Source-linked AI summary
MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications
Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, Hartwig Adam
TL;DR
MobileNets address the need for accurate vision models that also meet the size and speed constraints of mobile and embedded platforms. They use depthwise separable convolutions and two model-shrinking hyperparameters, achieving strong efficiency and accuracy across comparisons and applications. The paper concludes that MobileNets offer favorable size, speed, and accuracy characteristics across varied tasks.
Problem
Vision applications increasingly require timely recognition on computationally limited platforms, while accuracy improvements have not necessarily improved network size and speed.
Method
MobileNets use depthwise separable convolutions and width and resolution multipliers to construct smaller, more efficient models.
Results
MobileNets achieve strong size, speed, and accuracy characteristics compared with popular models and remain effective across a wide variety of tasks.
Takeaways & Limitations
The architecture and hyperparameters let model developers trade reasonable accuracy for reduced size and latency according to application requirements.
Takeaways & Limitations
The model-shrinking analysis assumes square input and output feature maps with matching spatial dimensions, although the authors state the results generalize to arbitrary sizes and aspect ratios.
Abstract
from arXiv · showhide
We present a class of efficient models called MobileNets for mobile and embedded vision applications. MobileNets are based on a streamlined architecture that uses depth-wise separable convolutions to build light weight deep neural networks. We introduce two simple global hyper-parameters that efficiently trade off between latency and accuracy. These hyper-parameters allow the model builder to choose the right sized model for their application based on the constraints of the problem. We present extensive experiments on resource and accuracy tradeoffs and show strong performance compared to other popular models on ImageNet classification. We then demonstrate the effectiveness of MobileNets across a wide range of applications and use cases including object detection, finegrain classification, face attributes and large scale geo-localization.
1. Introduction
The paper addresses the gap between increasingly accurate but inefficient vision networks and the real-time constraints of mobile and embedded applications. It introduces an efficient architecture and two hyperparameters for matching model size and latency to application requirements.
- Deepening and complicating networks has improved accuracy, but not necessarily efficiency in size and speed.Real-world robotics, self-driving, and augmented-reality systems require timely recognition on computationally limited platforms.
- MobileNets provide an efficient architecture and two hyperparameters for building very small, low-latency models matched to application requirements.
- The paper studies MobileNet architecture and model-shrinking hyperparameters, then evaluates them on ImageNet and diverse applications.The applications and use cases are described alongside the ImageNet experiments.
2. Prior Work
Prior work on efficient neural networks includes compressing pretrained models and training small networks directly. MobileNets instead target application-specific resource matching, primarily optimizing latency while also producing small models.
- Small-network methods generally compress pretrained networks or train small networks directly.
- MobileNets let developers choose a small architecture that matches application resource restrictions such as latency and size.The approach primarily focuses on latency while also yielding small networks.
- Depthwise separable convolutions build MobileNets, extending factorized-convolution approaches used in earlier efficient architectures.
- Compression, factorization, pruning, quantization, hashing, coding, and distillation represent complementary approaches to obtaining smaller or faster networks.Distillation is described as complementary to the MobileNet approach and is used in some applications.
3. MobileNet Architecture
MobileNet is built around depthwise separable convolutions and a structured architecture designed for efficient mobile and embedded vision models. Width and resolution multipliers further shrink computation, parameters, and model size while exposing resource–accuracy trade-offs.
- Depthwise Separable Convolution: Depthwise separable convolution factorizes standard convolution into depthwise filtering followed by 1 × 1 pointwise channel combination.The depthwise layer applies one filter per input channel; the pointwise layer combines the resulting channels.
- Depthwise Separable Convolution: MobileNet’s computation depends on channel counts, kernel size, and feature-map size, while factorization separates filtering from channel combination.This separation breaks the interaction between output-channel count and kernel size and reduces computation.
- Depthwise Separable Convolution: 3 × 3 depthwise separable convolutions use 8 to 9 times less computation than standard convolutions with only a small accuracy reduction.
- Network Structure and Training: The MobileNet body uses depthwise separable convolutions except in the first layer, with batch normalization and ReLU after convolutional layers.Strided depthwise convolutions perform downsampling, and the final average pooling reduces spatial resolution before the fully connected classifier.
- Width Multiplier: Thinner Models: The width multiplier α uniformly reduces input and output channels, lowering computational cost and parameter count by roughly α^2.Typical settings are α = 1, 0.75, 0.5, and 0.25; reduced models are trained from scratch.
4. Experiments
Experiments evaluate MobileNet’s architectural choices, resource–accuracy trade-offs, comparisons with popular models, and performance across several vision applications.
- Model Choices: At similar computation and parameter counts, making MobileNets thinner is 3% better than making them shallower.The thinner models use the width multiplier, whereas shallower models remove five separable-filter layers.
- Model Shrinking Hyperparameters: Accuracy drops smoothly as width or input resolution decreases, with a sharp degradation when the width multiplier reaches α = 0.25.Across 16 width–resolution combinations, the accuracy–computation relationship is log linear except for a jump when α = 0.25.
- Model Comparisons: Full MobileNet is nearly as accurate as VGG16 while being 32 times smaller and 27 times less compute intensive.It is also more accurate than GoogleNet while being smaller and requiring more than 2.5 times less computation.
- Model Comparisons: A reduced MobileNet is 4% better than AlexNet while being 45× smaller and 9.4× less compute intensive.It is also 4% better than Squeezenet at about the same size and 22× less computation.
- Applications: Across fine-grained recognition, geolocation, face attributes, and object detection, MobileNet retains competitive performance with substantially reduced model size or computation.The MobileNet PlaNet version has 13 million parameters and 0.58 million Mult-Adds, while face-attribute classification achieves similar mean AP using 1% of the Mult-Adds.
5. Conclusion
The conclusion presents MobileNets as efficient depthwise-separable architectures and highlights width and resolution multipliers for trading accuracy against size and latency across applications.
- Conclusion: MobileNets use depthwise separable convolutions and two multipliers to construct smaller and faster models with reasonable accuracy trade-offs.The width and resolution multipliers let developers match models to application resource constraints.
- Conclusion: Experiments compare MobileNets with popular models and demonstrate effectiveness across a wide variety of tasks.The conclusion identifies size, speed, and accuracy characteristics as the comparison focus.