KerasHub: a pretrained modeling library
4 months ago
- #Keras
- #Machine Learning
- #Pretrained Models
- KerasHub is a pretrained modeling library that is simple, flexible, and fast.
- Provides Keras 3 implementations of popular model architectures with pretrained checkpoints on Kaggle Models.
- Models can be used for training and inference on TensorFlow, Jax, and Torch backends.
- KerasHub extends the core Keras API, with components provided as keras.layers.Layer and keras.Model implementations.
- Installation: `pip install --upgrade keras-hub` for the latest release or `pip install --upgrade keras-hub-nightly` for nightly changes.
- Currently, installing KerasHub pulls in TensorFlow for tf.data API preprocessing, but training can occur on any backend.
- Example usage includes loading a ResNet model for image classification and a BERT model for fine-tuning on IMDb reviews.
- Follows Semantic Versioning but may break compatibility during pre-release 0.y.z development.
- Pre-trained models are provided 'as is' without warranties.
- Citation available for research use.