EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks

Proposed compound scaling, systematically balancing network depth, width, and image resolution using a fixed scaling coefficient.

Highly popular feature extractor and classification backbone for production vision APIs due to optimal accuracy-to-latency tradeoffs.

Assumptions

• Balancing network depth, width, and input resolution achieves substantially higher accuracy under fixed resource constraints

Limitations

• High image resolutions in larger variants (B7: 600x600) dramatically increase GPU memory requirements and reduce batch sizes