State-of-the-art vision and image-based tasks such as image classification (Krizhevsky et al., 2012; Simonyan & Zisserman, 2015; He et al., 2016), object detection (Ren et al., 2017; Huang et al., 2017) and segmentation (Long et al., 2015) are all built upon deep convolutional neural networks (CNNs). While CNN architectures have evolved to become more efficient, the general trend has been to use larger models with greater memory and compute requirements to achieve higher accuracy.
One common approach to reduce costs is to prune over-parameterized CNNs. If performed in a coarse-grain manner this approach is known as channel pruning (Ye et al., 2018; He et al., 2017; Liu et al., 2017; Wen et al., 2016; Alvarez & Salzmann, 2016)
. Channel pruning evaluates channel saliency measures and removes all input and output connections from unimportant channels – generating a smaller dense model. A saliency-based pruning method, however, has threefold disadvantages. Firstly, by removing channels, the capabilities of CNNs are permanently lost, and the resulting CNN may never regain its accuracy for difficult inputs for which the removed channels were responsible. Secondly, despite the fact that channel pruning may drastically shrink model size, without careful design, computational resources cannot be effectively reduced in a CNN without a detrimental impact on its accuracy. Finally, the saliency of a neuron is not static, which can be illustrated by the feature visualization inFigure 1. Here, a CNN is shown a set of input images, channel neurons in a convolutional output may get highly excited, whereas another set of images elicit little response from the same channel. This is in line with our understanding of CNNs that neurons in a convolutional layer specialize in recognizing distinct features, and the relative importance of a neuron depends heavily on the inputs.
The above shortcomings prompts the question: why should we prune by static importance, if the importance is highly input-dependent? Surely, a more promising alternative is to prune dynamically depending on the current input. A dynamic channel pruning strategy allows the network to learn to prioritize certain convolutional channels and ignore irrelevant ones. Instead of simply reducing model size at the cost of accuracy with pruning, we can accelerate convolution by selectively computing only a subset of channels predicted to be important at run-time, while considering the sparse input from the preceding convolution layer. By doing this, not only do we save computational resources, we also preserve all neurons of the full model, which minimizes the impact on model accuracy.
In this paper, we propose feature boosting and suppression (FBS) to dynamically amplify and suppress output channels computed by the convolutional layer. Intuitively, we can imagine that the flow of information of each output channel can be amplified or restricted under the control of a “valve”. This allows salient information to flow freely while we stop all information from unimportant channels and skip their computation. Unlike pruning statically, the valves use features from the previous layer to predict the saliency of output channels. With conventional stochastic gradient descent (SGD) methods, the predictor can learn to adapt itself by observing the input and output features of the convolution operation.
FBS introduces tiny auxiliary connections to existing convolutional layers. The minimal overhead added to the existing model is thus negligible when compared to the potential speed up provided by the dynamic sparsity. Existing dynamic computation strategies in CNNs (Lin et al., 2017; Odena et al., 2017; Bolukbasi et al., 2017)
produce on/off pruning decisions or execution path selections. Training them thus often resorts to reinforcement learning, which in practice is often computationally expensive. Even though our models similarly use non-differentiable functions, contrary to these methods, our unified losses are still well-minimized with conventional SGD.
We apply FBS to a custom CIFAR-10 (Krizhevsky et al., 2014)classifier and popular CNN models such as VGG-16 (Simonyan & Zisserman, 2015) and ResNet-18 (He et al., 2016) trained on the ImageNet ILSVRC2012 dataset (Deng et al., 2009). Empirical results show that under the same speed-up constraints, our strategy can produce models with validation accuracies surpassing all other static channel pruning and dynamic conditional execution methods examined in the paper.
