Read the original document by opening this link in a new tab.
Table of Contents
Abstract
1. Introduction
2. Related Work
3. Split-Attention Networks
3.1. Split-Attention Block
3.2. Efficient Radix-major Implementation
4. Network and Training
4.1. Network Tweaks
4.2. Training Strategy
Summary
ResNeSt is a modularized architecture that applies channel-wise attention on different network branches to enhance visual recognition. The model outperforms EfficientNet in accuracy and latency trade-off on image classification. It introduces a Split-Attention Block that captures cross-feature interactions effectively. The network achieves superior performance on various benchmarks and has been adopted by winning entries in challenges. The training strategy involves large mini-batch distributed training, label smoothing, auto-augmentation, and mixup training for data augmentation.