.KERAS

Introduction

A .keras file is a whole model saved by Keras, the deep learning library that is also TensorFlow’s high-level interface. It holds the model’s architecture, its weights and, if it was compiled, its training settings, so keras.saving.load_model can rebuild it in one call. It became the default in 2023 and is the only full-model format in Keras 3.

Before that, Keras saved models as HDF5 files with the extension .h5. These are still common in older tutorials and model collections, and Keras can still read them.

Structure

A .keras file is a ZIP archive with three entries:

• config.json: the model’s architecture, layer by layer, with the class name and settings of each
• metadata.json: the Keras version and the date it was saved
• model.weights.h5: the weights, in HDF5

A legacy Keras .h5 file is a single HDF5 file. Its root carries the architecture as a JSON string in the model_config attribute, with keras_version and backend attributes beside it; the weights are in a model_weights group with one subgroup per layer. A file saved with save_weights holds only the weight groups.

Also used for

• Hierarchical Data Format version 5 .h5

Links