2 Related Work
2.1 Structured Sparsity
Since LeCun et al. (1990) introduced optimal brain damage, the idea of creating more compact and efficient CNNs by removing connections or neurons has received significant attention. Early literature on pruning deep CNNs zero out individual weight parameters (Hassibi et al., 1994; Guo et al., 2016). This results in highly irregular sparse connections, which were notoriously difficult for GPUs to exploit. This has prompted custom accelerator solutions that exploit sparse weights (Parashar et al., 2017; Han et al., 2016). Although supporting both sparse and dense convolutions efficiently normally involves some compromises in terms of efficiency or performance.
Alternatively, recent work has thus increasingly focused on introducing structured sparsity (Wen et al., 2016; Ye et al., 2018; Alvarez & Salzmann, 2016; Zhou et al., 2016), which can be exploited by GPUs and allows custom accelerators to focus solely on efficient dense operations. Wen et al. (2016)
added group Lasso on channel weights to the model’s training loss function. This has the effect of reducing the magnitude of channel weights to diminish during training, and remove connections from zeroed-out channels. To facilitate this process,Alvarez & Salzmann (2016) additionally used proximal gradient descent, while Li et al. (2017) and He et al. (2018) proposed to prune channels by thresholds, i.e. they set unimportant channels to zero, and fine-tune the resulting CNN. The objective to induce sparsity in groups of weights may present difficulties for gradient-based methods, given the large number of weights that need to be optimized. A common approach to overcome this is to solve (He et al., 2017) or learn (Liu et al., 2017; Ye et al., 2018) channel saliencies to drive the sparsification of CNNs. He et al. (2017) solved an optimization problem which limits the number of active convolutional channels while minimizing the reconstruction error on the convolutional output. Liu et al. (2017) used Lasso regularization on channel saliencies to induce sparsity and prune channels with a global threshold. Ye et al. (2018)
learned to sparsify CNNs with an iterative shrinkage/thresholding algorithm applied to the scaling factors in batch normalization.Huang et al. (2018) adopted reinforcement learning to train agents that take the layer weights as inputs and produce binary channel pruning decisions. PerforatedCNNs, proposed by Figurnov et al. (2016), use predefined masks that are model-agnostic to skip the output pixels in convolutional layers.
2.2 Dynamic Execution
In a pruned model produced by structured sparsity methods, the capabilities of the pruned neurons and connections are permanently lost. Therefore, many propose to use dynamic networks as an alternative to structured sparsity. During inference, a dynamic network can use the input data to choose parts of the network to evaluate.
Convolutional layers are usually spatially sparse, i.e. their activation outputs may contain only small patches of salient regions. A number of recent publications exploit this for acceleration. Dong et al. (2017) introduced low-cost collaborative layers which induce spatial sparsity in cheap convolutions, so that the main expensive ones can use the same sparsity information. Figurnov et al. (2017) proposed spatially adaptive computation time for residual networks (He et al., 2016), which learns the number of residual blocks required to compute a certain spatial location. Almahairi et al. (2016) presented dynamic capacity networks, which use the gradient of a coarse output’s entropy to select salient locations in the input image for refinement. Ren et al. (2018) assumed the availability of a priori spatial sparsity in the input image, and accelerated the convolutional layer by computing non-sparse regions.
There are dynamic networks that make binary decisions or multiple choices for the inference paths taken. BlockDrop, proposed by Wu et al. (2018), trains a policy network to skip blocks in residual networks. Liu & Deng (2018) proposed conditional branches in deep neural networks (DNNs), and use Q-learning to train the branching policies. Odena et al. (2017)
designed a DNN with layers containing multiple modules, and decided which module to use with a recurrent neural network (RNN).Lin et al. (2017) learned an RNN to adaptively prune channels in convolutional layers. The on/off decisions commonly used in these networks cannot be represented by differentiable functions, hence the gradients are not well-defined. Consequently, the dynamic networks above train their policy functions by reinforcement learning. There exist, however, methods that workaround such limitations. Shazeer et al. (2017)
introduced sparsely-gated mixture-of-experts and used a noisy ranking on the backpropagate-able gating networks to select the expensive experts to evaluate.Bolukbasi et al. (2017) trained differentiable policy functions to implement early exits in a DNN. Hua et al. (2018) learned binary policies that decide whether partial or all input channels are used for convolution, but approximate the gradients of the non-differentiable policy functions with continuous ones.
