MobileNets

Post date: May 06, 2021 11:27:50 PM

Original paper on MobileNet;

• 1. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications on arXiv.org and Google AI Blog

MobileNetv2:

• 1. MobileNet V2: Inverted Residuals and Linear Bottlenecks on arXiv.org and Google AI Blog

  2. It is part of the TensorFlow-Slim Image Classification Library

  3. Also available on Google Colab

Explaining both networks:

• 1. Pretrained MobileNet model V2 & V3

  2. Google's MobileNets on the iPhone by Matthijs Hollemans. He wanted to know how fast it runs on the iPhone. The results: fast! 0.05 s per inference, 3x faster than Inception, 10x faster than VGGNet-16. Uses 4 million parameters (compared to 24 million for Inception-v3 and 130 million for VGGNet-16).

  3. MobileNet version 2 blog post by Matthijs Hollemans  V2 uses

  4. Side note: Matthijs previously implemented VGGNet on the iPhone. VGGNet requires 2-10 s to initialize 550 MB parameters. Inference takes 0.25 - 0.3 s per image, on iPhone 6s. Inception initalizes in 0.5 s and takes 0.15 - 0.2 s per inference. Inception is more complex, but uses less memory & runs faster than VGGnet.

V2 uses less MAC operations, less parameters and runs faster.