ExecuTorch deployment block

A custom Edge Impulse deployment block (enterprise) that packages a trained impulse into an ExecuTorch program (.pte) with a small Python runtime harness.

Because ExecuTorch consumes PyTorch programs while Edge Impulse emits TensorFlow/TFLite/ONNX, the block converts the exported ONNX model to PyTorch (via onnx2torch) and lowers it to a .pte with the chosen backend. If conversion tooling is unavailable, it still packages model.onnx plus a conversion note so a deploy never hard-fails.

Compatible Learn Blocks Classification, Timeseries, Object Detection FOMO & KWS

Name	link
Timeseries	https://github.com/edgeimpulse/executorch-pytorch-timeseries-block
Classification	https://github.com/edgeimpulse/executorch-pytorch-classification-block
FOMO	https://github.com/edgeimpulse/executorch-pytorch-object-detection-fomo-block
KWS	https://github.com/edgeimpulse/executorch-pytorch-kws-block

Deploy - Run on Android :

repo - https://github.com/edgeimpulse/executorch-android-app

Sample apks - https://github.com/edgeimpulse/executorch-android-app/releases/edit/v1.0.0-samples

Files

File	Purpose
parameters.json	Deploy block metadata + the backend selector shown in Studio.
build.py	Entrypoint. Reads deployment-metadata.json, converts, and writes deploy.zip.
app/run_pte.py	Runtime harness: loads a .pte and runs one forward pass.
app/convert.py	Offline ONNX → .pte converter (same logic as the block).
Dockerfile	Builds the block container.
requirements.txt	Pinned dependencies (onnx, onnx2torch, torch, executorch).

Block interface

• Input (--metadata): deployment-metadata.json, which points to folders.input (SDK + trained model in several formats) and folders.output.
• Parameter: --backend (xnnpack or portable).
• Output: deploy.zip in folders.output, containing model.pte (when converted), model.onnx, labels.txt, and the app/ harness.

Test locally

# Download the deployment inputs for a project into ./input
edge-impulse-blocks runner --download-data input/

docker build -t executorch-deploy .
docker run --rm -v "$PWD":/home executorch-deploy \
    --metadata /home/input/deployment-metadata.json

The resulting deploy.zip is written to the output directory reported in the metadata.

Push to Edge Impulse

edge-impulse-blocks init   # first time only; choose "Deployment block"
edge-impulse-blocks push

It then appears as a Custom block option on the project Deployment page.

Status

This is an MVP. The ONNX → PyTorch → ExecuTorch path works for common CNN graphs; exotic ops may need a custom partitioner or manual conversion via app/convert.py. Validate on your target model before relying on it in production.