The SynapRT Python package allows you to run real-time AI pipelines on your Synaptics Astra board in just a few lines of code. Built on the official SyNAP Python API, SynapRT leverages NPU acceleration for efficient, low-latency inference.

[Optional] Create a venv to manage dependencies:

python3 -m venv .venv --system-site-packages
source .venv/bin/activate

Next, install the latest release directly on your Astra board:

pip install https://github.com/synaptics-synap/synap-rt/releases/download/v0.0.2-preview/synap_rt-0.0.2-py3-none-any.whl

A SynapRT pipeline requires three key components: a task, a SyNAP model, and an input source. Additionally, it is recommended to define a handler function to process inference results or integrate them into a broader workflow.

"""A simple object detection pipeline example."""

import json
import sys

from synapRT.pipelines import pipeline


def handle_inference_result(results, inference_time=None):
    # simple handler that writes the results to stdout
    sys.stdout.write("\033[H\033[J")
    formatted_json = json.dumps(results, indent=4)
    print(formatted_json, flush=True)
    sys.stdout.flush()
    if inference_time:
        print(f"Avg. inference time: {inference_time} ms", flush=True)
    sys.stdout.flush()

# define pipeline
pipe = pipeline(
    task="object-detection",
    model="/usr/share/synap/models/object_detection/coco/model/yolov8s-640x384/model.synap",
    profile=True, # enable inference time profiling
    handler=handle_inference_result,
)

# run pipeline with inputs
pipe(sys.argv[1])

Run the above pipeline with:

python3 simple.py <input>

<input> can be image(s), or a video source like camera, MP4 file, or RTSP stream.

Explore the examples directory for additional pipeline use cases, such as:

• Running a pipeline in a separate thread to avoid blocking execution.
• Polling a pipeline for results instead of using a handler function.

For application-level implementations using SynapRT pipelines, check out the official Astra examples repository.