N = n=2 Nn . The dimension is increased by a
step factor of α/N , and the output dimension of the final the following section.
unit of each group becomes 16 + (n − 1)α/3 with same In terms of shortcut connections, many researchers ei-
number of residual units in each group. The details of our ther use those based on identity mapping, or those employ-
network architecture are presented in Table 1. ing convolution-based projection. However, as the feature
The above equations are based on an addition-based map dimension of PyramidNet is increased at every unit,
widening step factor α for increasing dimensions. However, we can only consider two options: zero-padded identity-
of course, multiplication-based widening (i.e., the process mapping shortcuts, and projection shortcuts conducted by
of multiplying by a factor to increase the channel dimen- 1×1 convolutions. However, as mentioned in the work of
sion geometrically) presents another possibility for creating He et al. [8], the 1×1 convolutional shortcut produces a
a pyramid-like structure. Then, eq.(2) can be transformed poor result when there are too many residual units, i.e., this
as follows: shortcut is unsuitable for very deep network architectures.
( Therefore, we select zero-padded identity-mapping short-
16, if k = 1, cuts for all residual units. Further discussions about the
Dk = 1 (3)
bDk−1 · α N c, if 2 ≤ k ≤ N + 1. zero-padded shortcut are provided in the following section.
(a) (b) tions and the number of ReLUs. This could be discussed
with original ResNets [7], for which it was shown that
Figure 5. Structure of residual unit (a) with zero-padded identity- the performance increases as the network becomes deeper;
mapping shortcut, (b) unraveled view of (a) showing that the zero- however, if the depth exceeds 1,000 layers, overfitting still
padded identity-mapping shortcut constitutes a mixture of a resid- occurs and the result is less accurate than that generated by
ual network with a shortcut connection and a plain network. shallower ResNets.
First, we note that using ReLUs after the addition of
dimensions of the k-th residual unit. From eq.(4), zero- residual units adversely affects performance:
padded elements of the identity-mapping shortcut for in-
xlk = ReLU (F(k ,l) (xlk −1 ) + xlk −1 ), (5)
creasing dimension let xlk contain the outputs of both resid-
ual networks and plain networks. Therefore, we could con- where the ReLUs seem to have the function of filtering non-
jecture that each zero-padded identity-mapping shortcut can negative elements. Gross and Wilber [5] found that simply
provide a mixture of the residual network and plain net- removing ReLUs from the original ResNet [7] after each
work, as shown in Figure 5. Furthermore, our Pyramid- addition with the shortcut connection leads to small perfor-
Net increases the channel dimension at every residual unit, mance improvements. This could be understood by consid-
and the mixture effect of the residual network and plain net- ering that, after addition, ReLUs provide non-negative in-
work increases markedly. Figure 4 supports the conclusion put to the subsequent residual units, and therefore the short-
that the test error of PyramidNet does not oscillate as much cut connection is always non-negative and the convolutional
as that of the pre-activation ResNet. Finally, we investigate layers would take responsibility for producing negative out-
several types of shortcuts including proposed zero-padded put before addition; this may decrease the overall capability
identity-mapping shortcut in Table 2. of the network architecture as analyzed in [8]. The pre-
3.3. A New Building Block activation ResNets proposed by He et al. [8] also overcame
this issue with pre-activated residual units that place BN
To maximize the capability of the network, it is natural layers and ReLUs before (instead of after) the convolutional
to ask the following question: “Can we design a better layers:
building block by altering the stacked elements inside
the building block in more principled way?”. The first xlk = F(k,l) (xlk−1 ) + xlk−1 , (6)
building block types were proposed in the original paper on
ResNets [7], and another type of building block was subse- where ReLUs are removed after addition to create an iden-
quently proposed in the paper on pre-activation ResNets [8], tity path. Consequently, the overall performance has in-
to answer the question. Moreover, pre-activation ResNets creased by a large margin without overfitting, even at depths
attempted to solve the backward gradient flowing problem exceeding 1,000 layers. Furthermore, Shen et al. [24] pro-
[8] by redesigning residual modules; this proved to be suc- posed a weighted residual network architecture, which lo-
cessful in trials. However, although the pre-activation resid- cates a ReLU inside a residual unit (instead of locating
ual unit was discovered with empirically improved perfor- ReLU after addition) to create an identity path, and showed
mance, further investigation over the possible combinations that this structure also does not overfit even at depths of
is not yet performed, leaving a potential room for improve- more than 1,000 layers.