3 Feature Boosting and Suppression
We start with a high-level illustration (Figure 2) of how FBS accelerates a convolutional layer with batch normalization (BN). The auxiliary components (in red) predicts the importance of each output channel based on the input features, and amplify the output features accordingly. Moreover, certain output channels are predicted to be entirely suppressed (or zero-valued as represented by ), such output sparsity information can advise the convolution operation to skip the computation of these channels, as indicated by the dashed arrow. It is notable that the expensive convolution can be doubly accelerated by skipping the inactive channels from both the input features and the predicted output channel saliencies. The rest of this section provides detailed explanation of the components in Figure 2.
For simplicity, we consider a deep sequential batch-normalized (Ioffe & Szegedy, 2015) CNN with convolutional layers, i.e. , where the layer computes the features , which comprise of channels of features with height and width . The layer is thus defined as:
Here, additions () and multiplications () are element-wise,
denotes the ReLU activation,are trainable parameters, normalizes each channel of features across the population of , with
respectively containing the population mean and variance of each channel, and a smallprevents division by zero:
Additionally, computes the convolution of input features
using the weight tensor, where is the kernel size. Specifically, FBS concerns the optimization of functions, as a CNN spends the majority of its inference time in them, using multiply-accumulate operations (MACs) for the layer.
3.2 Designing a Dynamic Layer
Consider the following generalization of a layer with dynamic execution:
where and respectively use weight parameters and and may have additional inputs, and compute tensors of the same output shape, denoted by and . Intuitively, the expensive can always be skipped for any index whenever the cost-effective evaluates to . Here, the superscript is used to index the slice of the tensor. For example, if we have features containing channels of -by- features, retrieves the feature image. We can further sparsify and accelerate the layer by adding, for instance, a Lasso on to the total loss, where is the expectation of over :
Despite the simplicity of this formulation, it is however very tricky to design properly. Under the right conditions, we can arbitrarily minimize the Lasso while maintaining the same output from the layer by scaling parameters. For example, in low-cost collaborative layers (Dong et al., 2017), and are simply convolutions (with or without ReLU activation) that respectively have weights and . Since and are homogeneous functions, one can always halve and double to decrease (4) while the network output remains the same. In other words, the optimal network must have , which is infeasible in finite-precision arithmetic. For the above reasons, Dong et al. (2017) observed that the additional loss in (4) always degrades the CNN’s task performance. Ye et al. (2018) pointed out that gradient-based training algorithms are highly inefficient in exploring such reparameterization patterns, and channel pruning methods may experience similar difficulties. Shazeer et al. (2017) avoided this limitation by finishing with a softmax normalization, but (4) can no longer be used as the softmax renders the -norm, which now evaluates to 1, useless. In addition, similar to sigmoid, softmax (without the cross entropy) is easily saturated, and thus may equally suffer from vanishing gradients. Many instead design to produce on/off decisions and train them with reinforcement learning as discussed in Section 2.
3.3 Feature Boosting and Suppression with Channel Saliencies
Instead of imposing sparsity on features or convolutional weight parameters (e.g. Wen et al. (2016); Alvarez & Salzmann (2016); Li et al. (2017); He et al. (2018)), recent channel pruning methods (Liu et al., 2017; Ye et al., 2018) induce sparsity on the BN scaling factors . Inspired by them, FBS similarly generate a channel-wise importance measure. Yet contrary to them, instead of using the constant BN scaling factors , we predict channel importance and dynamically amplify or suppress channels with a parametric function dependent on the output from the previous layer . Here, we propose to replace layer definition for all with which employs dynamic channel pruning:
where a low-overhead policy evaluates the pruning decisions for the computationally demanding :
Here, is a -winners-take-all function, i.e. it returns a tensor identical to , except that we zero out entries in that are smaller than the largest entries in absolute magnitude. In other words, provides a pruning strategy that computes only most salient channels predicted by , and suppresses the remaining channels with zeros. In Section 3.4, we provide a detailed explanation of how we design a cheap that learns to predict channel saliencies.
