EfficientNet, but make it Keras (and tf.keras too)

What it does This repository ports Google’s EfficientNet architecture—a family of convolutional networks built by compound-scaling a mobile-size baseline—to Keras and TensorFlow Keras. It provides eight model variants (B0 through B7) with pre-trained weights converted from the original TensorFlow TPU checkpoints, ready for standard Keras inference or transfer learning.

The interesting bit The library functions as a bilingual adapter: it exposes the same models through both legacy keras and modern tensorflow.keras namespaces, a compatibility pivot significant enough to break backward compatibility in 2019. It also includes the weight-conversion plumbing, so the community isn’t stuck waiting for official Keras HDF5s.

Key highlights

• Pre-trained weights for all eight EfficientNet variants, including standard ImageNet and the stronger Noisy Student semi-supervised checkpoints.
• Converted weights hit the expected top-1 accuracy marks on ImageNet validation; B7 reaches 0.844 (standard) and 0.869 (Noisy Student).
• Supports both legacy Keras and tf.keras workflows via separate module paths (efficientnet.keras and efficientnet.tfkeras).
• Bundles a shell script to convert original TensorFlow TPU checkpoints into Keras-compatible HDF5 weights.

Caveats

• Models saved before the July 2019 refactor require pinning to version 0.0.4; the update broke backward compatibility.
• Loading a saved model requires first importing the specific module (efficientnet.keras or efficientnet.tfkeras) used during training, because the implementation relies on custom objects.
• The codebase is explicitly described as “heavily inspired by” Google’s TPU reference rather than an independent rewrite.

Verdict Grab it if you need EfficientNet in a Keras or tf.keras pipeline. Skip it if you live in PyTorch or Google’s native TPU ecosystem.