ment. We next attempt to answer the question from two Second, we found that the use of a large number of Re-
points of view by considering Rectified Linear Units (Re- LUs in the blocks of each residual unit may negatively af-
LUs) [20] and Batch Normalization (BN) [11] layers. fect performance. Removing the first ReLU in the blocks
of each residual unit, as shown in Figure 6 (b) and (d), was
found to enhance performance compared with the blocks
3.3.1 ReLUs in a Building Block
shown in Figure 6 (a) and (c). Experimentally, we found
Including ReLUs [20] in the building blocks of residual that removal of the first ReLU in the stack is preferable and
units is essential for nonlinearity; however, we found empir- that the other ReLU should remain to ensure nonlinearity.
ically that the performance can vary depending on the loca- Removing the second ReLU in Figure 6 (a) changes the
(a) (b) (c) (d)
Figure 6. Various types of basic and bottleneck residual units. “BatchNorm” denotes a Batch Normalization (BN) layer. (a) original
pre-activation ResNets [8], (b) pre-activation ResNets removing the first ReLU, (c) pre-activation ResNets with a BN layer after the final
convolutional layer, and (d) pre-activation ResNets removing the fist ReLU with a BN layer after the final convolutional layer.
The main role of a BN layer is to normalize the activations of freedom obtained by involving γ and β from the BN lay-
for fast convergence and to improve performance. The ex- ers could improve the capability of the network architecture.
perimental results of the four structures provided in Table 3 The results in Table 3 support the conclusion that adding a
show that the BN layer can be used to maximize the capabil- BN layer at the end of each building block, as in type (c)
ity of a single residual unit. A BN layer conducts an affine and (d) in Figure 6, improves the performance. Note that
transformation with the following equation: the aforementioned network removing the first ReLU is also
improved by adding a BN layer after the final convolutional
y = γx + β, (7) layer. Furthermore, the results in Table 3 show that both
where γ and β are learned for every activation in feature PyramidNet and a new building block improve the perfor-
maps. We experimentally found that the learned γ and β mance significantly.
could closely approximate 0. This implies that if the learned
γ and β are both close to 0, then the corresponding acti- 4. Experimental Results
vation is considered not to be useful. Weighted ResNets
[24], in which the learnable weights occur at the end of We evaluate and compare the performance of our algo-
their building blocks, are also similarly learned to determine rithm with that of existing algorithms [7, 8, 18, 24, 34] using
whether the corresponding residual unit is useful. Thus, the representative benchmark datasets: CIFAR-10 and CIFAR-
BN layers at the end of each residual unit are a generalized 100 [12]. CIFAR-10 and CIFAR-100 each contain 32×32-
version including [24] to enable decisions to be made as to pixel color images, consists of 50,000 training images and
whether each residual unit is helpful. Therefore, the degrees 10,000 testing images. But in case of CIFAR-10, it includes
Network # of Params Output Feat. Dim. Depth Training Mem. CIFAR-10 CIFAR-100
NiN [18] - - - - 8.81 35.68
All-CNN [27] - - - - 7.25 33.71
DSN [17] - - - - 7.97 34.57
FitNet [21] - - - - 8.39 35.04
Highway [29] - - - - 7.72 32.39
Fractional Max-pooling [4] - - - - 4.50 27.62
ELU [29] - - - - 6.55 24.28
ResNet [7] 1.7M 64 110 547MB 6.43 25.16