It is notable that our strategy prune least salient output channels from layer, where the density can be varied to sweep the trade-off relationship between performance and accuracy. Moreover, pruned channels contain all-zero values. This allows the subsequent
layer to trivially make use of input-side sparsity, since all-zero features can be safely skipped even for zero-padded layers. Because all convolutions can exploit both input- and output-side sparsity, the speed-up gained from pruning is quadratic with respect to the pruning ratio. For instance, dynamically pruning half of the channels in all layers gives rise to a dynamic CNN that uses approximatelyof the original MACs.
Theoretically, FBS does not introduce the reparameterization discussed in Section 3.2. By batch normalizing the convolution output, the convolution kernel is invariant to scaling. Computationally, it is more efficient to train. Many alternative methods use non-differentiable functions that produce on/off decisions. In general, DNNs with these policy functions are incompatible with SGD, and resort to reinforcement learning for training. In contrast, (6) allows end-to-end training, as is a piecewise differentiable and continuous function like ReLU. Srivastava et al. (2015) suggested that in general, a network is easier and faster to train for complex tasks and less prone to catastrophic forgetting, if it uses functions such as that promote local competition between many subnetworks.
3.4 Learning to Predict Channel Saliencies
This section explains the design of the channel saliency predictor . To avoid significant computational cost in , we subsample by reducing the spatial dimensions of each channel to a scalar using the following function :
where reduces the channel of to a scalar using, for instance, the -norm , -norm, -norm, or the variance of . The results in Section 4 use the -norm by default, which is equivalent to global average pooling for the ReLU activated . We then design to predict channel saliencies with a fully connected layer following the subsampled activations , where is the weight tensor of the layer:
We generally initialize with and apply He et al. (2015)’s initialization to . To induce sparsity in , we use ReLU as shown in (8), then follow Liu et al. (2017) and Ye et al. (2018) and regularize all layers with the Lasso in the total loss, where in our experiments.
We ran extensive experiments on CIFAR-10 (Krizhevsky et al., 2014) and the ImageNet ILSVRC2012 (Deng et al., 2009), two popular image classification datasets. We designed M-CifarNet, a custom 8-layer CNN for CIFAR-10 (see Appendix B for its structure), using only parameters with and top-1 and top-5 accuracies respectively. M-CifarNet is much smaller than a VGG-16 on CIFAR-10 (Liu et al., 2017), which uses parameters and only more accurate. Because of its compactness, our CNN is more challenging to accelerate. By faithfully reimplementing Network Slimming (NS) (Liu et al., 2017), we closely compare FBS with NS under various speedup constraints. For ILSVRC2012, we augment two popular CNN variants, ResNet-18 (He et al., 2016) and VGG-16 (Simonyan & Zisserman, 2015), and provide detailed accuracy/MACs trade-off comparison against recent structured pruning and dynamic execution methods. we provide the detailed experimental setup in Appendix A, and further examine and explain the effectiveness of FBS on CNNs by exploring the model design space in Appendix C.
Our method begins by first replacing all convolutional layer computations with (5), and initializing the new convolutional kernels with previous parameters. Initially, we do not suppress any channel computations by using density in (6) and fine-tune the resulting network. For fair comparison against NS, we then follow Liu et al. (2017) by iteratively decrementing the overall density of the network by in each step, and thus gradually using fewer channels to sweep the accuracy/performance trade-off. The difference is that NS prunes channels by ranking globally, while FBS prunes around of each layer.
. The heat map in (fig:cifar10:heatmap) reveals the individual probability of skipping a channel for each channel (-axis), when an image of a category (-axis) is shown to the network with .
