Skip to content

Input formats accepted by inference server

Why should I care?

The Roboflow team has designed the inference server to be as user-friendly and straightforward to integrate as possible. We understand that some users prioritize ease of use, which is why we did not restrict the use of potentially less secure data loading methods. This approach caters to those who prefer a simple and accessible serving mechanism without the need for rigorous security measures.

However, we also recognize the importance of having a production-ready solution. Therefore, we offer configuration options that allow users to disable potentially unsafe behaviors.

In this document, we explain how to configure the server to either enhance security or enable more flexible behaviors, depending on your needs.

Deserialization of pickled numpy objects

One of the ways to send requests to the inference server is via serialized numpy objects:

import cv2
import pickle
import requests

image = cv2.imread("...")
img_str = pickle.dumps(image)

infer_payload = {
    "model_id": "{project_id}/{model_version}",
    "image": {
        "type": "numpy",
        "value": img_str,
    },
    "api_key": "YOUR-API-KEY",
}

res = requests.post(
    "http://localhost:9001/infer/{task}",
    json=infer_payload,
)

Starting from version v0.14.0, deserialization of this type of payload is disabled by default. However, you can enable it by setting an environmental variable, ALLOW_NUMPY_INPUT=True. Check inference cli docs to see how to run the server with that flag. This option is not available in Roboflow's Hosted Inference API.

Warning

Roboflow advises all users hosting the inference server in production environments not to enable this option if the server is open to requests from the open Internet or is not locked down to accept only authenticated requests from your workspace's API key.

Sending URLs to inference images

Making GET requests to obtain images from URLs can expose the server to server-side request forgery (SSRF) attacks. However, it is also very convenient to simply provide an image URL for requests: 

import requests


infer_payload = {
    "model_id": "{project_id}/{model_version}",
    "image": {
        "type": "numpy",
        "value": "https://some.com/image.jpg",
    },
    "api_key": "YOUR-API-KEY",
}

res = requests.post(
    "http://localhost:9001/infer/{task}",
    json=infer_payload,
)

This option is enabled by default, but we recommend configuring the server to enhance security using one or more of the following environment variables: * ALLOW_URL_INPUT - Set to False disable image URLs of any kind to be accepted by server - default: True. * ALLOW_NON_HTTPS_URL_INPUT - set to False to only allow https protocol in URLs (useful to make sure domain names are not maliciously resolved) - default: False * ALLOW_URL_INPUT_WITHOUT_FQDN - set to False to enforce URLs with fully qualified domain names only - and reject URLs based on IPs - default: False * WHITELISTED_DESTINATIONS_FOR_URL_INPUT - Optionally, you can specify a comma-separated list of allowed destinations for URL requests. For example: WHITELISTED_DESTINATIONS_FOR_URL_INPUT=192.168.0.15,some.site.com. URLs pointing to other targets will be rejected. * BLACKLISTED_DESTINATIONS_FOR_URL_INPUT - Optionally, you can specify a comma-separated list of forbidden destinations for URL requests. For example: BLACKLISTED_DESTINATIONS_FOR_URL_INPUT=192.168.0.15,some.site.com. URLs pointing to these targets will be rejected. * ALLOW_LOADING_IMAGES_FROM_LOCAL_FILESYSTEM - Set to False to disable local filesystem access to images - default: True.

Check inference cli docs to see how to run server with specific flags.