ResNet [7] 10.2M 64 1001 2,921MB - 27.82
ResNet [7] 19.4M 64 1202 2,069MB 7.93 -
Pre-activation ResNet [8] 1.7M 64 164 841MB 5.46 24.33
Pre-activation ResNet [8] 10.2M 64 1001 2,921MB 4.62 22.71
Stochastic Depth [10] 1.7M 64 110 547MB 5.23 24.58
Stochastic Depth [10] 10.2M 64 1202 2,069MB 4.91 -
FractalNet [14] 38.6M 1,024 21 - 4.60 23.73
SwapOut v2 (width×4) [26] 7.4M 256 32 - 4.76 22.72
Wide ResNet (width×4) [34] 8.7M 256 40 775MB 4.97 22.89
Wide ResNet (width×10) [34] 36.5M 640 28 1,383MB 4.17 20.50
Weighted ResNet [24] 19.1M 64 1192 - 5.10 -
DenseNet (k = 24) [9] 27.2M 2,352 100 4,381MB 3.74 19.25
DenseNet-BC (k = 40) [9] 25.6M 2,190 190 7,247MB 3.46 17.18
PyramidNet (α = 48) 1.7M 64 110 655MB 4.58±0.06 23.12±0.04
PyramidNet (α = 84) 3.8M 100 110 781MB 4.26±0.23 20.66±0.40
PyramidNet (α = 270) 28.3M 286 110 1,437MB 3.73±0.04 18.25±0.10
PyramidNet (bottleneck, α = 270) 27.0M 1,144 164 4,169MB 3.48±0.20 17.01±0.39
PyramidNet (bottleneck, α = 240) 26.6M 1,024 200 4,451MB 3.44±0.11 16.51±0.13
PyramidNet (bottleneck, α = 220) 26.8M 944 236 4,767MB 3.40±0.07 16.37±0.29
PyramidNet (bottleneck, α = 200) 26.0M 864 272 5,005MB 3.31±0.08 16.35±0.24
Table 4. Top-1 error rates (%) on CIFAR datasets. All the results of PyramidNets are produced with additive PyramidNets, and α denotes
the widening factor. “Output Feat. Dim.” denotes the feature dimension of just before the last softmax classifier. The best results are
highlighted in red.
Table 5. Comparisons of single-model, single-crop error (%) on the ILSVRC 2012 validation set. All the results of PyramidNets are
produced with additive PyramidNets. “asp ratio” means the aspect ratio applied for data augmention, and “Output feat. dim.” denotes the
feature dimension of just after the last global pooling layer. ∗ denotes the models which applied dropout method, and † denotes the results
obtained from https://github.com/facebook/fb.resnet.torch.
chitectures do not have significant structural differences. We train our models for 120 epochs with a batch size of
As the number of parameters increases, they start to show 128, and the initial learning rate is set to 0.05, divided by 10
a more marked difference in terms of the feature map di- at 60, 90 and 105 epochs. We use the same weight decay,
mension configuration. Because the feature map dimension momentum, and initialization settings as those of CIFAR
increases linearly in the case of additive PyramidNets, the datasets. We train our model by using a standard data aug-
feature map dimensions of the input-side layers tend to be mentation with scale jittering and aspect ratio as suggested
larger, and those of the output-side layers tend to be smaller, in Szegedy et al. [31]. Table 5 shows the results of our
compared with multiplicative PyramidNets as illustrated in PyramidNets in ImageNet dataset compared with the state-
Figure 2. of-the-art models. The experimental results show that our
Previous works [7, 25] typically set multiplicative scal- PyramidNet with α = 300 has a top-1 error rate of 20.5%,
ing of feature map dimension for downsampling modules, which is 1.2% lower than the pre-activation ResNet-200 [8]
which is implemented to give a larger degree of freedom to which has a similar number of parameters but higher out-
the classification part by increasing the feature map dimen- put feature dimension than our model. We also notice that
sion of the output-side layers. However, for our Pyramid- increasing α with an appropriate regularization method can
Net, the results in Figure 7 implies that increasing the model further improve the performance.