By respectively applying NS and FBS to our CIFAR-10 classifier and incrementally increasing sparsity, we produce the trade-off relationships between number of operations (measured in MACs) and the classification accuracy as shown in Figure 2(a). FBS clearly surpasses NS in its ability to retain the task accuracy under an increasingly stringent computational budget. Besides comparing FBS against NS, we are interested in combining both methods, which demonstrates the effectiveness of FBS if the model is already less redundant, i.e. it cannot be pruned further using NS without degrading the accuracy by more than . The composite method (NS+FBS) is shown to successfully regain most of the lost accuracy due to NS, producing a trade-off curve closely matching FBS. It is notable that under the same accuracy constraints, FBS, NS+FBS, and NS respectively achieve , , and speed-up ratios. Conversely for a speed-up target, they respectively produce models with accuracies not lower than , and .
Figure 2(b) demonstrates that our FBS can effectively learn to amplify and suppress channels when dealing with different input images. The 8 heat maps respectively represent the channel skipping probabilities of the 8 convolutional layers. The brightness of the pixel at location denotes the probability of skipping the channel when looking at an image of the category. The heat maps verify our belief that the auxiliary network learned to predict which channels specialize to which features, as channels may have drastically distinct probabilites of being used for images of different categories. The model here is a M-CifarNet using FBS with , which has a top-1 accuracy of (top-5 ). Moreover, channels in the heat maps are sorted so the channels that are on average least frequently evaluated are placed on the left, and channels shaded in stripes are never evaluated. The network in Figure 2(b) is not only approximately faster than the original, by removing the unused channels, we also reduce the number of weights by . This reveals that FBS naturally subsumes channel pruning strategies such as NS, as we can simply prune away channels that are skipped regardless of the input. It is notable that even though we specified a universal density , FBS learned to adjust its dynamicity across all layers, and prune different ratios of channels from the convolutional layers.
4.2 ImageNet ILSVRC2012 Classification
Residual networks (He et al., 2016) adopt sequential structure of residual blocks: , where is the output of the block, is either an identity function or a downsampling convolution, and consists of a sequence of convolutions. For residual networks, we directly apply FBS to all convolutional layers, with a difference in the way we handle the feature summation. Because the block receives as input the sum of the two features with sparse channels and , the channels of this sum is considered sparse only when the same channels in both features are sparse.
By applying FBS and NS respectively to ResNet-18, we saw that the ILSVRC2012 validation accuracy of FBS consistently outperforms NS under different speed-up constraints (see Appendix C for the trade-off curves). For instance, at , it utilizes only MACs ( speed-up) to achieve a top-1 error rate of , while NS requires MACs ( faster) for a similar error rate of . When compared across recent dynamic execution methods examined in Table 1, FBS demonstrates simultaneously the highest speed-up and the lowest error rates. It is notable that the baseline accuracies for FBS refer to a network that has been augmented with the auxiliary layers featuring FBS but suppress no channels, i.e. . We found that this method brings immediate accuracy improvements, an increase of in top-1 and in top-5 accuracies, to the baseline network, which is in line with our observation on M-CifarNet.
Since VGG-16 is computationally intensive with over MACs, We first applied NS on VGG-16 to reduce the computational and memory requirements, and ease the training of the FBS-augmented variant. We assigned a budget in top-5 accuracy degradation and compressed the network using NS, which gave us a smaller VGG-16 with of all channels pruned. The resulting network is a lot less redundant, which almost halves the compute requirements, with only MACs remaining. We then apply FBS to the well-compressed network. In Table 2, we compare the performances of different structured pruning and dynamic execution methods to FBS. At a speed-up constraint of , FBS shows a minimal increase of in top-5 error rate. At speed-ups of and , FBS only degrades the top-5 error rates by and respectively.
