MobileNetV2

MobileNetV2 is a lightweight convolutional neural network from Google designed for efficient image classification on resource-constrained hardware. Its memorable building block is the inverted residual with a linear bottleneck: channels expand for computation, then compress again, with shortcuts joining narrow representations where possible. That is different from ResNet-50’s heavier residual design and from ViT’s patch tokens. ‘Mobile’ describes the efficiency goal, not a requirement to run only on phones. TT-Metal lists MobileNetV2 for 224x224 images on a single n150 Wormhole at batch 10 and on n300 Wormhole with DP=2 at batch 20. The two rows are one model at single- and two-device data-parallel scales.