Densely Connected Convolutional Networks

1 . Densely Connected Convolutional Networks Gao Huang∗ Zhuang Liu∗ Laurens van der Maaten Cornell University Tsinghua University Facebook AI Research gh349@cornell.edu liuzhuang13@mails.tsinghua.edu.cn lvdmaaten@fb.com Kilian Q. Weinberger Cornell University arXiv:1608.06993v5 [cs.CV] 28 Jan 2018 kqw4@cornell.edu Abstract x0 H1 Recent work has shown that convolutional networks can x1 be substantially deeper, more accurate, and efficient to train if they contain shorter connections between layers close to H2 x2 the input and those close to the output. In this paper, we embrace this observation and introduce the Dense Convo- H3 lutional Network (DenseNet), which connects each layer x3 to every other layer in a feed-forward fashion. Whereas H4 traditional convolutional networks with L layers have L x4 connections—one between each layer and its subsequent layer—our network has L(L+1) 2 direct connections. For each layer, the feature-maps of all preceding layers are used as inputs, and its own feature-maps are used as inputs into all subsequent layers. DenseNets have several com- pelling advantages: they alleviate the vanishing-gradient Figure 1: A 5-layer dense block with a growth rate of k = 4. problem, strengthen feature propagation, encourage fea- Each layer takes all preceding feature-maps as input. ture reuse, and substantially reduce the number of parame- ters. We evaluate our proposed architecture on four highly competitive object recognition benchmark tasks (CIFAR-10, Networks [34] and Residual Networks (ResNets) [11] have CIFAR-100, SVHN, and ImageNet). DenseNets obtain sig- surpassed the 100-layer barrier. nificant improvements over the state-of-the-art on most of them, whilst requiring less computation to achieve high per- As CNNs become increasingly deep, a new research formance. Code and pre-trained models are available at problem emerges: as information about the input or gra- https://github.com/liuzhuang13/DenseNet. dient passes through many layers, it can vanish and “wash out” by the time it reaches the end (or beginning) of the network. Many recent publications address this or related problems. ResNets [11] and Highway Networks [34] by- 1. Introduction pass signal from one layer to the next via identity connec- Convolutional neural networks (CNNs) have become tions. Stochastic depth [13] shortens ResNets by randomly the dominant machine learning approach for visual object dropping layers during training to allow better information recognition. Although they were originally introduced over and gradient flow. FractalNets [17] repeatedly combine sev- 20 years ago [18], improvements in computer hardware and eral parallel layer sequences with different number of con- network structure have enabled the training of truly deep volutional blocks to obtain a large nominal depth, while CNNs only recently. The original LeNet5 [19] consisted of maintaining many short paths in the network. Although 5 layers, VGG featured 19 [29], and only last year Highway these different approaches vary in network topology and training procedure, they all share a key characteristic: they ∗ Authors contributed equally create short paths from early layers to later layers. 1 

2 . In this paper, we propose an architecture that distills this eters than existing algorithms with comparable accuracy. insight into a simple connectivity pattern: to ensure maxi- Further, we significantly outperform the current state-of- mum information flow between layers in the network, we the-art results on most of the benchmark tasks. connect all layers (with matching feature-map sizes) di- rectly with each other. To preserve the feed-forward nature, 2. Related Work each layer obtains additional inputs from all preceding lay- ers and passes on its own feature-maps to all subsequent The exploration of network architectures has been a part layers. Figure 1 illustrates this layout schematically. Cru- of neural network research since their initial discovery. The cially, in contrast to ResNets, we never combine features recent resurgence in popularity of neural networks has also through summation before they are passed into a layer; in- revived this research domain. The increasing number of lay- stead, we combine features by concatenating them. Hence, ers in modern networks amplifies the differences between the th layer has inputs, consisting of the feature-maps architectures and motivates the exploration of different con- of all preceding convolutional blocks. Its own feature-maps nectivity patterns and the revisiting of old research ideas. are passed on to all L− subsequent layers. This introduces A cascade structure similar to our proposed dense net- L(L+1) work layout has already been studied in the neural networks 2 connections in an L-layer network, instead of just literature in the 1980s [3]. Their pioneering work focuses on L, as in traditional architectures. Because of its dense con- fully connected multi-layer perceptrons trained in a layer- nectivity pattern, we refer to our approach as Dense Convo- by-layer fashion. More recently, fully connected cascade lutional Network (DenseNet). networks to be trained with batch gradient descent were A possibly counter-intuitive effect of this dense connec- proposed [40]. Although effective on small datasets, this tivity pattern is that it requires fewer parameters than tra- approach only scales to networks with a few hundred pa- ditional convolutional networks, as there is no need to re- rameters. In [9, 23, 31, 41], utilizing multi-level features learn redundant feature-maps. Traditional feed-forward ar- in CNNs through skip-connnections has been found to be chitectures can be viewed as algorithms with a state, which effective for various vision tasks. Parallel to our work, [1] is passed on from layer to layer. Each layer reads the state derived a purely theoretical framework for networks with from its preceding layer and writes to the subsequent layer. cross-layer connections similar to ours. It changes the state but also passes on information that needs Highway Networks [34] were amongst the first architec- to be preserved. ResNets [11] make this information preser- tures that provided a means to effectively