In summary, we proposed feature boosting and suppression that helps CNNs to achieve significant reductions in the compute required while maintaining high accuracies. FBS fully preserves the capabilities of CNNs and predictively boosts important channels to help the accelerated models retain high accuracies. We demonstrated that FBS achieves around and speed-ups respectively on ResNet-18 and VGG-16 within loss of top-5 accuracy. Under the same performance constraints, the accuracy gained by FBS surpasses all recent structured pruning and dynamic execution methods examined in this paper. In addition, it can serve as an off-the-shelf technique for accelerating many popular CNN networks and the fine-tuning process is unified in the traditional SGD which requires no algorithmic changes in training. The implementation of FBS and the optimized networks are fully open source, and will be released to the public.
Almahairi et al. (2016)
Amjad Almahairi, Nicolas Ballas, Tim Cooijmans, Yin Zheng, Hugo Larochelle, and
Dynamic capacity networks.
Proceedings of the 33rd International Conference on International Conference on Machine Learning (ICML), pp. 2549–2558, 2016.
- Alvarez & Salzmann (2016) Jose M Alvarez and Mathieu Salzmann. Learning the number of neurons in deep networks. In D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett (eds.), Advances in Neural Information Processing Systems (NIPS), pp. 2270–2278. 2016.
- Bolukbasi et al. (2017) Tolga Bolukbasi, Joseph Wang, Ofer Dekel, and Venkatesh Saligrama. Adaptive neural networks for efficient inference. In Proceedings of the 34th International Conference on Machine Learning (ICML), pp. 527–536, 2017.
- Deng et al. (2009) J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A large-scale hierarchical image database. In
- Dong et al. (2017) Xuanyi Dong, Junshi Huang, Yi Yang, and Shuicheng Yan. More is less: A more complicated network with less inference complexity. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017.
- Figurnov et al. (2017) Michael Figurnov, Maxwell D. Collins, Yukun Zhu, Li Zhang, Jonathan Huang, Dmitry Vetrov, and Ruslan Salakhutdinov. Spatially adaptive computation time for residual networks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017.
- Figurnov et al. (2016) Mikhail Figurnov, Aizhan Ibraimova, Dmitry P Vetrov, and Pushmeet Kohli. PerforatedCNNs: Acceleration through elimination of redundant convolutions. In Advances in Neural Information Processing Systems (NIPS), pp. 947–955, 2016.
- Guo et al. (2016) Yiwen Guo, Anbang Yao, and Yurong Chen. Dynamic network surgery for efficient DNNs. In Advances in Neural Information Processing Systems (NIPS), 2016.
- Han et al. (2016) Song Han, Xingyu Liu, Huizi Mao, Jing Pu, Ardavan Pedram, Mark A Horowitz, and William J Dally. Eie: efficient inference engine on compressed deep neural network. In Computer Architecture (ISCA), 2016 ACM/IEEE 43rd Annual International Symposium on, pp. 243–254. IEEE, 2016.
- Hassibi et al. (1994) Babak Hassibi, David G. Stork, and Gregory Wolff. Optimal brain surgeon: Extensions and performance comparisons. In J. D. Cowan, G. Tesauro, and J. Alspector (eds.), Advances in Neural Information Processing Systems (NIPS), pp. 263–270. 1994.
- He et al. (2015) Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the 2015 IEEE International Conference on Computer Vision (ICCV), ICCV ’15, pp. 1026–1034, Washington, DC, USA, 2015. IEEE Computer Society. ISBN 978-1-4673-8391-2. doi: 10.1109/ICCV.2015.123.
- He et al. (2016) Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016.
He et al. (2018)
Yang He, Guoliang Kang, Xuanyi Dong, Yanwei Fu, and Yi Yang.
Soft filter pruning for accelerating deep convolutional neural
International Joint Conference on Artificial Intelligence (IJCAI), pp. 2234–2240, 2018.
- He et al. (2017) Yihui He, Xiangyu Zhang, and Jian Sun. Channel pruning for accelerating very deep neural networks. IEEE International Conference on Computer Vision (ICCV), pp. 1398–1406, 2017.
- Hua et al. (2018) Weizhe Hua, Christopher De Sa, Zhiru Zhang, and G. Edward Suh. Channel gating neural networks. CoRR, abs/1805.12549, 2018. URL http://arxiv.org/abs/1805.12549.
