Fd-Mobilenet: Improved Mobilenet with a Fast Downsampling Strategy

Zheng Qin, Zhaoning Zhang, Xiaotao Chen, Changjian Wang, Yuxing Peng

TL;DR

This paper introduces FD-MobileNet, a novel convolutional neural network architecture designed for extremely resource-constrained devices. By employing a "fast downsampling" strategy early in the network, FD-MobileNet significantly reduces computational cost while increasing information capacity, leading to improved accuracy and faster inference speeds compared to existing compact networks like MobileNet and ShuffleNet.

What Problem Does This Paper Address?

Modern deep convolutional neural networks (CNNs) achieve state-of-the-art performance in computer vision tasks but demand substantial computational resources and large model sizes. This makes them unsuitable for deployment on mobile or embedded devices with limited processing power, memory, and battery life. While various compression and acceleration techniques exist (tensor decomposition, quantization, pruning), this paper focuses on designing inherently compact and efficient network architectures. Specifically, it addresses the performance degradation observed in existing compact networks like MobileNet when operating under very low computational budgets (e.g., 10-140 MFLOPs), where simply reducing network width leads to insufficient information capacity.

Methods

The core innovation of FD-MobileNet lies in its fast downsampling strategy. Unlike traditional CNNs that gradually reduce spatial resolution, FD-MobileNet aggressively downsamples the input image within the initial layers. This is achieved by using depthwise separable convolutions with large strides early in the network.

Key aspects of the method:

  • Depthwise Separable Convolutions: Inherited from MobileNet, these are used as the building blocks. They factorize a standard convolution into a depthwise convolution and a pointwise convolution, drastically reducing computation and parameters.
  • Fast Downsampling: Instead of spreading downsampling across many layers, FD-MobileNet performs a significant spatial reduction (e.g., 32x downsampling within the first 12 layers). This is achieved by using depthwise separable convolutions with strides of 2 in the early stages. The authors explicitly avoid max-pooling in this early stage, finding it doesn't improve performance but adds computation.
  • Increased Information Capacity: By downsampling early, the network can afford to have more channels in the subsequent layers. This increased channel count allows the network to store more information, which is crucial for maintaining performance at very low computational budgets.
  • Simple Architecture: FD-MobileNet maintains the simple, sequential stacking of layers found in MobileNet, which is beneficial for efficient engineering implementation and avoids the complex memory/cache switching issues that can arise with more intricate structures like ShuffleNet's bypass connections.

Datasets

  • ILSVRC 2012: Used for image classification tasks.
  • PASCAL VOC 2007: Used for object detection tasks.

Key Insights and Ideas

  • Downsampling Strategy is Crucial for Low-Budget Networks: The paper highlights that when downsampling occurs is as important as how much it occurs, especially for networks with very limited computational budgets. Aggressive early downsampling can be more beneficial than gradual reduction.
  • Information Capacity vs. Spatial Resolution: For extremely low FLOPs, the ability of a network to encode information (channel count) becomes more critical than preserving fine-grained spatial details in early layers. FD-MobileNet balances this by downsampling early to save computation, then using the saved budget to increase channel width.
  • Engineering Friendliness Matters: The simple, sequential architecture of FD-MobileNet, inherited from MobileNet, leads to efficient memory and cache utilization on actual hardware, resulting in faster inference speeds than networks with more complex control flow (like ShuffleNet's bypass connections).

Key Results

  • Image Classification (ILSVRC 2012):
    • FD-MobileNet consistently outperforms MobileNet across various computational budgets.
  • At 12 MFLOPs, FD-MobileNet achieves 5.5% higher top-1 accuracy than MobileNet.
  • At 40 MFLOPs, FD-MobileNet is 5.6% better than MobileNet.
  • Performance is comparable to ShuffleNet under similar budgets.
  • Object Detection (PASCAL VOC 2007):
    • FD-MobileNet detector shows significant improvements over MobileNet.
  • At 12 MFLOPs, FD-MobileNet achieves 3.6% higher mAP than MobileNet.
  • At 40 MFLOPs, the gap widens to 2.8% mAP.
  • FD-MobileNet shows consistent improvements on challenging classes for MobileNet.
  • Inference Speed (ARM-based device):
    • FD-MobileNet achieves approximately 1.1x speedup over MobileNet across tested budgets.
  • FD-MobileNet is significantly faster than ShuffleNet: 1.82x speedup at 12 MFLOPs, and 1.33x-1.47x at higher budgets, despite ShuffleNet sometimes having fewer FLOPs. This is attributed to FD-MobileNet's efficient architecture avoiding cache misses.

Limitations and Future Work

  • Limitations: While FD-MobileNet achieves comparable or slightly worse results than ShuffleNet in some cases, the authors attribute this to ShuffleNet's potentially more powerful bypass connection structure. However, they also note that this structure leads to slower actual inference on mobile devices due to cache inefficiencies.
  • Future Work: The authors plan to explore applying the fast downsampling strategy to other compact network architectures, such as ShuffleNet, to potentially combine the benefits of both approaches.

Summary updated

Abstract

We present Fast-Downsampling MobileNet (FD-MobileNet), an efficient and accurate network for very limited computational budgets (e.g., 10-140 MFLOPs). Our key idea is applying a fast downsampling strategy to Mobile Net framework. In FD-Mobile Net, we perform 32× downsampling within 12 layers, only half the layers in the original MobileNet. This design brings three advantages: (i) It remarkably reduces the computational cost. (ii) It increases the information capacity and achieves significant performance improvements. (iii) It is engineering-friendly and provides fast actual inference speed. Experiments on ILSVRC 2012 and PASCAL VOC datasets demonstrate that FD-Mobile Net consistently outperforms MobileNet and achieves comparable results with ShufflieNet under different computational budgets, for instance, surpassing Mobile-Net by 5.5% on the ILSVRC 2012 top-l accuracy and 8.3% on the VOC 2007 mAP under a complexity of 12 MFLOPs. On an ARM-based device, FD-Mobile Net achieves 1.11× inference speedup over Mobile Net and 1.82× over Shufflie Net under the same complexity.
BibTeX
@inproceedings{qin2018fd,
  author = {Zheng Qin and Zhaoning Zhang and Xiaotao Chen and Changjian Wang and Yuxing Peng},
  title = {{Fd-Mobilenet: Improved Mobilenet with a Fast Downsampling Strategy}},
  booktitle = {ICIP},
  year = {2018},
  doi = {10.1109/icip.2018.8451355},
}
Figure