One API to wrangle two dozen Keras vision models

What it does This library is a unified storefront for pretrained ImageNet classification models in Keras and TensorFlow Keras. It exposes architectures like VGG, ResNet, DenseNet, and NASNet through a single Classifiers.get() call that hands you both the model and its preprocessing logic. Weights are ready to load for standard 1000-class ImageNet, with 11k-class variants available for ResNet50 and ResNet152.

The interesting bit The heavy lifting is weight conversion: checkpoints have been ported from PyTorch, MXNet, and Keras Applications into one consistent Keras format. The README also publishes a full benchmark table—top-1, top-5, and inference time—for every model, evaluated on the same hardware, which saves you from guessing whether SE-ResNeXt50 is worth the latency hit.

Key highlights

• 16+ architectures: ResNeXt, SE-ResNet, SENet, Xception, MobileNet v2, etc.
• Dual framework support: works with standalone keras and tensorflow.keras.
• Weights drawn from three different upstream sources and converted to Keras HDF5.
• Benchmarks cover Acc@1, Acc@5, and wall-clock inference time for direct comparison.
• Extra weight sets: 11k-class ImageNet and ImageNet+Place365 for select ResNets.

Caveats

• The README warns of a breaking update on “05 of August” (year unspecified) and advises pinning an old release to load legacy models.
• The PyPI package is named image-classifiers, while the repository is classification_models; the mismatch is easy to miss.
• TensorFlow is deliberately omitted from requirements, so you must install it manually.

Verdict Handy if you want a single, consistent API to audition ResNet variants or SENet in Keras without wrestling with weight conversions. Skip it if you are already comfortable pulling models directly from torchvision or tf.keras.applications.