Build for the edge.
Make it your own.
Build Edge AI models for deployment on Ambiq silicon. Combine reusable Keras architectures, preprocessing, metrics and training components in your own workflow, from preparing data to exporting your model.
TensorFlow and PyTorch through Keras. Support varies by component.
A WORKSPACE FOR YOUR NEXT STEP
Find your way into the code.
Start with your task or explore a stage. Each example connects to a practical guide.
Give your model a consistent input.
Normalize signals and compose transforms with explicit training behavior. Keep selected targets and masks aligned with the data.
Prepare your dataimport keras
from helia_edge.layers.preprocessing import Normalization1D
signals = keras.ops.ones((2, 128, 1))
normalize = Normalization1D(mean=0., variance=4.)
inputs = normalize(signals)A small tensor example. Select your Keras backend before importing.
A model family. Your configuration.
Start with a configurable architecture, then make the input shape and output classes your own. Keep the parameters alongside the trained weights.
Choose an architecturefrom helia_edge.models import ModelSpec, build, compact_tcn_params
params = compact_tcn_params(filters=8, num_classes=2)
spec = ModelSpec(params=params, input_shape=(240, 14))
model = build(spec)Constructs a model with fresh weights. Training data comes from your task.
Keep your Keras training workflow.
Use compile() and fit() with the components your task needs. Add progress reporting or explore a specialized training objective.
Explore trainingfrom helia_edge.callbacks import TQDMProgressBar
history = model.fit(
train_x, train_y,
validation_data=(validation_x, validation_y),
callbacks=[TQDMProgressBar()],
epochs=10, verbose=0,
)Uses your compiled model and training/validation arrays.
Measure what matters to your task.
Inspect classification errors or signal reconstruction quality. Choose metrics for the meaning of your outputs, not just a single aggregate score.
Choose a metricfrom helia_edge.metrics import Snr
metric = Snr()
metric.update_state(reference_signals, predicted_signals)
print(metric.result())
metric.reset_state()SNR example for aligned reference and predicted signals.
Take a measured step toward inference.
Export a supported TensorFlow-backed model to LiteRT, inspect its predictions and keep the settings with the artifact.
Export a modelfrom helia_edge.export import ExportSpec, export_model
spec = ExportSpec(precision="a8w8", io_dtype="int8",
mode="concrete")
result = export_model(model, spec,
calibration=representative_x.astype("float32"))
print(result.sha256)Requires a trained TensorFlow-backed model and representative inputs.
BUILT TO WORK TOGETHER
One toolkit. No fixed recipe.
Use a single component or connect a workflow. Your application owns the data and training recipe.
CHOOSE YOUR ENVIRONMENT
TensorFlow and PyTorch.
Through Keras.
Choose the backend before importing Keras. Portable components and selected training paths span both backends. TensorFlow provides the LiteRT conversion and dataset utilities.
FROM READING TO BUILDING