train end-to-end vation explicit through additive identity transformations. networks with more than 100 layers. Using bypassing paths Recent variations of ResNets [13] show that many layers along with gating units, Highway Networks with hundreds contribute very little and can in fact be randomly dropped of layers can be optimized without difficulty. The bypass- during training. This makes the state of ResNets similar ing paths are presumed to be the key factor that eases the to (unrolled) recurrent neural networks [21], but the num- training of these very deep networks. This point is further ber of parameters of ResNets is substantially larger because supported by ResNets [11], in which pure identity mappings each layer has its own weights. Our proposed DenseNet ar- are used as bypassing paths. ResNets have achieved im- chitecture explicitly differentiates between information that pressive, record-breaking performance on many challeng- is added to the network and information that is preserved. ing image recognition, localization, and detection tasks, DenseNet layers are very narrow (e.g., 12 filters per layer), such as ImageNet and COCO object detection [11]. Re- adding only a small set of feature-maps to the “collective cently, stochastic depth was proposed as a way to success- knowledge” of the network and keep the remaining feature- fully train a 1202-layer ResNet [13]. Stochastic depth im- maps unchanged—and the final classifier makes a decision proves the training of deep residual networks by dropping based on all feature-maps in the network. layers randomly during training. This shows that not all Besides better parameter efficiency, one big advantage of layers may be needed and highlights that there is a great DenseNets is their improved flow of information and gra- amount of redundancy in deep (residual) networks. Our pa- dients throughout the network, which makes them easy to per was partly inspired by that observation. ResNets with train. Each layer has direct access to the gradients from the pre-activation also facilitate the training of state-of-the-art loss function and the original input signal, leading to an im- networks with > 1000 layers [12]. plicit deep supervision [20]. This helps training of deeper An orthogonal approach to making networks deeper network architectures. Further, we also observe that dense (e.g., with the help of skip connections) is to increase the connections have a regularizing effect, which reduces over- network width. The GoogLeNet [36, 37] uses an “Incep- fitting on tasks with smaller training set sizes. tion module” which concatenates feature-maps produced We evaluate DenseNets on four highly competitive by filters of different sizes. In [38], a variant of ResNets benchmark datasets (CIFAR-10, CIFAR-100, SVHN, and with wide generalized residual blocks was proposed. In ImageNet). Our models tend to require much fewer param- fact, simply increasing the number of filters in each layer of 

3 . Input Prediction C o Dense Block 1 C o P Dense Block 2 C o P Dense Block 3 P “horse” n n o n o o L i v v o v o o n o l o l l i o l l i l i e u t u t n u t n n a r i o i o g i o g g n n n Figure 2: A deep DenseNet with three dense blocks. The layers between two adjacent blocks are referred to as transition layers and change feature-map sizes via convolution and pooling. ResNets can improve its performance provided the depth is An advantage of ResNets is that the gradient can flow di- sufficient [42]. FractalNets also achieve competitive results rectly through the identity function from later layers to the on several datasets using a wide network structure [17]. earlier layers. However, the identity function and the output Instead of drawing representational power from ex- of H are combined by summation, which may impede the tremely deep or wide architectures, DenseNets exploit the information flow in the network. potential of the network through feature reuse, yielding con- Dense connectivity. To further improve the information densed models that are easy to train and highly parameter- flow between layers we propose a different connectivity efficient. Concatenating feature-maps learned by different pattern: we introduce direct connections from any layer layers increases variation in the input of subsequent layers to all subsequent layers. Figure 1 illustrates the layout of and improves efficiency. This constitutes a major difference the resulting DenseNet schematically. Consequently, the between DenseNets and ResNets. Compared to Inception th layer receives the feature-maps of all preceding layers, networks [36, 37], which also concatenate features from dif- x0 , . . . , x −1 , as input: ferent layers, DenseNets are simpler and more efficient. There are other notable network architecture innovations x = H ([x0 , x1 , . . . , x −1 ]), (2) which have yielded competitive results. The Network in where [x0 , x1 , . . . , x −1 ] refers to the concatenation of the Network (NIN) [22] structure includes micro multi-layer feature-maps produced in layers 0, . . . , − 1. Because of its perceptrons into the filters of convolutional layers to ex- dense connectivity we refer to this network architecture as tract more complicated features. In Deeply Supervised Net- Dense Convolutional Network (DenseNet). For ease of im- work (DSN) [20], internal layers are directly supervised plementation, we concatenate the multiple inputs of H (·) by auxiliary classifiers, which can strengthen the gradients in eq. (2) into a single tensor. received by earlier layers. Ladder Networks [27, 25] in- troduce lateral connections into autoencoders, producing Composite function. Motivated by [12], we define H (·) impressive accuracies on semi-supervised learning tasks. as a composite function of three consecutive operations: In [39], Deeply-Fused Nets (DFNs) were proposed to im- batch normalization (BN) [14], followed by a rectified lin- prove information flow by combining intermediate layers ear unit (ReLU) [6] and a 3 × 3 convolution (Conv). of different base networks. The augmentation of networks Pooling layers. The concatenation operation used in with pathways that minimize reconstruction losses was also Eq. (2) is not viable when the size of feature-maps