- Huang et al. (2017) J. Huang, V. Rathod, C. Sun, M. Zhu, A. Korattikara, A. Fathi, I. Fischer, Z. Wojna, Y. Song, S. Guadarrama, and K. Murphy. Speed/accuracy trade-offs for modern convolutional object detectors. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 3296–3297, July 2017.
- Huang et al. (2018) Qiangui Huang, Kevin Zhou, Suya You, and Ulrich Neumann. Learning to prune filters in convolutional neural networks. In IEEE Winter Conference on Computer Vision. 2018.
- Ioffe & Szegedy (2015) Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Proceedings of the 32Nd International Conference on International Conference on Machine Learning (ICML), pp. 448–456, 2015.
- Krizhevsky et al. (2012) Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems (NIPS). 2012.
- Krizhevsky et al. (2014) Alex Krizhevsky, Vinod Nair, and Geoffrey Hinton. The CIFAR-10 and CIFAR-100 datasets. http://www.cs.toronto.edu/ kriz/cifar.html, 2014.
- LeCun et al. (1990) Yann LeCun, John S. Denker, and Sara A. Solla. Optimal brain damage. In Advances in Neural Information Processing Systems (NIPS), pp. 598–605. 1990.
- Li et al. (2017) Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. 2017.
- Lin et al. (2017) Ji Lin, Yongming Rao, Jiwen Lu, and Jie Zhou. Runtime neural pruning. In Advances in Neural Information Processing Systems (NIPS), pp. 2181–2191. 2017.
- Liu & Deng (2018) Lanlan Liu and Jia Deng. Dynamic deep neural networks: Optimizing accuracy-efficiency trade-offs by selective execution. 2018.
- Liu et al. (2017) Zhuang Liu, Jianguo Li, Zhiqiang Shen, Gao Huang, Shoumeng Yan, and Changshui Zhang. Learning efficient convolutional networks through network slimming. In International Conference on Computer Vision (ICCV), 2017.
- Long et al. (2015) J. Long, E. Shelhamer, and T. Darrell. Fully convolutional networks for semantic segmentation. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 3431–3440, June 2015.
- Odena et al. (2017) Augustus Odena, Dieterich Lawson, and Christopher Olah. Changing model behavior at test-time using reinforcement learning. 2017.
- Parashar et al. (2017) Angshuman Parashar, Minsoo Rhu, Anurag Mukkara, Antonio Puglielli, Rangharajan Venkatesan, Brucek Khailany, Joel Emer, Stephen W Keckler, and William J Dally. Scnn: An accelerator for compressed-sparse convolutional neural networks. In ACM SIGARCH Computer Architecture News, volume 45, pp. 27–40. ACM, 2017.
- Ren et al. (2018) Mengye Ren, Andrei Pokrovsky, Bin Yang, and Raquel Urtasun. SBNet: Sparse blocks network for fast inference. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018.
- Ren et al. (2017) S. Ren, K. He, R. Girshick, and J. Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1137–1149, June 2017.
- Shazeer et al. (2017) Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. 2017.
- Simonyan & Zisserman (2015) Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations (ICLR), 2015.
- Srivastava et al. (2015) Rupesh Kumar Srivastava, Jonathan Masci, Faustino J. Gomez, and Jürgen Schmidhuber. Understanding locally competitive networks. In International Conference on Learning Representations (ICLR), 2015.
- Wen et al. (2016) Wei Wen, Chunpeng Wu, Yandan Wang, Yiran Chen, and Hai Li. Learning structured sparsity in deep neural networks. In Advances in Neural Information Processing Systems (NIPS), pp. 2074–2082. 2016.
- Wu et al. (2018) Zuxuan Wu, Tushar Nagarajan, Abhishek Kumar, Steven Rennie, Larry S. Davis, Kristen Grauman, and Rogerio Feris. BlockDrop: Dynamic inference paths in residual networks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018.
