MnasNet: Platform-Aware Neural Architecture Search for Mobile

Abstract

Designing convolutional neural networks (CNN) for mobile devices is challenging because mobile models need to be small and fast, yet still accurate. Although significant efforts have been dedicated to design and improve mobile CNNs on all dimensions, it is very difficult to manually balance these trade-offs when there are so many architectural possibilities to consider. In this paper, we propose an automated mobile neural architecture search (MNAS) approach, which explicitly incorporate model latency into the main objective so that the search can identify a model that achieves a good trade-off between accuracy and latency. Unlike previous work, where latency is considered via another, often inaccurate proxy (e.g., FLOPS), our approach directly measures real-world inference latency by executing the model on mobile phones. To further strike the right balance between flexibility and search space size, we propose a novel factorized hierarchical search space that encourages layer diversity throughout the network. Experimental results show that our approach consistently outperforms state-of-the-art mobile CNN models across multiple vision tasks. On the ImageNet classification task, our MnasNet achieves 75.2% top-1 accuracy with 78ms latency on a Pixel phone, which is 1.8× faster than MobileNetV2 [29] with 0.5% higher accuracy and 2.3× faster than NASNet [36] with 1.2% higher accuracy. Our MnasNet also achieves better mAP quality than MobileNets for COCO object detection. Code is at https

Differentiable Learning-to-Group Channels via Groupable Convolutional Neural Networks

“…val accuracy). These methods were extended in many ways such as progressive searching [13], parameter sharing [21], network transformation [3], resource-constrained searching [30], and differentiable searching like DARTS [15] and SNAS [34]. Evolutionary algorithm is an alternative to RL.…”

“…a is a weighted product to approximate the Pareto optimal problem [30] and a is a constant value. We have a = 0 if ζ ≤ o, implying that the complexity constraint is satisfied.…”

“…Otherwise, a = α is used to penalize the model complexity when ζ > o. For the value of α, [30] empirically set α = −1 or −0.07 and this setting works well in reinforcement learning by using rewards. However, these empirical values make the regularizer too sensitive in our problem.…”

Hardware/Software Co-Exploration of Neural Architectures

“…Architecture Search Space: For CIFAR-10, we use convolutional architectures as the backbone. For every convolutional layer, we first determine the filter size in [24,36,48,64], the kernel size in [1,3,5,7], and the strides. Two sets of experiments are carried out to determine the strides: (1) by exploring the child networks with a fixed stride of 1; (2) by allowing the controller to predict the strides in [1,2].…”

Non-Discriminative Data or Weak Model? On the Relative Importance of Data and Model Resolution

“…We concentrate on the architecture as described in table 2. We note that this only scratches the surface of possible architectures, and also potentially provides a useful target for NAS-style approaches [16,14] to find better isometric models. In our experiments, we use isometric networks consisting of Mo-bileNetV3 + SE bottleneck blocks [5].…”