changes. shown to improve image classification models [43]. However, an essential part of convolutional networks is down-sampling layers that change the size of feature-maps. 3. DenseNets To facilitate down-sampling in our architecture we divide Consider a single image x0 that is passed through a con- the network into multiple densely connected dense blocks; volutional network. The network comprises L layers, each see Figure 2. We refer to layers between blocks as transition of which implements a non-linear transformation H (·), layers, which do convolution and pooling. The transition where indexes the layer. H (·) can be a composite func- layers used in our experiments consist of a batch normal- tion of operations such as Batch Normalization (BN) [14], ization layer and an 1×1 convolutional layer followed by a rectified linear units (ReLU) [6], Pooling [19], or Convolu- 2×2 average pooling layer. tion (Conv). We denote the output of the th layer as x . Growth rate. If each function H produces k feature- ResNets. Traditional convolutional feed-forward net- maps, it follows that the th layer has k0 + k × ( − 1) input works connect the output of the th layer as input to the feature-maps, where k0 is the number of channels in the in- ( + 1)th layer [16], which gives rise to the following put layer. An important difference between DenseNet and layer transition: x = H (x −1 ). ResNets [11] add a existing network architectures is that DenseNet can have skip-connection that bypasses the non-linear transforma- very narrow layers, e.g., k = 12. We refer to the hyper- tions with an identity function: parameter k as the growth rate of the network. We show in x = H (x −1 ) + x −1 . (1) Section 4 that a relatively small growth rate is sufficient to 

4 . Layers Output Size DenseNet-121 DenseNet-169 DenseNet-201 DenseNet-264 Convolution 112 × 112 7 × 7 conv, stride 2 Pooling 56 × 56 3 × 3 max pool, stride 2 Dense Block 1 × 1 conv 1 × 1 conv 1 × 1 conv 1 × 1 conv 56 × 56 ×6 ×6 ×6 ×6 (1) 3 × 3 conv 3 × 3 conv 3 × 3 conv 3 × 3 conv Transition Layer 56 × 56 1 × 1 conv (1) 28 × 28 2 × 2 average pool, stride 2 Dense Block 1 × 1 conv 1 × 1 conv 1 × 1 conv 1 × 1 conv 28 × 28 × 12 × 12 × 12 × 12 (2) 3 × 3 conv 3 × 3 conv 3 × 3 conv 3 × 3 conv Transition Layer 28 × 28 1 × 1 conv (2) 14 × 14 2 × 2 average pool, stride 2 Dense Block 1 × 1 conv 1 × 1 conv 1 × 1 conv 1 × 1 conv 14 × 14 × 24 × 32 × 48 × 64 (3) 3 × 3 conv 3 × 3 conv 3 × 3 conv 3 × 3 conv Transition Layer 14 × 14 1 × 1 conv (3) 7×7 2 × 2 average pool, stride 2 Dense Block 1 × 1 conv 1 × 1 conv 1 × 1 conv 1 × 1 conv 7×7 × 16 × 32 × 32 × 48 (4) 3 × 3 conv 3 × 3 conv 3 × 3 conv 3 × 3 conv Classification 1×1 7 × 7 global average pool Layer 1000D fully-connected, softmax Table 1: DenseNet architectures for ImageNet. The growth rate for all the networks is k = 32. Note that each “conv” layer shown in the table corresponds the sequence BN-ReLU-Conv. obtain state-of-the-art results on the datasets that we tested Implementation Details. On all datasets except Ima- on. One explanation for this is that each layer has access geNet, the DenseNet used in our experiments has three to all the preceding feature-maps in its block and, therefore, dense blocks that each has an equal number of layers. Be- to the network’s “collective knowledge”. One can view the fore entering the first dense block, a convolution with 16 (or feature-maps as the global state of the network. Each layer twice the growth rate for DenseNet-BC) output channels is adds k feature-maps of its own to this state. The growth performed on the input images. For convolutional layers rate regulates how much new information each layer con- with kernel size 3×3, each side of the inputs is zero-padded tributes to the global state. The global state, once written, by one pixel to keep the feature-map size fixed. We use 1×1 can be accessed from everywhere within the network and, convolution followed by 2×2 average pooling as transition unlike in traditional network architectures, there is no need layers between two contiguous dense blocks. At the end of to replicate it from layer to layer. the last dense block, a global average pooling is performed and then a softmax classifier is attached. The feature-map Bottleneck layers. Although each layer only produces k sizes in the three dense blocks are 32× 32, 16×16, and output feature-maps, it typically has many more inputs. It 8×8, respectively. We experiment with the basic DenseNet has been noted in [37, 11] that a 1×1 convolution can be in- structure with configurations {L = 40, k = 12}, {L = troduced as bottleneck layer before each 3×3 convolution 100, k = 12} and {L = 100, k = 24}. For DenseNet- to reduce the number of input feature-maps, and thus to BC, the networks with configurations {L = 100, k = 12}, improve computational efficiency. We find this design es- {L = 250, k = 24} and {L = 190, k = 40} are evaluated. pecially effective for DenseNet and we refer to our network with such a bottleneck layer, i.e., to the BN-ReLU-Conv(1× In our experiments on ImageNet, we use a DenseNet-BC 1)-BN-ReLU-Conv(3×3) version of H , as DenseNet-B. In structure with 4 dense blocks on 224×224 input images. our experiments, we let each 1×1 convolution produce 4k The initial convolution layer comprises 2k convolutions of feature-maps. size 7×7 with stride 2; the number of feature-maps in all other layers also follow from setting k. The exact network Compression. To further improve model compactness, configurations we used on ImageNet are shown in Table 1. we can reduce the number of feature-maps at transition layers. If a dense block contains m feature-maps, we let the following transition layer generate θm output feature- maps, where 0 < θ ≤ 1 is referred to as the compression fac- 4. Experiments tor. When θ = 1, the number of feature-maps across transi- tion layers remains unchanged. We refer the DenseNet with We empirically demonstrate DenseNet’s effectiveness on θ < 1 as DenseNet-C, and we set θ = 0.5 in our experiment. several benchmark datasets and compare with state-of-the- When both the bottleneck and transition layers with θ < 1 art architectures, especially with ResNet and its variants. are used, we refer to our model as DenseNet-BC. 