capacity of the input-side layers would lead to a better per- For comparison with the Inception-ResNet [30] that uses
formance improvement than using a conventional way of a testing crop with 299 × 299 size, we test our model on a
multiplicative scaling of feature map dimension. 320 × 320 crop, by the same reason with the work of He et
We also note that, although the use of regularization al. [8]. Our PyramidNet with α = 300 shows a top-1 error
methods such as dropout [28] or stochastic depth [10] could rate of 19.6%, which outperforms both the pre-activation
further improve the performance of our model, we did not ResNet [8] and the Inception-ResNet-v2 [30] models.
involve those methods to ensure a fair comparison with
other models. 5. Conclusion
4.3. ImageNet The main idea of the novel deep network architecture
described in this paper involves increasing the feature map
1,000-class ImageNet dataset [22] used for ILSVRC dimension gradually, in order to construct so-called Pyra-
contains more than one million training images and 50,000 midNets along with the concept of ResNets. We also devel-
validation images. We use additive PyramidNets with the oped a novel residual unit, which includes a new building
pyramidal bottleneck residual units, deleting the first ReLU block for a residual unit with a zero-padded shortcut; this
and adding a BN layer at the last layer as described in Sec- design leads to significantly improved generalization abil-
tion 3.3 and shown in Figure 6 (d) for further performance ity. In tests using CIFAR-10, CIFAR-100, and ImageNet-
improvement. 1k datasets, our PyramidNets outperform all previous state-
of-the-art deep network architectures. Furthermore, the in- [16] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-
sights in this paper could be utilized by any network archi- based learning applied to document recognition. Proceed-
tecture, to improve their capacity for better performance. ings of the IEEE, 86(11):2278–2324, 1998. 1
In future work, we will develop methods of optimizing pa- [17] C.-Y. Lee, S. Xie, P. Gallagher, Z. Zhang, and Z. Tu. Deeply-
rameters such as feature map dimensions in more principled supervised nets. In AISTATS, 2015. 7
ways with proper cost functions that give insight into the na- [18] M. Lin, Q. Chen, and S. Yan. Network in network. In ICLR,
ture of residual networks. 2014. 6, 7
[19] J. Long, E. Shelhamer, and T. Darrell. Fully convolutional
Acknowledgements: This work was supported by the ICT
networks for semantic segmentation. In CVPR, 2015. 1
R&D program of MSIP/IITP, 2016-0-00563, Research on
[20] V. Nair and G. E. Hinton. Rectified linear units improve re-
Adaptive Machine Learning Technology Development for
stricted boltzmann machines. In ICML, 2010. 5
Intelligent Autonomous Digital Companion.
[21] A. Romero, N. Ballas, S. E. Kahou, A. Chassang, C. Gatta,
and Y. Bengio. Fitnets: Hints for thin deep nets. In ICLR,
References 2015. 7
[22] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh,
[1] R. Collobert, K. Kavukcuoglu, and C. Farabet. Torch7: A
S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein,
matlab-like environment for machine learning. In BigLearn,
A. C. Berg, and L. Fei-Fei. ImageNet Large Scale Visual
NIPS Workshop, 2011. 7
Recognition Challenge. International Journal of Computer
[2] J. Donahue, Y. Jia, O. Vinyals, J. Hoffman, N. Zhang, Vision, 115(3):211–252, 2015. 1, 8
E. Tzeng, and T. Darrell. Decaf: A deep convolutional acti- [23] P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus,
vation feature for generic visual recognition. In ICML, 2014. and Y. LeCun. Overfeat: Integrated recognition, localization
1 and detection using convolutional networks. In ICLR, 2014.
