Call a Model API¶

Once your model is deployed, you can call it using the requests library. MLFlow builds an API endpoint for your model, the endpoints are the following:

• POST /invocations: An inference endpoint that accepts POST requests with input data and returns predictions.
• GET /ping: Used for health checks.
• GET /health: Same as /ping
• GET /version: Returns the MLflow version.

Check the model status¶

First, we need to check if the model is running. We can do this by calling the /ping endpoint.

In [1]:

import requests

ENDPOINT = "http://localhost:5001/ping"

response = requests.get(ENDPOINT)
response

import requests ENDPOINT = "http://localhost:5001/ping" response = requests.get(ENDPOINT) response

Out[1]:

<Response [200]>

Check the model version¶

We can also check the model version by calling the /version endpoint.

In [2]:

import requests

ENDPOINT = "http://127.0.0.1:5001/version"

response = requests.get(ENDPOINT)
response.text

import requests ENDPOINT = "http://127.0.0.1:5001/version" response = requests.get(ENDPOINT) response.text

Out[2]:

'2.17.0'

Call the model¶

The model is ready to receive requests. We can call the /invocations endpoint with the input data to get the predictions. Let's go step by step:

1. Define the endpoint URL.¶

In [3]:

# Define the URL and payload (JSON data)
ENDPOINT = 'http://localhost:5001/invocations'

# Define the URL and payload (JSON data) ENDPOINT = 'http://localhost:5001/invocations'

2. Prepare the data to be sent¶

Now we prepare the data to be sent to the model. When sending data remember we have 2 parts:

• Body: also called payload, it is the data we want to send to the model. The data should be in JSON format (in python, it is a dictionary). The JSON has a single key inputs and the value is a list of vectors. Each vector has 8 elements, those are the features of the model. The model will return a list of predictions, one for each input vector. The body is
• Headers: we need to specify the content type of the data we are sending. In this case, it is application/json.

In [4]:

# we build de body (payload) of the request
headers = {'Content-Type': 'application/json'}
features = [[0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8]]  # list of lists (vectors)
body = {'inputs': features}

# we build de body (payload) of the request headers = {'Content-Type': 'application/json'} features = [[0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8]] # list of lists (vectors) body = {'inputs': features}

3. Convert the data to JSON¶

We need to convert the data to JSON format. We can use the json library to do this.

In [5]:

import json

# Convert the payload to JSON format
body_json = json.dumps(body)

import json # Convert the payload to JSON format body_json = json.dumps(body)

4. Send the request¶

Now we send the request (POST) to the model using the requests library. The response will be a JSON with the predictions.

In [6]:

# Make a POST request
response = requests.post(ENDPOINT, headers=headers, data=body_json)
response.json()

# Make a POST request response = requests.post(ENDPOINT, headers=headers, data=body_json) response.json()

Out[6]:

{'predictions': [-37.34314240221075]}

We can check that the status code of the response is 200, which means the request was successful.

In [7]:

# Check the response
response.status_code

# Check the response response.status_code

Out[7]:

200