5 . Method Depth Params C10 C10+ C100 C100+ SVHN Network in Network [22] - - 10.41 8.81 35.68 - 2.35 All-CNN [32] - - 9.08 7.25 - 33.71 - Deeply Supervised Net [20] - - 9.69 7.97 - 34.57 1.92 Highway Network [34] - - - 7.72 - 32.39 - FractalNet [17] 21 38.6M 10.18 5.22 35.34 23.30 2.01 with Dropout/Drop-path 21 38.6M 7.33 4.60 28.20 23.73 1.87 ResNet [11] 110 1.7M - 6.61 - - - ResNet (reported by [13]) 110 1.7M 13.63 6.41 44.74 27.22 2.01 ResNet with Stochastic Depth [13] 110 1.7M 11.66 5.23 37.80 24.58 1.75 1202 10.2M - 4.91 - - - Wide ResNet [42] 16 11.0M - 4.81 - 22.07 - 28 36.5M - 4.17 - 20.50 - with Dropout 16 2.7M - - - - 1.64 ResNet (pre-activation) [12] 164 1.7M 11.26∗ 5.46 35.58∗ 24.33 - 1001 10.2M 10.56∗ 4.62 33.47∗ 22.71 - DenseNet (k = 12) 40 1.0M 7.00 5.24 27.55 24.42 1.79 DenseNet (k = 12) 100 7.0M 5.77 4.10 23.79 20.20 1.67 DenseNet (k = 24) 100 27.2M 5.83 3.74 23.42 19.25 1.59 DenseNet-BC (k = 12) 100 0.8M 5.92 4.51 24.15 22.27 1.76 DenseNet-BC (k = 24) 250 15.3M 5.19 3.62 19.64 17.60 1.74 DenseNet-BC (k = 40) 190 25.6M - 3.46 - 17.18 - Table 2: Error rates (%) on CIFAR and SVHN datasets. k denotes network’s growth rate. Results that surpass all competing methods are bold and the overall best results are blue. “+” indicates standard data augmentation (translation and/or mirroring). ∗ indicates results run by ourselves. All the results of DenseNets without data augmentation (C10, C100, SVHN) are obtained using Dropout. DenseNets achieve lower error rates while using fewer parameters than ResNet. Without data augmentation, DenseNet performs better by a large margin. 4.1. Datasets ImageNet. The ILSVRC 2012 classification dataset [2] consists 1.2 million images for training, and 50,000 for val- CIFAR. The two CIFAR datasets [15] consist of colored idation, from 1, 000 classes. We adopt the same data aug- natural images with 32×32 pixels. CIFAR-10 (C10) con- mentation scheme for training images as in [8, 11, 12], and sists of images drawn from 10 and CIFAR-100 (C100) from apply a single-crop or 10-crop with size 224×224 at test 100 classes. The training and test sets contain 50,000 and time. Following [11, 12, 13], we report classification errors 10,000 images respectively, and we hold out 5,000 training on the validation set. images as a validation set. We adopt a standard data aug- mentation scheme (mirroring/shifting) that is widely used 4.2. Training for these two datasets [11, 13, 17, 22, 28, 20, 32, 34]. We denote this data augmentation scheme by a “+” mark at the All the networks are trained using stochastic gradient de- end of the dataset name (e.g., C10+). For preprocessing, scent (SGD). On CIFAR and SVHN we train using batch we normalize the data using the channel means and stan- size 64 for 300 and 40 epochs, respectively. The initial dard deviations. For the final run we use all 50,000 training learning rate is set to 0.1, and is divided by 10 at 50% and images and report the final test error at the end of training. 75% of the total number of training epochs. On ImageNet, we train models for 90 epochs with a batch size of 256. SVHN. The Street View House Numbers (SVHN) dataset The learning rate is set to 0.1 initially, and is lowered by [24] contains 32×32 colored digit images. There are 73,257 10 times at epoch 30 and 60. Note that a naive implemen- images in the training set, 26,032 images in the test set, and tation of DenseNet may contain memory inefficiencies. To 531,131 images for additional training. Following common reduce the memory consumption on GPUs, please refer to practice [7, 13, 20, 22, 30] we use all the training data with- our technical report on the memory-efficient implementa- out any data augmentation, and a validation set with 6,000 tion of DenseNets [26]. images is split from the training set. We select the model Following [8], we use a weight decay of 10−4 and a with the lowest validation error during training and report Nesterov momentum [35] of 0.9 without dampening. We the test error. We follow [42] and divide the pixel values by adopt the weight initialization introduced by [10]. For the 255 so they are in the [0, 1] range. three datasets without data augmentation, i.e., C10, C100 

