ShuffleNet is a convolutional neural network architecture designed specifically for mobile devices with very limited computing power. ShuffleNet aims to achieve high accuracy while minimizing computation cost. ShuffleNet introduces this operation to reduce computation. It combines pointwise convolutions, 1×1 convolutions with group convolutions. The group convolutions allow efficient parallelization across channels. It groups input channels and performs independent convolutions within each group. Channel shuffle helps maintain accuracy while reducing computation. It enhances information exchange between different parts of the network.
SqueezeNet is a remarkable deep neural network architecture designed for image classification with a focus on minimal model size and fewer parameters. SqueezeNet employs fire modules, which consist of two layers:
Squeeze Layer: A 1×1 convolutional layer, for squeeze operation reduces the number of input channels.
Expand Layer: Two parallel 1×1 and 3×3 convolutional layers for expand operation capture spatial information.
The squeeze layer compresses feature maps, effectively “squeezing” parameters. Lightweight depth wise convolutions further reduce model complexity. SqueezeNet achieves competitive accuracy on the ImageNet dataset while having just 5MB of parameters.
ShuffleNet is a remarkable CNN architecture that balances efficiency and accuracy, making it an excellent choice for mobile vision applications with limited computational resources.
SqueezeNet’s compactness and accuracy make it an excellent choice for resource-constrained environments, including edge devices and mobile applications.