Creating a Predictor

Create and share your predictors.

Function uses Jupyter Notebooks as a self-contained format for creating predictors. Let's create a predictor that says a greeting to a user.

Creating the Predictor

First, create a predictor.ipynb notebook and add this function in a Python cell:

# Function requires your prediction function to be called "predict"
def predict (first_name: str):
    return f"Hello {first_name}!"

Our prediction function accepts a first_name string and returns a greeting. With this, we can create the predictor on Function to run it from anywhere. Open a terminal and run the following command:

# Create the predictor
fxn create @username/greeting predictor.ipynb --overwrite

Once you run this command, Function does three things:

1. Creates a predictor signature by analyzing your code.
2. Invokes the code in your notebook, up until the predict function.
3. Containerizes and deploys your code.

Defining the Predictor Signature

The predictor signature provides information about the input and output values of your prediction function. It is the second most important component of any predictor, second to the actual code.

Function relies solely on code annotations to infer the signature of your predictor. For example, Function will infer the following signature from our predictor above:

{
  "inputs": [
    {
      "name": "first_name",
      "type": "string"
    }
  ],
  "outputs": []
}

Function is also able to infer more of your predictor signature from your prediction function's docstring. For example, let's add an age parameter to our original function:

def predict (first_name: str, age: int):
    """
    Say a greeting to a user.

    Parameters:
        first_name (str): The user's first name.
        age (int): The user's age. This should be more than 13 but less than 80.
    """
    return f"Hello {first_name}! You are {age} years old"

With the added docstring, Function will infer the following signature from the predictor:

{
  "inputs": [
    {
      "name": "first_name",
      "type": "string",
      "description": "The user's first name."
    },
    {
      "name": "age",
      "type": "int32",
      "description": "The user's age. This should be more than 13 but less than 80.",
      "range": [14, 79] // Function infers this from your docstring
    }
  ],
  "outputs": []
}

Defining the Predictor Card

The predictor card is effectively a readme for the predictor. It informs prospective users about what the predictor does, any considerations on its usage, the predictor license, and more. It is shown prominently on fxn.ai ↗, and will be the first thing that users will read.

To define a predictor card, simply add a Markdown cell as the very first cell in your predictor notebook. During provisioning, Function will automatically pull it out as the predictor card.

Installing Dependencies

When creating a prediction function, you will often need to use third-party libraries. Function relies on Jupyter magic commands to install these dependencies:

Installing Python Packages

Recreate the modified predictor from above on Function by opening a terminal and running the following command:

# Recreate the predictor
fxn create @username/greeting predictor.ipynb --overwrite

Once the predictor is active, run the following command to make a prediction:

# Make a prediction with age=1
fxn predict @username/greeting --name "John" --age 1

You should see the following result:

{
  ...,
  "results": [
    "Hello John! You are 1 years old"
  ]
}

We have a typo! We aren't properly pluralizing the word "years". Thankfully, we can use the pluralizer ↗ Python package to fix this.

To install the package, add a code cell before the predict function and add the following:

# Install the pluralizer package
%pip install pluralizer

Then modify the predict function to use the pluralizer:

from pluralizer import Pluralizer

# Create a pluralizer
pluralizer = Pluralizer()

def predict (first_name: str, age: int):
    """
    Say a greeting to a user.

    Parameters:
        first_name (str): The user's first name.
        age (int): The user's age. This should be more than 13 but less than 80.
    """
    # Pluralize "year"
    years_str = pluralizer.pluralize("year", age)
    # Return the greeting
    return f"Hello {first_name}! You are {age} {years_str} old"

You can test the updated predictor right in the notebook. Simply create a new cell and call the predict function:

# Test out the updated predictor
greeting = predict("John", 1)
print(greeting)
#: Hello John! You are 1 year old

Installing System Packages

You can install system packages to be installed while your predictor is being provisioned. Simply use the apt-get package manager tool:

# Install FFmpeg so that our predictor can use it
!apt-get install ffmpeg