6 . 27.5 27.5 ResNets ResNets ResNet−34 DenseNets−BC ResNet−34 DenseNets−BC 26.5 26.5 Model top-1 top-5 validation error (%) validation error (%) 25.5 25.5 DenseNet−121 DenseNet−121 DenseNet-121 25.02 / 23.61 7.71 / 6.66 24.5 ResNet−50 24.5 ResNet−50 DenseNet-169 23.80 / 22.08 6.85 / 5.92 DenseNet−169 DenseNet−169 23.5 23.5 DenseNet−201 ResNet−101 DenseNet−201 ResNet−101 DenseNet-201 22.58 / 21.46 6.34 / 5.54 22.5 ResNet−152 22.5 ResNet−152 DenseNet−264 DenseNet−264 DenseNet-264 22.15 / 20.80 6.12 / 5.29 21.5 21.5 0 1 2 3 4 5 6 7 8 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 #parameters 7 #flops 10 x 10 x 10 Table 3: The top-1 and top-5 error rates on the Figure 3: Comparison of the DenseNets and ResNets top-1 error rates (single-crop ImageNet validation set, with single-crop / 10- testing) on the ImageNet validation dataset as a function of learned parameters (left) crop testing. and FLOPs during test-time (right). and SVHN, we add a dropout layer [33] after each convolu- Parameter Efficiency. The results in Table 2 indicate that tional layer (except the first one) and set the dropout rate to DenseNets utilize parameters more efficiently than alterna- 0.2. The test errors were only evaluated once for each task tive architectures (in particular, ResNets). The DenseNet- and model setting. BC with bottleneck structure and dimension reduction at transition layers is particularly parameter-efficient. For ex- 4.3. Classification Results on CIFAR and SVHN ample, our 250-layer model only has 15.3M parameters, but We train DenseNets with different depths, L, and growth it consistently outperforms other models such as FractalNet rates, k. The main results on CIFAR and SVHN are shown and Wide ResNets that have more than 30M parameters. We in Table 2. To highlight general trends, we mark all results also highlight that DenseNet-BC with L = 100 and k = 12 that outperform the existing state-of-the-art in boldface and achieves comparable performance (e.g., 4.51% vs 4.62% er- the overall best result in blue. ror on C10+, 22.27% vs 22.71% error on C100+) as the 1001-layer pre-activation ResNet using 90% fewer parame- Accuracy. Possibly the most noticeable trend may orig- ters. Figure 4 (right panel) shows the training loss and test inate from the bottom row of Table 2, which shows that errors of these two networks on C10+. The 1001-layer deep DenseNet-BC with L = 190 and k = 40 outperforms ResNet converges to a lower training loss value but a similar the existing state-of-the-art consistently on all the CIFAR test error. We analyze this effect in more detail below. datasets. Its error rates of 3.46% on C10+ and 17.18% on Overfitting. One positive side-effect of the more efficient C100+ are significantly lower than the error rates achieved use of parameters is a tendency of DenseNets to be less by wide ResNet architecture [42]. Our best results on prone to overfitting. We observe that on the datasets without C10 and C100 (without data augmentation) are even more data augmentation, the improvements of DenseNet architec- encouraging: both are close to 30% lower than Fractal- tures over prior work are particularly pronounced. On C10, Net with drop-path regularization [17]. On SVHN, with the improvement denotes a 29% relative reduction in error dropout, the DenseNet with L = 100 and k = 24 also from 7.33% to 5.19%. On C100, the reduction is about 30% surpasses the current best result achieved by wide ResNet. from 28.20% to 19.64%. In our experiments, we observed However, the 250-layer DenseNet-BC doesn’t further im- potential overfitting in a single setting: on C10, a 4× growth prove the performance over its shorter counterpart. This of parameters produced by increasing k = 12 to k = 24 lead may be explained by that SVHN is a relatively easy task, to a modest increase in error from 5.77% to 5.83%. The and extremely deep models may overfit to the training set. DenseNet-BC bottleneck and compression layers appear to Capacity. Without compression or bottleneck layers, be an effective way to counter this trend. there is a general trend that DenseNets perform better as 4.4. Classification Results on ImageNet L and k increase. We attribute this primarily to the corre- sponding growth in model capacity. This is best demon- We evaluate DenseNet-BC with different depths and strated by the column of C10+ and C100+. On C10+, the growth rates on the ImageNet classification task, and com- error drops from 5.24% to 4.10% and finally to 3.74% as pare it with state-of-the-art ResNet architectures. To en- the number of parameters increases from 1.0M, over 7.0M sure a fair comparison between the two architectures, we to 27.2M. On C100+, we observe a similar trend. This sug- eliminate all other factors such as differences in data pre- gests that DenseNets can utilize the increased representa- processing and optimization settings by adopting the pub- tional power of bigger and deeper models. It also indicates licly available Torch implementation for ResNet by [8]1 . that they do not suffer from overfitting or the optimization 1 https://github.com/facebook/fb.resnet.torch difficulties of residual networks [11]. 

