Source-linked AI summary
ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices
Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, Jian Sun
TL;DR
Mobile CNNs need high accuracy under very limited computation, but existing efficient designs leave substantial 1 × 1 convolution cost and restrict information flow across channel groups. ShuffleNet combines pointwise group convolution with channel shuffle, achieving 7.8% lower ImageNet top-1 error than MobileNet at 40 MFLOPs and ~13× actual speedup over AlexNet with comparable accuracy.
Problem
Efficient CNN architectures must improve accuracy within tens or hundreds of MFLOPs, while existing designs leave substantial pointwise-convolution cost and restrict cross-group information flow.
Method
ShuffleNet combines pointwise group convolution with channel shuffle in an efficient architecture designed for small networks.
Results
7.8% lower ImageNet top-1 error than MobileNet at 40 MFLOPs, while achieving ~13× actual speedup over AlexNet with comparable accuracy.
Takeaways & Limitations
ShuffleNet provides superior performance across ImageNet classification and MS COCO object detection benchmarks under constrained computation budgets.
Takeaways & Limitations
The authors present a simple reference design and note that further hyper-parameter tuning might produce better results.
Abstract
from arXiv · showhide
We introduce an extremely computation-efficient CNN architecture named ShuffleNet, which is designed specially for mobile devices with very limited computing power (e.g., 10-150 MFLOPs). The new architecture utilizes two new operations, pointwise group convolution and channel shuffle, to greatly reduce computation cost while maintaining accuracy. Experiments on ImageNet classification and MS COCO object detection demonstrate the superior performance of ShuffleNet over other structures, e.g. lower top-1 error (absolute 7.8%) than recent MobileNet on ImageNet classification task, under the computation budget of 40 MFLOPs. On an ARM-based mobile device, ShuffleNet achieves ~13x actual speedup over AlexNet while maintaining comparable accuracy.
1. Introduction
ShuffleNet targets accurate CNN inference under tens- to hundreds-of-MFLOPs budgets for mobile platforms. It combines pointwise group convolution with channel shuffle to reduce computation while preserving cross-channel information flow, achieving strong ImageNet, MS COCO, and hardware results.
- Motivation: Tens to hundreds of MFLOPs define the paper’s target regime, contrasting with accurate CNNs that require billions of FLOPs.The focus is common mobile platforms such as drones and robots.
- Problem: Dense 1 × 1 convolutions make Xception and ResNeXt inefficient in extremely small networks.The paper identifies costly dense 1 × 1 convolutions as the bottleneck motivating pointwise group convolutions.
- Method: Pointwise group convolutions reduce 1 × 1 convolution complexity, while channel shuffle enables information flow across feature channels.Together, these techniques form the ShuffleNet architecture, which allows more feature map channels at a fixed computation budget.
- Experiments: 7.8% absolute lower ImageNet top-1 error at 40 MFLOPs gives ShuffleNet a significant margin over MobileNet.The models are evaluated on ImageNet classification and MS COCO object detection, with controlled experiments supporting the design principles.
- Hardware Results: ∼13× actual speedup over AlexNet is achieved on an off-the-shelf ARM-based computing core while maintaining comparable accuracy.Theoretical speedup is 18×.
2. Related Work
Related work develops efficient model designs for embedded devices through architectural innovations, lightweight convolutions, and model search. It also covers channel shuffling and inference-acceleration methods such as pruning, quantization, and factorization.
- Efficient Model Designs: Efficient model designs target high-quality deep neural networks on embedded devices, with GoogLeNet and SqueezeNet reducing complexity or computation through architectural changes.The passage identifies model design as important for computer vision and motivates efficient designs by embedded-device constraints.
- Efficient Model Designs: 26.0% @ 564 MFLOPs vs. 26.3% @ 524 MFLOPs compares mobile NASNet with ShuffleNet on ImageNet classification error.The cited work uses reinforcement learning and model search, but does not report models below 150 MFLOPs or actual mobile-device inference time.
- Group Convolution: Group convolution originated in AlexNet, was demonstrated in ResNeXt, and underlies lightweight approaches including Xception and MobileNet’s depthwise separable convolutions.MobileNet is described as achieving state-of-the-art results among lightweight models.
