User guide

Build an Edge AI workflow with the components you need. Choose a task below for practical guidance, then use the API reference for exact parameters and contracts.

Build the model

Choose an architecture

Compare model families for signals, images and task-specific workloads.

Configure your model

Define construction parameters and keep configuration separate from trained weights.

Import weights

Load weights trained in ONNX, safetensors or PyTorch through a checked mapping pinned to the source file.

Prepare the data

Preprocessing and augmentation

Select transforms and keep signals, targets and masks aligned.

Build an input pipeline

Read records with Grain for either backend, or turn Python generators into TensorFlow datasets.

Work with structured samples Extend preprocessing

Train and evaluate

Training and callbacks

Combine Keras training with progress reporting and specialized objectives.

Metrics and evaluation

Choose classification or signal metrics and inspect model predictions.

Train a masked autoencoder

Save and deploy

Save and load models

Preserve weights and configuration, and restore custom Keras components.

Export and quantization

Convert a TensorFlow-backed model to LiteRT and validate the converted predictions.

Reference and support

Find a specific component, check backend boundaries or set up a contribution.

Browse Python APIs Check backend support Development setup