7 . 16 16 16 DenseNet ResNet Test error: ResNet-1001 (10.2M) 100 14 DenseNet-C 14 DenseNet-BC 14 Test error: DenseNet-BC-100 (0.8M) DenseNet-B Training loss: ResNet-1001 (10.2M) DenseNet-BC Training loss: DenseNet-BC-100 (0.8M) 12 12 12 test error (%) test error (%) test error (%) training loss 10−1 10 10 10 8 8 8 10−2 3x fewer parameters 6 6 6 4 4 4 10−3 0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8 0 50 100 150 200 250 300 #parameters ×105 #parameters ⇥105 epoch Figure 4: Left: Comparison of the parameter efficiency on C10+ between DenseNet variations. Middle: Comparison of the parameter efficiency between DenseNet-BC and (pre-activation) ResNets. DenseNet-BC requires about 1/3 of the parameters as ResNet to achieve comparable accuracy. Right: Training and testing curves of the 1001-layer pre-activation ResNet [12] with more than 10M parameters and a 100-layer DenseNet with only 0.8M parameters. We simply replace the ResNet model with the DenseNet- ResNet architecture (middle). We train multiple small net- BC network, and keep all the experiment settings exactly works with varying depths on C10+ and plot their test ac- the same as those used for ResNet. curacies as a function of network parameters. In com- We report the single-crop and 10-crop validation errors parison with other popular network architectures, such as of DenseNets on ImageNet in Table 3. Figure 3 shows AlexNet [16] or VGG-net [29], ResNets with pre-activation the single-crop top-1 validation errors of DenseNets and use fewer parameters while typically achieving better re- ResNets as a function of the number of parameters (left) and sults [12]. Hence, we compare DenseNet (k = 12) against FLOPs (right). The results presented in the figure reveal that this architecture. The training setting for DenseNet is kept DenseNets perform on par with the state-of-the-art ResNets, the same as in the previous section. whilst requiring significantly fewer parameters and compu- The graph shows that DenseNet-BC is consistently the tation to achieve comparable performance. For example, a most parameter efficient variant of DenseNet. Further, to DenseNet-201 with 20M parameters model yields similar achieve the same level of accuracy, DenseNet-BC only re- validation error as a 101-layer ResNet with more than 40M quires around 1/3 of the parameters of ResNets (middle parameters. Similar trends can be observed from the right plot). This result is in line with the results on ImageNet panel, which plots the validation error as a function of the we presented in Figure 3. The right plot in Figure 4 shows number of FLOPs: a DenseNet that requires as much com- that a DenseNet-BC with only 0.8M trainable parameters putation as a ResNet-50 performs on par with a ResNet-101, is able to achieve comparable accuracy as the 1001-layer which requires twice as much computation. (pre-activation) ResNet [12] with 10.2M parameters. It is worth noting that our experimental setup implies that we use hyperparameter settings that are optimized for Implicit Deep Supervision. One explanation for the im- ResNets but not for DenseNets. It is conceivable that more proved accuracy of dense convolutional networks may be extensive hyper-parameter searches may further improve that individual layers receive additional supervision from the performance of DenseNet on ImageNet. the loss function through the shorter connections. One can interpret DenseNets to perform a kind of “deep supervi- sion”. The benefits of deep supervision have previously 5. Discussion been shown in deeply-supervised nets (DSN; [20]), which Superficially, DenseNets are quite similar to ResNets: have classifiers attached to every hidden layer, enforcing the Eq. (2) differs from Eq. (1) only in that the inputs to H (·) intermediate layers to learn discriminative features. are concatenated instead of summed. However, the implica- DenseNets perform a similar deep supervision in an im- tions of this seemingly small modification lead to substan- plicit fashion: a single classifier on top of the network pro- tially different behaviors of the two network architectures. vides direct supervision to all layers through at most two or three transition layers. However, the loss function and gra- Model compactness. As a direct consequence of the in- dient of DenseNets are substantially less complicated, as the put concatenation, the feature-maps learned by any of the same loss function is shared between all layers. DenseNet layers can be accessed by all subsequent layers. This encourages feature reuse throughout the network, and Stochastic vs. deterministic connection. There is an leads to more compact models. interesting connection between dense convolutional net- The left two plots in Figure 4 show the result of an works and stochastic depth regularization of residual net- experiment that aims to compare the parameter efficiency works [13]. In stochastic depth, layers in residual networks of all variants of DenseNets (left) and also a comparable are randomly dropped, which creates direct connections be- 

