efficientnet

Why we included this project

EfficientNet is a lightweight convolutional network that reaches state-of-the-art accuracy on ImageNet while using an order of magnitude fewer parameters and FLOPs than comparable models. This Keras reimplementation ships pretrained checkpoints for the whole B0-to-B7 family, so you can pick a smaller variant when inference speed or memory matters and a larger one when accuracy does. The weights cover ImageNet and five other transfer-learning datasets, which makes fine-tuning for a custom classifier fairly painless. The code mirrors the original TensorFlow implementation closely, so it is also a useful reference if you want to study or modify the architecture rather than use it as a black box.