MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications
- Publication
- Publication venue not yet verified
- First submitted to arXiv
- arXiv:1704.04861
Cite this paper · BibTeX
The citation identifies the linked source version. It cites the original paper and its authors.
@misc{lightweightmobilenetv12017,
title = {{MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications}},
author = {Andrew G. Howard and Menglong Zhu and Bo Chen and Dmitry Kalenichenko and Weijun Wang and Tobias Weyand and Marco Andreetto and Hartwig Adam},
year = {2017},
eprint = {1704.04861},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/1704.04861}
}
Depthwise separable convolution splits a standard convolution into spatial filtering within each channel and a 1×1 convolution that mixes channels. It can reduce both multiply-adds and parameters, making ConvNets more practical on mobile devices. This paper asks how to choose a useful model based on accuracy, size, and latency.
MobileNet did not invent the operation. Its paper cites earlier work by Sifre:
MobileNet’s contribution is to use this factorization systematically in a lightweight network, offer a family of models with adjustable width and input resolution, and evaluate their tradeoffs on ImageNet and other tasks.
That is the now famous MobileNet-V1.
Problem Definition
Interest in small, efficient neural networks was growing. But as the authors point out, many papers on small networks focused on size without considering efficiency. Clearly, we need a network that is both small and fast. Robotics, autonomous driving, and augmented reality need timely recognition on limited hardware, so model selection should consider size, multiply-adds, and measured latency.
Solution
Depthwise Separable Convolution
Part (b) assumes depth multiplier 1. With input channels, it uses spatial filters of size , one per channel. Part (c) then uses pointwise 1×1 filters to combine those channels into output channels. Spatial and channel mixing become separate steps.
Let the output feature map have side length . A standard convolution requires multiply-adds. The depthwise step requires and the pointwise step , giving this cost ratio:
For , , and , the paper’s Table 3 gives about 462M multiply-adds for standard convolution and 52.3M for the factorized layer, with parameters falling from about 2.36M to 0.27M. This is a theoretical single-layer resource comparison. If stride reduces spatial size, here is the output dimension. Whole-network accuracy and speed require separate experiments.
Network Architecture
The first MobileNet layer is a standard 3×3 convolution. Its later main convolutional blocks pair a depthwise 3×3 convolution with a pointwise 1×1 convolution:
The MobileNet structure is built on depthwise separable convolutions as mentioned in the previous section except for the first layer which is a full convolution.
Starting with 224×224×3, the first layer yields 112×112×32. The following depthwise/pointwise pair keeps 112×112 spatial resolution while changing 32 channels to 64. Reading the table this way shows which operation changes spatial size and which changes channel count. Counting depthwise and pointwise convolutions separately, the authors describe a 28-layer network.
One row near the end of Table 1 marks a stride-2 depthwise convolution at 7×7×1024, yet the next row still lists a 7×7×1024 input. Those dimensions cannot both hold, so this stride should not be copied into a precise shape trace without checking the implementation. Paper Table 2 attributes 94.86% of network multiply-adds and 74.59% of parameters to 1×1 convolutions; depthwise 3×3 uses only 3.06% and 1.06%, respectively. Pointwise convolution dominates total cost.
Training Techniques
The authors describe several choices for training MobileNet-V1:
- Use less regularization and data augmentation than for larger models. They omit auxiliary heads and label smoothing and limit very small random crops.
- Apply little or no weight decay to the depthwise filters, which contain few parameters.
- Train with RMSprop in TensorFlow and asynchronous gradient descent similar to Inception-V3.
Discussion
Besides the baseline network, the authors give model builders two controls for shrinking it:
- Width multiplier changes input and output channels to . Typical settings are 1, 0.75, 0.5, and 0.25. Since pointwise convolution dominates, parameters and multiply-adds fall roughly with . The resulting narrower network must be trained afresh.
- Resolution multiplier reduces the input from 224 to 192, 160, or 128 pixels, which also shrinks later feature maps. Convolution multiply-adds fall roughly with , while a fixed model’s parameter count does not change with input size.
With both controls, one depthwise-plus-pointwise layer costs multiply-adds. Width and resolution therefore affect different resources. Tables 6 and 7 show that accuracy falls sharply at ; theoretical savings alone do not pick the best model.
Cost of Depthwise Separable Convolution
Replacing the main convolutions costs 1.1 percentage points of ImageNet top-1 accuracy, while whole-network multiply-adds fall about 8.6-fold and parameters about 7-fold. Device latency requires separate measurement.
Depth or Width?
The models have similar cost, yet the narrower one is 3.1 percentage points more accurate. Relative to the full MobileNet at 70.6%, narrowing loses 2.2 points and removing layers loses 5.3. These are different comparison baselines. The ablation favors narrowing for this model and cost range, not for every architecture.
Comparing Other Models
The full MobileNet in Table 8 has 70.6% top-1, 569M multiply-adds, and 4.2M parameters. VGG16 has 71.5%, 15,300M, and 138M: MobileNet is 0.9 percentage points less accurate, with about 1/33 the parameters and 1/27 the multiply-adds. GoogLeNet has 69.8%, 1,550M, and 6.8M; MobileNet is 0.8 points more accurate while using fewer resources. These are model size and operation counts, not mobile-device speed measurements.
In Table 9, 0.5 MobileNet-160 has 60.2%, 76M, and 1.32M; SqueezeNet has 57.5%, 1,700M, and 1.25M. The smaller MobileNet is 2.7 percentage points more accurate and needs about 1/22 the multiply-adds, but has slightly more parameters. The paper’s prose says “about 4%,” which conflicts with its table, so the difference here is calculated from the rows. Do not carry this reduced model’s result over to the full MobileNet in Table 8.
Conclusion
MobileNet-V1 brought depthwise separable convolution into the spotlight. It kept relatively high accuracy while greatly reducing model size and computation, giving resource-constrained devices a practical CNN architecture. The paper's ImageNet results support this trade-off, although accuracy falls as width, resolution, or block count is reduced.
Its design also inspired later lightweight architectures and helped bring deep learning models to mobile and edge devices. For a deployment, select and input size based on acceptable accuracy, then measure latency and memory on the target device. Theoretical multiply-adds help rank candidates; they do not replace deployment measurements.