YOLOv8 添加EfficientViT替换Backbone

### YOLO Model Backbone Architecture In the evolution of YOLO models, particularly with versions like YOLOv10 and beyond, significant improvements have been made to the backbone architecture by incorporating more efficient network structures such as MobileNetV1[^1], MobileNetV2[^2], and EfficientViT[^3]. These changes aim at enhancing performance while reducing computational requirements. #### Utilizing MobileNetV1 for Lightweight Networks The transition towards lightweight networks has seen the adoption of MobileNetV1 within YOLO's backbone structure. This approach simplifies deployment on mobile devices without compromising much on accuracy or speed. However, integrating this change can be complex due to certain architectural nuances that cannot easily be represented through YAML configuration files alone. #### Advancements with MobileNetV2 Further advancements led to the incorporation of MobileNetV2 into YOLO’s backbone design. Training configurations now allow users to specify paths directly related to these new backbones via command-line instructions tailored specifically for projects utilizing MobileNetV2 as their core component: ```bash cd yolov11项目所在的路径 yolo detect train data=coco128.yaml model=ultralytics/cfg/models/11/yolo11n_MobileNetV2.yaml epochs=100 imgsz=640 batch=16 device=cpu project=yolov11 ``` #### Introduction of EfficientViT For even greater efficiency gains, Microsoft introduced EfficientViT—a highly optimized vision transformer designed explicitly for real-time applications. When applied to YOLO architectures, it not only improves upon traditional CNN-based designs but also integrates advanced features like cascaded group attention mechanisms which further boost its capabilities. ```yaml backbone: # [from, number, module, args] [ [-1, 1, EfficientViT_M0, []], [-1, 1, SPPF, [1024, 5]] ] ``` #### Code Implementation Details To implement modifications effectively across different parts of a neural network including both `backbone` and `head`, specific code adjustments are necessary. For instance, when iterating over components defined under either section (`d["backbone"] + d["head"]`), flags like `is_backbone` help distinguish between sections during runtime processing[^4]: ```python is_backbone = False for i, (f, n, m, args) in enumerate(d["backbone"] + d["head"]): t = m m = getattr(torch.nn, m[3:]) if "nn." in m else globals()[m] # get module ```