🎓Lightweight neural network model

[kurisaW]

轻量级神经网络模型

[kurisaW]

1.MobileNet: MobileNet is a small and efficient CNN architecture that was specifically designed for mobile and embedded devices. It uses depthwise separable convolutions to reduce the number of parameters and computations. The final model size can be around 16-17 MB.

2.MobileNetV2: MobileNetV2 is a popular neural network architecture designed for mobile and embedded vision applications. It consists of depthwise separable convolution layers, which allow for significant reduction in model size without sacrificing accuracy. The final model size is around 14MB.

3.MobileNetV3: MobileNetV3 is the latest iteration in the MobileNet series of lightweight neural networks. It introduces several new features, including dynamic convolution and h-swish activation function, which further reduce model size while improving accuracy. The final model size is around 5MB.

4.EfficientNet-Lite: EfficientNet-Lite is a family of lightweight models derived from the EfficientNet architecture. These models achieve state-of-the-art accuracy on image classification tasks while being significantly smaller than their counterparts. The final model size can range from 4 to 8 MB depending on the specific variant used.

5.SqueezeNet: SqueezeNet is another lightweight CNN architecture that uses a combination of 1x1 and 3x3 filters to reduce the number of parameters. It has fewer layers than conventional CNNs, which helps in faster inference. The final model size can be around 5 MB.

6.ShuffleNet: ShuffleNet is a neural network architecture that uses group convolutions and channel shuffling to reduce computational complexity while preserving accuracy. It has fewer parameters and is suitable for low-power devices. The final model size can be around 14-15 MB.

7.ShuffleNetV2: ShuffleNetV2 is an efficient neural network architecture designed for mobile devices, which leverages channel shuffling to reduce computation cost. The final model size is around 15MB.

8.Tiny YOLO: YOLO (You Only Look Once) is a popular object detection algorithm that can also be adapted for fight recognition. Tiny YOLO is a lightweight version of YOLO that uses fewer layers and smaller filters. It has fewer parameters and requires less memory. The final model size can be around 35-40 MB.

9.MnasNet: MnasNet is a mobile-oriented neural network architecture which uses a combination of depthwise convolutions and squeeze-and-excitation blocks to achieve high accuracy while keeping the model size small. The final model size is around 12MB.

10.ESPNetv2: ESPNetv2 is a highly efficient segmentation network designed for low-power embedded devices. It uses a spatial pyramid pooling module and a pyramid feature fusion module to achieve high accuracy with minimal computational cost. The final model size is around 1.3MB.

11.EfficientDet-Lite: EfficientDet-Lite is a family of lightweight models derived from the EfficientDet architecture, which is used for object detection tasks. These models use a combination of depthwise separable convolutions and bi-directional feature pyramid networks to achieve high accuracy with minimal computational cost. The final model size can range from 10MB to 30MB depending on the specific variant used.

12.GhostNet: GhostNet is another lightweight neural network architecture that uses ghost modules to reduce computational cost. It achieves high accuracy on image classification tasks while keeping the model size small. The final model size is around 6MB.