- Ye et al. (2018) Jianbo Ye, Xin Lu, Zhe L. Lin, and James Z. Wang. Rethinking the smaller-norm-less-informative assumption in channel pruning of convolution layers. In International Conference on Learning Representations (ICLR), 2018.
- Zhou et al. (2016) Hao Zhou, Jose M. Alvarez, and Fatih Porikli. Less is more: Towards compact cnns. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling (eds.), Computer Vision – ECCV 2016, pp. 662–677, Cham, 2016. Springer International Publishing. ISBN 978-3-319-46493-0.
Appendix A Experiment Setup
We trained M-CifarNet (see Appendix B) with a learning rate and a batch size. We reduced the learning rate by a factor of for every epochs. To compare FBS against NS fairly, every model with a new target MACs budget were consecutively initialized with the previous model, and trained for a maximum of epochs, which is enough for all models to converge to the best obtainable accuracies. For NS, we follow Liu et al. (2017) and start training with an -norm sparsity regularization weighted by on the BN scaling factors. We then prune at epochs and fine-tune the resulting network without the sparsity regularization.
We additionally employed image augmentation procedures from Krizhevsky et al. (2012) to preprocess each training example. Each CIFAR-10 example was randomly horizontal flipped and slightly perturbed in the brightness, saturation and hue.
ILSVRC2012 classifiers, i.e. ResNet-18 and VGG-16, were trained with a procedure similar to the one above. The difference was that they were trained for a maximum of epochs, the learning rate was decayed for every epochs, and NS models were all pruned at epochs. For image preprocessing, we additionally cropped and stretched/squeezed images randomly following Krizhevsky et al. (2012).
Appendix B Details of M-CifarNet on CIFAR-10
For the CIFAR-10 classification task, we use M-CifarNet, a custom designed CNN, with less than parameters and takes MACs to perform inference for a -by- RGB image. The architecture is illustrated in Table 3, where all convolutional layers use kernels, the Shape column shows the shapes of each layer’s features, and pool7 is a global average pooling layer. Table 3 additionally provides further comparisons of layer-wise compute costs between FBS, NS, and the composition of the two methods (NS+FBS). It is notable that the FBS column has two different output channel counts, where the former is the number of computed channels for each inference, and the latter is the number of channels remaining in the layer after removing the unused channels.
|Layer||Shape||Number of MACs (Output Channels)|
|conv0||1.5 M (64)||1.3 M (52)||893 k (32/62)||860 k (32)|
|conv1||33.2 M (64)||27.0 M (64)||8.4 M (32/42)||10.2 M (39)|
|conv2||16.6 M (128)||15.9 M (123)||4.2 M (64/67)||5.9 M (74)|
|conv3||33.2 M (128)||31.9 M (128)||8.3 M (64/79)||11.6 M (77)|
|conv4||33.2 M (128)||33.1 M (128)||8.3 M (64/83)||12.1 M (77)|
|conv5||14.1 M (192)||13.4 M (182)||3.6 M (96/128)||4.9 M (110)|
|conv6||21.2 M (192)||11.6 M (111)||5.4 M (96/152)||4.3 M (67)|
|conv7||21.2 M (192)||12.3 M (192)||5.4 M (96/96)||4.5 M (116)|
|fc||1.9 k (10)||1.9 k (10)||960 (10)||1.1 k (10)|
|Total||174.3 M||146.5 M||44.3 M||54.2 M|
Figure 4 shows how the skipping probabilites heat maps of the convolutional layer conv4 evolve as we train M-CifarNet. The network was trained for 12 epochs, and we saved the model at every epoch. The heat maps are generated with the saved models in sequence, where we apply the same reordering to all heat map channels with the sorted result from the first epoch. It can be observed that as we train the network, the channel skipping probabilites become more pronounced.
Appendix C Additional Results on ILSVRC2012
Finally, Figure 5 compares the accuracy/performance trade-off curves between FBS and NS.