Use the encoders

For encoder users. You run Perception's gateway locally and call encoders over HTTP from your own code.

Before you start

• MedOmni installed, and Docker.
• Inputs in a supported family. Each encoder reads exactly one family; see Encoders and Families.
• A GPU for some encoders. Each encoder's model card says what it needs.

Steps

1. Build the image of each encoder you need. The check builds it and runs it on the encoder's test data.

   medomni perception check perception/encoders/ecg-12lead-500hz-mv/ecgfounder
   

2. Start the gateway on the built encoders, pointing it at the folder that holds your inputs. The folder is mounted read-only into every encoder container.

   medomni perception up --data-root ~/samples
   

   It listens on 127.0.0.1:8000.

3. List the encoders and read a model card before you rely on an encoder.

   curl "localhost:8000/encoders?family=ecg-12lead-500hz-mv"
   

   curl localhost:8000/cards/ecg-12lead-500hz-mv__ecgfounder
   

4. Send an input and read the evidence. The path is relative to the data root.

   curl -X POST localhost:8000/encode -H "Content-Type: application/json" -d '{"encoder_id": "ecg-12lead-500hz-mv__ecgfounder", "path": "ecg/record.hea"}'
   

Check

A successful encode returns evidence: the encoder's outputs with their provenance, and the reliability and limits from the model card. The first call to an encoder waits for its container to start; the response reports the wait in runtime.container_start_ms. The request and response shapes are in Gateway API.

Where next

• Encoders: the model card format and every encoder in this version.
• Add an encoder: package a new encoder for a family.