Fast backbone in NAS: Train faster, edge-ready models

You can now train smaller, lower-latency object detection models with a new Fast backbone in Neural Architecture Search, alongside the existing Standard and Plus backbones. The Fast backbone searches for the most lightweight architectures, so it is the best starting point when you plan to deploy to a CPU or an edge device.

You choose the backbone before training, then run the search and compare the resulting models on the speed-versus-accuracy graph as usual. Fast, Standard, and Plus are separate backbones, so a run started on one does not share checkpoints with the others.

Learn more about Neural Architecture Search