Collecting data without model

inference without model? Is that possible?¶

Inference offers a way to expose models stub - which will not produce any meaningful predictions, but can be used for several purposes: * initial integration on your end with inference serving * collecting dataset via inference Active Learning capabilities

How stubs work?¶

Simply, create workspace and project at Roboflow platform. Once you are done - use the client to send request to the API:

import cv2
from inference_sdk import InferenceHTTPClient

CLIENT = InferenceHTTPClient(
    api_url="http://localhost:9001",  # if inference docker container is running locally
    api_key="XXX"
)

image = cv2.imread(...)
CLIENT.infer(image, model_id="YOUR-PROJECT-NAME/0")   # use version "0" to denote that you want stub model

As a result - you will receive the following response:

{
    "time": 0.0002442499971948564,
    "is_stub": true,
    "model_id": "asl-poly-instance-seg/0",
    "task_type": "instance-segmentation"
}

You should not rely on response format, as it will change once you train and deploy a model, but utilising stubs let you avoid integration cold start.