- Channel Shuffle Operation: Channel shuffle was rarely discussed in prior efficient-model work, although cuda-convnet supported an equivalent random sparse convolution operation.The passage distinguishes that earlier operation’s purpose and limited later use from channel shuffle’s role in grouped convolution.
- Model Acceleration: Model acceleration reduces inference cost while preserving a pre-trained model’s accuracy through pruning, quantization, and factorization.Pruning removes redundant network connections or channels, while quantization and factorization reduce calculation redundancy.
3. Approach
ShuffleNet addresses the limitations of stacked group convolutions with pointwise group convolution and channel shuffle, then combines them with depthwise convolution in an efficient unit. The architecture stacks these units into three stages and uses group number and channel scaling to control complexity.
- Motivation: Pointwise group convolution reduces computation, but stacked group convolutions restrict outputs to a small fraction of input channels.This connectivity problem motivates channel shuffle.
- Channel Shuffle: Channel shuffle rearranges channels across groups so subsequent group convolutions receive different channel subgroups and fully relate input and output channels.The operation enables more powerful structures with multiple group-convolutional layers.
- ShuffleNet Unit: The ShuffleNet unit replaces the first 1 × 1 layer with pointwise group convolution and channel shuffle, while applying 3 × 3 depthwise convolution to bottleneck features.It is based on a residual bottleneck design.
- Complexity: hw(2cm/g + 9m) FLOPs is required by ShuffleNet, versus hw(2cm + 9m) for ResNet and hw(2cm + 9m^2/g) for ResNeXt.Here c × h × w is the input size, m is the bottleneck width, and g is the number of convolution groups.
- Complexity: Given a computational budget, ShuffleNet can use wider feature maps, which is critical because tiny networks often have too few channels to process information.The wider maps result from the unit’s lower complexity under comparable settings.
- Architecture: The architecture stacks ShuffleNet units into three stages, using stride = 2 for each stage’s first block and doubling output channels for the next stage.Bottleneck channels are set to 1/4 of the output channels, and the design is intended as a simple reference configuration.
- Architecture: Scaling ShuffleNet 1× by s multiplies its filters by s, making overall complexity roughly s2 times that of ShuffleNet 1×.This provides a simple way to customize the network to a desired complexity.
4. Experiments
Experiments show that pointwise group convolution and channel shuffle improve ShuffleNet’s classification performance under fixed complexity, while ShuffleNet generally outperforms alternative efficient architectures. The results also examine accuracy–width–group trade-offs and practical mobile inference efficiency.
- Pointwise Group Convolution: Group convolutions consistently outperform g = 1 counterparts, with gains of 1.2%, 3.5%, and 4.4% for ShuffleNet 1×, 0.5×, and 0.25×, respectively.Smaller models benefit more from group convolutions under the same complexity constraint.
- Pointwise Group Convolution: Larger group numbers can saturate or reduce accuracy, but they consistently improve smaller ShuffleNet 0.25× models by enabling wider feature maps.Fewer input channels per filter may harm representation capability at large group counts.
- Channel Shuffle: Channel shuffle consistently boosts classification scores across complexity settings, especially when the group number is relatively large.Its purpose is to enable cross-group information flow across multiple group convolution layers.
- Comparison with Other Structures: At 38 MFLOPs, ShuffleNet uses 576 Stage 4 output channels versus 50–288 for the compared structures, supporting its accuracy advantage under fixed computation.The compared values are VGG-like 50, ResNet 192, ResNeXt 192, Xception-like 288, and ShuffleNet 576.
- Comparison with MobileNet: 7.8%: ShuffleNet surpasses MobileNet at approximately 40 MFLOPs, while it is 3.1% more accurate than MobileNet 1× at a cost of 500 MFLOPs.ShuffleNet is reported as superior to MobileNet across the evaluated complexity levels.
- Efficiency and Inference Speed: 18×: ShuffleNet 0.5× is theoretically faster than AlexNet with a comparable classification score.The paper separately evaluates actual running time on an ARM-based mobile device.