[3] R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich fea- 1
ture hierarchies for accurate object detection and semantic [24] F. Shen and G. Zeng. Weighted residuals for very deep net-
segmentation. In CVPR, 2014. 1 works. arXiv preprint arXiv:1605.08831, 2016. 5, 6, 7
[4] B. Graham. Fractional max-pooling. arXiv preprint [25] K. Simonyan and A. Zisserman. Very deep convolutional
arXiv:1412.6071, 2014. 7 networks for large-scale image recognition. In ICLR, 2015.
[5] S. Gross and M. Wilber. Training and investigating residual 1, 2, 3, 4, 8
nets. 2016. http://torch.ch/blog/2016/02/04/resnets.html. 5 [26] S. Singh, D. Hoiem, and D. Forsyth. Swapout: Learning an
[6] K. He, X. Zhang, S. Ren, and J. Sun. Delving deep into ensemble of deep architectures. In NIPS, 2016. 7
rectifiers: Surpassing human-level performance on imagenet [27] J. T. Springenberg, A. Dosovitskiy, T. Brox, and M. Ried-
classification. In ICCV, 2015. 7 miller. Striving for simplicity: The all convolutional net. In
[7] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning ICLR Workshop, 2015. 7
for image recognition. In CVPR, 2016. 1, 2, 3, 4, 5, 6, 7, 8 [28] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and
[8] K. He, X. Zhang, S. Ren, and J. Sun. Identity mappings in R. Salakhutdinov. Dropout: A simple way to prevent neu-
deep residual networks. In ECCV, 2016. 1, 2, 3, 4, 5, 6, 7, 8 ral networks from overfitting. Journal of Machine Learning
[9] G. Huang, Z. Liu, and K. Q. Weinberger. Densely connected Research, 15:1929–1958, 2014. 8
convolutional networks. arXiv preprint arXiv:1608.06993, [29] R. K. Srivastava, K. Greff, and J. Schmidhuber. Training
2016. 7 very deep networks. In NIPS, 2015. 1, 7
[10] G. Huang, Y. Sun, Z. Liu, D. Sedra, and K. Weinberger. Deep [30] C. Szegedy, S. Ioffe, and V. Vanhoucke. Inception-v4,
networks with stochastic depth. In ECCV, 2016. 1, 2, 4, 7, 8 inception-resnet and the impact of residual connections on
learning. In ICLR Workshop, 2016. 1, 8
[11] S. Ioffe and C. Szegedy. Batch normalization: Accelerating
deep network training by reducing internal covariate shift. In [31] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed,
ICML, 2015. 5 D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich.
Going deeper with convolutions. In CVPR, 2015. 1, 2, 8
[12] A. Krizhevsky. Learning multiple layers of features from
[32] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna.
tiny images. In Tech Report, 2009. 2, 6
Rethinking the inception architecture for computer vision. In
[13] A. Krizhevsky, I. Sutskever, and G. E. Hinton. ImageNet CVPR, 2016. 8
Classification with Deep Convolutional Neural Networks. In
[33] A. Veit, M. Wilber, and S. Belongie. Residual networks be-
NIPS, 2012. 1, 2
have like ensembles of relatively shallow networks. In NIPS,
[14] G. Larsson, M. Maire, and G. Shakhnarovich. Fractalnet: 2016. 1, 2, 3, 4
Ultra-deep neural networks without residuals. arXiv preprint [34] S. Zagoruyko and N. Komodakis. Wide residual networks.
arXiv:1605.07648, 2016. 7 In BMVC, 2016. 2, 6, 7, 8
[15] Y. LeCun, B. Boser, J. S. Denker, D. Henderson, R. E. [35] M. D. Zeiler and R. Fergus. Visualizing and understanding
Howard, W. Hubbard, and L. D. Jackel. Backpropagation convolutional networks. In ECCV, 2014. 1, 2
applied to handwritten zip code recognition. Neural compu-
tation, 1(4):541–551, 1989. 7