Getting Started¶

Getting started is as easy as:

from porter.datascience import WrappedModel
from porter.services import ModelApp, PredictionService

my_model = WrappedModel.from_file('my-model.pkl')
prediction_service = PredictionService(
    model=my_model,
    name='my-model',
    api_version='v1')

app = ModelApp([prediction_service])
app.run()

Now just send a POST request to the endpoint /my-model/v1/prediction to get a prediction. Behind the scenes porter will convert the POST data to a pandas.DataFrame, pass this to my_model.predict() and return the result to the user as formatted below:

{
    "model_context": {
        "api_version": "v1",
        "model_meta": {},
        "model_name": "my-model"
    },
    "predictions": [
        {
            "id": 1,
            "prediction": 0
        }
    ],
    "request_id": "0f86644edee546ee9c495a9a71b0746c"
}

The model can be any Python object with a .predict(X) method, where X is a DataFrame and the return value is a sequence with one element per row of X.

WrappedModel.from_file() supports .pkl files via joblib and .h5 files for keras models.

Multiple models can be served by a single app simply by passing additional services to porter.services.ModelApp.

Error handling comes for free when exposing models with ModelApp. For example, by default, if the POST data sent to the prediction endpoint can’t be parsed the user will receive a response with a 400 status code and a payload describing the error.

{
    "error": {
        "messages": [
            "The browser (or proxy) sent a request that this server could not understand."
        ],
        "name": "BadRequest"
    },
    "model_context": {
        "api_version": "v1",
        "model_meta": {},
        "model_name": "my-model"
    },
    "request_id": "852ca09d578b447aa3d41d70b8cc4431"
}