Example: Implement a container model adapter

The following is an example model adapter defined for a container-backed model. It assumes the image used is a simple flask server listening at the /mirror endpoint to take in a request object with "text" as the only field. The model adapter will return that text verbatim in the field "returnedText" with a response object.

You can view the full definition of the container_context object in the API: ModelAdapter reference documentation.

Copied!
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
import requests
import json
import pandas as pd

import palantir_models as pm


class ExampleIdentityFunctionModelAdapter(pm.ContainerModelAdapter):
    """
        :display-name: Example Identity Function Model Adapter
        :description: Reference example of a model adapter for container-backed model
    """
    def __init__(self, shared_volume_path, model_host_and_port):
        self.shared_volume_path = shared_volume_path
        self.model_host_and_port = model_host_and_port

    @classmethod
    def init_container(cls, container_context):
        shared_volume_path = container_context.shared_empty_dir_mount_path
        # Note this adapter only expects one container name with one provided service URI.
        model_host_and_port = list(container_context.services.values())[0][0]
        return cls(shared_volume_path, model_host_and_port)

    @classmethod
    def api(cls):
        inputs = {"input_df": pm.Pandas(columns=[("text", str)])}
        outputs = {"output_df": pm.Pandas(columns=[("text", str), ("returnedText", str)])}
        return inputs, outputs

    def predict(self, input_df):
        def run_inference_on_row(row):
            request = {"text": row.text}
            response = requests.post("http://" + self.model_host_and_port + "/mirror", json=request)
            json_res = json.loads(response.content.decode("utf-8"))
            return (row.text, json_res["returnedText"])

        results = [run_inference_on_row(row) for row in input_df.itertuples()]
        columns = ["text", "returnedText"]
        return pd.DataFrame(results, columns=columns)