8 . Dense Block 1 Dense Block 2 Dense Block 3 tween the surrounding layers. As the pooling layers are 1 1 1 1 0.9 never dropped, the network results in a similar connectiv- 0.8 3 3 3 ity pattern as DenseNet: there is a small probability for Source layer (s) 0.7 5 5 5 0.6 any two layers, between the same pooling layers, to be di- 0.5 7 7 7 0.4 rectly connected—if all intermediate layers are randomly 0.3 9 9 9 dropped. Although the methods are ultimately quite dif- 0.2 0.1 11 11 1 ferent, the DenseNet interpretation of stochastic depth may Transition layer 1 Transition layer 2 Classification layer 0 2 4 6 8 10 12 2 4 6 8 10 12 2 4 6 8 10 12 provide insights into the success of this regularizer. Target layer () Target layer () Target layer () Figure 5: The average absolute filter weights of convolutional lay- Feature Reuse. By design, DenseNets allow layers ac- ers in a trained DenseNet. The color of pixel (s, ) encodes the av- cess to feature-maps from all of its preceding layers (al- erage L1 norm (normalized by number of input feature-maps) of though sometimes through transition layers). We conduct the weights connecting convolutional layer s to within a dense an experiment to investigate if a trained network takes ad- block. Three columns highlighted by black rectangles correspond vantage of this opportunity. We first train a DenseNet on to two transition layers and the classification layer. The first row C10+ with L = 40 and k = 12. For each convolutional encodes weights connected to the input layer of the dense block. layer within a block, we compute the average (absolute) weight assigned to connections with layer s. Figure 5 shows DenseNets tend to yield consistent improvement in accu- a heat-map for all three dense blocks. The average absolute racy with growing number of parameters, without any signs weight serves as a surrogate for the dependency of a convo- of performance degradation or overfitting. Under multi- lutional layer on its preceding layers. A red dot in position ple settings, it achieved state-of-the-art results across sev- ( , s) indicates that the layer makes, on average, strong use eral highly competitive datasets. Moreover, DenseNets of feature-maps produced s-layers before. Several observa- require substantially fewer parameters and less computa- tions can be made from the plot: tion to achieve state-of-the-art performances. Because we adopted hyperparameter settings optimized for residual net- 1. All layers spread their weights over many inputs within works in our study, we believe that further gains in accuracy the same block. This indicates that features extracted of DenseNets may be obtained by more detailed tuning of by very early layers are, indeed, directly used by deep hyperparameters and learning rate schedules. layers throughout the same dense block. Whilst following a simple connectivity rule, DenseNets 2. The weights of the transition layers also spread their naturally integrate the properties of identity mappings, deep weight across all layers within the preceding dense supervision, and diversified depth. They allow feature reuse block, indicating information flow from the first to the throughout the networks and can consequently learn more last layers of the DenseNet through few indirections. compact and, according to our experiments, more accurate 3. The layers within the second and third dense block models. Because of their compact internal representations consistently assign the least weight to the outputs of and reduced feature redundancy, DenseNets may be good the transition layer (the top row of the triangles), in- feature extractors for various computer vision tasks that dicating that the transition layer outputs many redun- build on convolutional features, e.g., [4, 5]. We plan to dant features (with low weight on average). This is in study such feature transfer with DenseNets in future work. keeping with the strong results of DenseNet-BC where exactly these outputs are compressed. Acknowledgements. The authors are supported in part by 4. Although the final classification layer, shown on the the NSF III-1618134, III-1526012, IIS-1149882, the Of- very right, also uses weights across the entire dense fice of Naval Research Grant N00014-17-1-2175 and the block, there seems to be a concentration towards final Bill and Melinda Gates foundation. GH is supported by feature-maps, suggesting that there may be some more the International Postdoctoral Exchange Fellowship Pro- high-level features produced late in the network. gram of China Postdoctoral Council (No.20150015). ZL is supported by the National Basic Research Program of China Grants 2011CBA00300, 2011CBA00301, the NSFC 6. Conclusion 61361136003. We also thank Daniel Sedra, Geoff Pleiss We proposed a new convolutional network architec- and Yu Sun for many insightful discussions. ture, which we refer to as Dense Convolutional Network (DenseNet). It introduces direct connections between any References two layers with the same feature-map size. We showed that [1] C. Cortes, X. Gonzalvo, V. Kuznetsov, M. Mohri, and DenseNets scale naturally to hundreds of layers, while ex- S. Yang. Adanet: Adaptive structural learning of artificial hibiting no optimization difficulties. In our experiments, neural networks. arXiv preprint arXiv:1607.01097, 2016. 2 

