[Paper Review] PydMobileNet: Improved Version of MobileNets with Pyramid Depthwise Separable Convolution

Van-Thanh Hoang, Kang-Hyun Jo|arXiv (Cornell University)|Nov 17, 2018

IoT and Edge/Fog Computing45 references4 citations

TL;DR

This paper proposes PydMobileNet, an improved MobileNet variant that replaces standard 3×3 depthwise separable convolutions with a pyramid kernel size strategy to capture richer spatial features. By combining feature maps via addition or concatenation and adjusting width multipliers, PydMobileNet achieves superior accuracy with fewer parameters and better latency trade-offs than MobileNet and other state-of-the-art models on CIFAR-10 and CIFAR-100.

ABSTRACT

Convolutional neural networks (CNNs) have shown remarkable performance in various computer vision tasks in recent years. However, the increasing model size has raised challenges in adopting them in real-time applications as well as mobile and embedded vision applications. Many works try to build networks as small as possible while still have acceptable performance. The state-of-the-art architecture is MobileNets. They use Depthwise Separable Convolution (DWConvolution) in place of standard Convolution to reduce the size of networks. This paper describes an improved version of MobileNet, called Pyramid Mobile Network. Instead of using just a $3 imes 3$ kernel size for DWConvolution like in MobileNet, the proposed network uses a pyramid kernel size to capture more spatial information. The proposed architecture is evaluated on two highly competitive object recognition benchmark datasets (CIFAR-10, CIFAR-100). The experiments demonstrate that the proposed network achieves better performance compared with MobileNet as well as other state-of-the-art networks. Additionally, it is more flexible in fine-tuning the trade-off between accuracy, latency and model size than MobileNets.

Motivation & Objective

• Address the challenge of deploying compact, accurate CNNs on mobile and embedded devices with limited computational resources.
• Improve upon MobileNet's reliance on fixed 3×3 depthwise separable convolutions by enabling multi-scale spatial feature extraction.
• Provide greater flexibility in balancing model accuracy, inference latency, and parameter count through adjustable network width and feature fusion strategies.
• Demonstrate that pyramid kernel depthwise separable convolutions can outperform standard depthwise separable convolutions in terms of accuracy and efficiency on benchmark datasets.

Proposed method

• Replace standard 3×3 depthwise separable convolutions in MobileNet with a pyramid kernel size strategy that applies multiple kernel sizes (e.g., 3×3, 5×5, 7×7) in parallel.
• Use bottleneck residual blocks to control channel dimensions and maintain computational efficiency.
• Combine feature maps from different kernel sizes using either element-wise addition or concatenation to preserve spatial and channel-wise information.
• Introduce a width multiplier hyperparameter to scale the number of channels and control model capacity, enabling fine-grained trade-offs between accuracy, size, and latency.
• Train the network end-to-end on CIFAR-10 and CIFAR-100 datasets using standard optimization protocols.
• Evaluate inference speed on both CPU and GPU to assess real-world deployment feasibility, particularly considering platform-specific inefficiencies in depthwise convolution implementations.

Experimental results

Research questions

• RQ1 Can a pyramid kernel size strategy in depthwise separable convolutions improve feature representation compared to fixed 3×3 kernels in MobileNet?
• RQ2 How does the choice between feature fusion via addition or concatenation affect the accuracy, parameter count, and inference speed of the resulting model?
• RQ3 To what extent can PydMobileNet achieve better accuracy with fewer parameters than MobileNet and other state-of-the-art models on CIFAR-10 and CIFAR-100?
• RQ4 How does the model’s inference speed compare to MobileNet and ResNet when deployed on CPU and GPU, especially given the current inefficiency of depthwise separable convolutions in deep learning frameworks?

Key findings

• PydMobileNet-Add-56-1 achieved a top-1 error rate of 10.7% on CIFAR-100 with only 0.382M parameters, outperforming MobileNet-56-1 (0.416M parameters, 11.1% error) and other state-of-the-art models.
• On CIFAR-10, PydMobileNet variants achieved comparable or lower error rates than MobileNet and ResNet while using significantly fewer parameters.
• The PydMobileNet-Concat-56-0.25 variant achieved 17.7% top-1 error on CIFAR-100 with just 0.175M parameters, demonstrating high efficiency.
• Inference speed evaluation showed that PydMobileNet-Concat variants achieved speeds comparable to MobileNet on both CPU and GPU, suggesting practical deployability despite deeper architecture.
• The model's performance was robust across different width multipliers and fusion strategies, confirming its flexibility in tuning the accuracy-latency-size trade-off.
• Training curves showed that a 56-layer PydMobileNet-0.25-Concat achieved similar test error to a 110-layer ResNet but with significantly fewer parameters and faster convergence.