EdgeNeXt
EdgeNeXt is an efficient hybrid CNN-Transformer architecture designed for mobile vision applications. It combines convolutional neural networks and transformer models to achieve a balance between accuracy and computational efficiency, making it suitable for deployment on edge devices with limited resources. The architecture introduces a split depth-wise transpose attention (SDTA) encoder that splits input tensors into multiple channel groups, using depth-wise convolution alongside self-attention across channel dimensions. This design increases the receptive field and encodes multi-scale features effectively. EdgeNeXt is available in multiple variants, including ultra-lightweight and base models. For example, the smallest variant with 1.3M parameters achieves 71.2% top-1 accuracy on ImageNet-1K, outperforming MobileViT by 2.2% with 28% fewer FLOPs. The base model with 5.6M parameters achieves 79.4% top-1 accuracy, while a larger ImageNet-21K pretrained variant reaches 83.31% accuracy. Beyond image classificati