9 . [2] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei- [24] Y. Netzer, T. Wang, A. Coates, A. Bissacco, B. Wu, and A. Y. Fei. Imagenet: A large-scale hierarchical image database. In Ng. Reading digits in natural images with unsupervised fea- CVPR, 2009. 5 ture learning, 2011. In NIPS Workshop, 2011. 5 [3] S. E. Fahlman and C. Lebiere. The cascade-correlation learn- [25] M. Pezeshki, L. Fan, P. Brakel, A. Courville, and Y. Bengio. ing architecture. In NIPS, 1989. 2 Deconstructing the ladder network architecture. In ICML, [4] J. R. Gardner, M. J. Kusner, Y. Li, P. Upchurch, K. Q. 2016. 3 Weinberger, and J. E. Hopcroft. Deep manifold traversal: [26] G. Pleiss, D. Chen, G. Huang, T. Li, L. van der Maaten, Changing labels with convolutional features. arXiv preprint and K. Q. Weinberger. Memory-efficient implementation of arXiv:1511.06421, 2015. 8 densenets. arXiv preprint arXiv:1707.06990, 2017. 5 [5] L. Gatys, A. Ecker, and M. Bethge. A neural algorithm of [27] A. Rasmus, M. Berglund, M. Honkala, H. Valpola, and artistic style. Nature Communications, 2015. 8 T. Raiko. Semi-supervised learning with ladder networks. [6] X. Glorot, A. Bordes, and Y. Bengio. Deep sparse rectifier In NIPS, 2015. 3 neural networks. In AISTATS, 2011. 3 [28] A. Romero, N. Ballas, S. E. Kahou, A. Chassang, C. Gatta, [7] I. Goodfellow, D. Warde-Farley, M. Mirza, A. Courville, and and Y. Bengio. Fitnets: Hints for thin deep nets. In ICLR, Y. Bengio. Maxout networks. In ICML, 2013. 5 2015. 5 [8] S. Gross and M. Wilber. Training and investigating residual [29] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, nets, 2016. 5, 7 S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, [9] B. Hariharan, P. Arbeláez, R. Girshick, and J. Malik. Hyper- et al. Imagenet large scale visual recognition challenge. columns for object segmentation and fine-grained localiza- IJCV. 1, 7 tion. In CVPR, 2015. 2 [30] P. Sermanet, S. Chintala, and Y. LeCun. Convolutional neu- [10] K. He, X. Zhang, S. Ren, and J. Sun. Delving deep into ral networks applied to house numbers digit classification. In rectifiers: Surpassing human-level performance on imagenet ICPR, pages 3288–3291. IEEE, 2012. 5 classification. In ICCV, 2015. 5 [31] P. Sermanet, K. Kavukcuoglu, S. Chintala, and Y. LeCun. [11] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning Pedestrian detection with unsupervised multi-stage feature for image recognition. In CVPR, 2016. 1, 2, 3, 4, 5, 6 learning. In CVPR, 2013. 2 [12] K. He, X. Zhang, S. Ren, and J. Sun. Identity mappings in [32] J. T. Springenberg, A. Dosovitskiy, T. Brox, and M. Ried- deep residual networks. In ECCV, 2016. 2, 3, 5, 7 miller. Striving for simplicity: The all convolutional net. [13] G. Huang, Y. Sun, Z. Liu, D. Sedra, and K. Q. Weinberger. arXiv preprint arXiv:1412.6806, 2014. 5 Deep networks with stochastic depth. In ECCV, 2016. 1, 2, [33] N. Srivastava, G. E. Hinton, A. Krizhevsky, I. Sutskever, and 5, 8 R. Salakhutdinov. Dropout: a simple way to prevent neural [14] S. Ioffe and C. Szegedy. Batch normalization: Accelerating networks from overfitting. JMLR, 2014. 6 deep network training by reducing internal covariate shift. In [34] R. K. Srivastava, K. Greff, and J. Schmidhuber. Training ICML, 2015. 3 very deep networks. In NIPS, 2015. 1, 2, 5 [15] A. Krizhevsky and G. Hinton. Learning multiple layers of [35] I. Sutskever, J. Martens, G. Dahl, and G. Hinton. On the features from tiny images. Tech Report, 2009. 5 importance of initialization and momentum in deep learning. In ICML, 2013. 5 [16] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In [36] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, NIPS, 2012. 3, 7 D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich. Going deeper with convolutions. In CVPR, 2015. 2, 3 [17] G. Larsson, M. Maire, and G. Shakhnarovich. Fractalnet: Ultra-deep neural networks without residuals. arXiv preprint [37] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna. arXiv:1605.07648, 2016. 1, 3, 5, 6 Rethinking the inception architecture for computer vision. In CVPR, 2016. 2, 3, 4 [18] Y. LeCun, B. Boser, J. S. Denker, D. Henderson, R. E. [38] S. Targ, D. Almeida, and K. Lyman. Resnet in Howard, W. Hubbard, and L. D. Jackel. Backpropagation resnet: Generalizing residual architectures. arXiv preprint applied to handwritten zip code recognition. Neural compu- arXiv:1603.08029, 2016. 2 tation, 1(4):541–551, 1989. 1 [39] J. Wang, Z. Wei, T. Zhang, and W. Zeng. Deeply-fused nets. [19] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradient- arXiv preprint arXiv:1605.07716, 2016. 3 based learning applied to document recognition. Proceed- ings of the IEEE, 86(11):2278–2324, 1998. 1, 3 [40] B. M. Wilamowski and H. Yu. Neural network learning without backpropagation. IEEE Transactions on Neural Net- [20] C.-Y. Lee, S. Xie, P. Gallagher, Z. Zhang, and Z. Tu. Deeply- works, 21(11):1793–1803, 2010. 2 supervised nets. In AISTATS, 2015. 2, 3, 5, 7 [41] S. Yang and D. Ramanan. Multi-scale recognition with dag- [21] Q. Liao and T. Poggio. Bridging the gaps between residual cnns. In ICCV, 2015. 2 learning, recurrent neural networks and visual cortex. arXiv [42] S. Zagoruyko and N. Komodakis. Wide residual networks. preprint arXiv:1604.03640, 2016. 2 arXiv preprint arXiv:1605.07146, 2016. 3, 5, 6 [22] M. Lin, Q. Chen, and S. Yan. Network in network. In ICLR, [43] Y. Zhang, K. Lee, and H. Lee. Augmenting supervised neural 2014. 3, 5 networks with unsupervised objectives for large-scale image [23] J. Long, E. Shelhamer, and T. Darrell. Fully convolutional classification. In ICML, 2016. 3 networks for semantic segmentation. In CVPR, 2015. 2