Saving and Multi-Backend

## What the archive contains `model.save("model.keras")`, equivalently `keras.saving.save_model(model, "model.keras")`, produces the `.keras` v3 archive. Despite the extension it is an ordinary zip file with three entries: - `metadata.json` - the Keras version that wrote it, plus a timestamp. - `config.json` - the model configuration: a JSON tree of class names and the constructor arguments of every layer, wired into the same graph. - `model.weights.h5` - an HDF5 file of the variable values (kernels, biases, normalization moving statistics, and the optimizer variables of a compiled model). The split matters: the config is code-free JSON describing how to rebuild, and the weights file is plain numeric arrays. Nothing in the archive is executable, which is why a saved model is portable but also why Keras must be able to find your own classes again at load time. ## The round trip ``` model.save("model.keras") restored = keras.saving.load_model("model.keras") ``` `load_model` parses `config.json`, instantiates each class with its recorded arguments, then streams the weights into the freshly created variables. The result is a normal model object: `restored.predict(x)` works, `restored.summary()` prints the same table, and because the optimizer variables travelled too, `restored.fit(...)` continues with the optimizer's accumulated state intact rather than from a cold start. That last point is the practical reason to prefer whole-model saving over weights-only saving when you may need to resume a run. `load_model` takes `compile=False` if you only want to serve the model; the layers and weights are restored, the optimizer and loss are not, and the model comes back uncompiled. ## Extensions and the legacy path Keras 3 validates the filepath. No extension raises a ValueError telling you to add `.keras` or `.h5`. A `.h5` or `.hdf5` suffix still works but takes the legacy HDF5 whole-model path, and Keras warns that the native format is recommended. Legacy HDF5 predates the v3 archive, stores weights by layer name, and carries custom objects far less reliably - treat it as a format you read old files with, not one you write new ones in. There is a third verb: `model.export(...)` writes an inference-only artifact for serving stacks rather than a resumable Keras model. Do not reach for it when what you want is to reload the model in Python. ## What can still go wrong The archive stores class names, not class code. A model built only from built-in Keras layers therefore always reloads. A model containing your own `Layer` subclass reloads only if Keras can map the recorded name back to a Python class at load time - otherwise the load fails while trying to locate the class, and the fix is to register the class or pass it in `custom_objects`. The other common surprise is version skew: an archive written by a much newer Keras may contain config keys an older Keras does not understand. Pin the Keras version alongside the artifact the way you pin any other dependency.