DIY Your EasyNAS for Vision: Convolution Operation Merging, Map Channel Reducing, and Search Space to Supernet...

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Differentiable Architecture Search (DARTS) is popular but computationally expensive for complex vision tasks due to over-parameterized supernets.
• Existing methods face limitations in memory and computational efficiency, hindering broader applicability.

Purpose of the Study:

• To introduce EasyNAS, a novel and efficient Neural Architecture Search (NAS) method.
• To overcome the computational and memory limitations of traditional DARTS for complex vision tasks.
• To enable NAS application across diverse tasks including classification, detection, and semantic segmentation.

Main Methods:

• Developed an operator merging technique to share and merge weights of candidate convolution operations.
• Implemented a dynamic channel refinement strategy for improved efficiency.
• Introduced a configurable search space-to-supernet conversion tool utilizing atomic search components.

Main Results:

• Achieved state-of-the-art accuracy (76.2%) on ImageNet for classification.
• Attained 40.1 mAP at 120 FPS on MS-COCO for detection.
• Reached 77.05 mAP₅₀ on DOTA for rotation detection with 21.1 M parameters.
• Secured 72.6% mIoU at 173 FPS on Cityscape for semantic segmentation after 0.7 GPU-day search.

Conclusions:

• EasyNAS significantly enhances the efficiency of Neural Architecture Search (NAS).
• The proposed methods enable state-of-the-art performance across various complex vision tasks.
• EasyNAS offers a practical and scalable solution for NAS in real-world applications.