# Save and load models

Save the model instance after training, and keep its architecture and preprocessing configuration alongside it. A parameter JSON file describes construction choices; it does not contain trained weights.

## Save a Keras model

```python
model.save("model.keras")
```

Use the EDGE loader in a fresh process after selecting the appropriate backend:

```python
from helia_edge.models import load_model
model = load_model("model.keras")
```

This loader registers supported EDGE custom objects. A bare `import helia_edge` uses lazy imports and does not register every Keras class as a side effect.

## Use the Keras loader directly

```python
import helia_edge
helia_edge.register_keras_serializables()

import keras
model = keras.saving.load_model("model.keras")
```

Models whose layer normalization runs over spatial axes (TCN, UNet and UNext with `norm="layer"`) save `helia_edge>LayerNormalization`, so plain `keras.saving.load_model` needs this registration first; `helia_edge.models.load_model` registers automatically. On the Torch backend, `helia_edge.models.load_model` also adapts `.keras` files saved by earlier helia-edge versions: it renames names containing `.` (layers, the model, and compile keys), which Torch rejects, and loads Keras `LayerNormalization` over spatial axes as the helia-edge class. Other formats, such as `.h5`, load unchanged.

Registration initializes the selected backend. It does not turn a TensorFlow-only component into a portable component. See the [backend guide](https://ambiqai.github.io/helia-edge/getting-started/backends/#loading-saved-custom-objects) for the tested round-trip boundaries.

## Reconstruct from configuration and weights

Construct the same architecture and input signature, then call `model.load_weights(checkpoint_path)`. Check that graph connections, shapes and layer configuration agree with the checkpoint. Do not use skipped mismatches as a substitute for compatibility.

The [model architectures guide](https://ambiqai.github.io/helia-edge/guide/architectures/#fixed-and-preset-families) explains reading older TCN configurations, and [specs and trained weights](https://ambiqai.github.io/helia-edge/guide/architectures/#specs-and-trained-weights) the handling of historical trained artifacts. Cross-backend loading depends on the model's custom components and naming conventions; preserve the original backend and dependency versions with the checkpoint.

## Include the experiment context

Record the model configuration, preprocessing policy, class or target definitions, package versions and evaluation results. Record seed policy too, but do not assume a seed guarantees identical weights or random sequences across backends.