Model adapter API

The model adapter's api() method specifies the expected inputs and outputs in order to execute this model adapter's inference logic. Inputs and outputs are specified separately.

At runtime, the model adapter's predict() method is called with the specified inputs.

Example api() implementation

The following example shows an API specifying one input, named input_dataframe, and one output, named output_dataframe. Both the input and output objects are specified as Pandas dataframes, where the input dataframe has one column of float type named input_feature, and the output dataframe has two columns: (1) a column named input_feature of float type and (2) a column named output_feature of int type.

Copied!
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
import palantir_models as pm


class ExampleModelAdapter(pm.ModelAdapter):
    ...

    @classmethod
    def api(cls):
        inputs = {
            "input_dataframe": pm.Pandas(columns=[("input_feature", float)])
        }
        outputs = {
            "output_dataframe": pm.Pandas(columns=[("input_feature", float), ("prediction", float)])
        }
        return inputs, outputs

    ...

The API definition can also be extended to support multiple inputs or outputs of arbitrary types:

Copied!
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
import palantir_models as pm


class ExampleModelAdapter(pm.ModelAdapter):
    ...

    @classmethod
    def api(cls):
        inputs = {
            "input_dataframe": pm.Pandas(columns=[("input_feature", float)]),
            "input_parameter": pm.Parameter(float, default=1.0)
        }
        outputs = {
            "output_dataframe": pm.Pandas(columns=[("input_feature", float), ("prediction", float)])
        }
        return inputs, outputs

    ...

API types

The types of inputs and outputs for the model adapter API can be specified with the following classes, defined in detail below:

• pm.Pandas, for Pandas Dataframes (*)
• pm.Spark, for Spark Dataframes (*)
• pm.Parameter, for constant, single-valued parameters
• pm.FileSystem, for Foundry Dataset filesystem access
• pm.MediaReference, for use with Media References
• pm.Object, for use with Ontology Objects
• pm.ObjectSet, for use with Onology Object Sets

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
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
# The following classes are accessible via `palantir_models` or `pm`

class Pandas:
    def __init__(self, columns: List[Union[str, Tuple[str, type]]]):
        """
        Defines an Pandas Dataframe input or output. Column name and type definitions can be specified as a parameter of this type.
        """

class Spark:
    def __init__(self, columns: List[Union[str, Tuple[str, type]]] = []):
        """
        Defines an Spark Dataframe (pyspark.sql.Dataframe) input or output. Column name and type definitions can be specified as a parameter of this type.
        """

class Parameter:
    def __init__(self, type: type = Any, default = None):
        """
        Defines a constant single-valued parameter input or output. The type of this parameter (default Any) and default value can be specified as parameters of this type.
        """

class FileSystem:
    def __init__(self):
        """
        Defines a FileSystem access input or output object. This type is only usable if the model adapter's `transform()` or `transform_write()` method is called with Foundry Dataset objects.

        If used as an input, the FileSystem representation of the dataset is returned.

        If used as an output, an object containing an `open()` method is used to write files to the output dataset.

        Note that FileSystem outputs are only usable via calling `.transform_write()`.
        """

class MediaReference:
    def __init__(self):
        """
        Defines an input object to be of MediaReference type. This input expects either a stringified JSON representation or a dictionary representation of a media reference object.

        This type is not supported as an API output.
        """

class Object:
    def __init__(self):
        """
        Defines an input object to be of Object type. This input expects either a primary key of the specified Object type, or an instance of the Object type itself.

        This type is not supported as an API output.
        """

class ObjectSet:
    def __init__(self):
        """
        Defines an input object to be of ObjectSet type. This input expects either an object set rid for the specified Object type, or an instance of an ObjectSet itself.

        This type is not supported as an API output.
        """

(*) Review the requirements described below on using these types when publishing the model as a function.

Specifying tabular columns

For Pandas or Spark inputs and outputs, columns can be specified as either a list of strings specifying the column name, or a list of two-object tuples in the format (<name>, <type>) where <name> is a string representing the column name and <type> is a Python type representing the type of the data in the column. If a string is provided for a column definition, its type will default to Any.

Column types

The following types are supported for tabular columns:

• str
• int
• float
• bool
• list (*)
• dict(*)
• typing.Any (*)
• MediaReference

Column types are not enforced and act as a way to signal to consumers of this model adapter what the expected column types are. The only exception is the MediaReference type, which expects each element in the column to be a media reference string and will convert each element to a MediaReference object before being passed to this model adapter's inference logic.

(*) Review the requirements described below on using these types when publishing the model as a function.

Parameter types

For Parameter inputs and outputs, the following types are supported:

• str
• int
• float
• bool
• list(*)
• dict(*)
• typing.Any(*)

Parameter types are enforced, and any parameter input to model.transform() that does not correspond to the designated type will throw a runtime error.

(*) Review the requirements described below on using these types when publishing the model as a function.

Object and ObjectSet types

For Object or ObjectSet inputs, the object type is specified when defining the input in the model adapter API. This object type will be imported from an Ontology SDK generated for a chosen Ontology:

Copied!
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
from ontology_sdk.ontology.objects import ExampleRestaurant


class ExampleModelAdapter(pm.ModelAdapter):
    ...

    @classmethod
    def api(cls):
        inputs = {
            "input_object": pm.Object(ExampleRestaurant),
            "input_object_set": pm.ObjectSet(ExampleRestaurant)
        }
        outputs = {
            "output_dataframe": pm.Pandas(columns=[("input_feature", float), ("prediction", float)])
        }
        return inputs, outputs

    def predict(self, input_object: ExampleRestaurant, input_object_set: ExampleRestaurantObjectSet):
        outputs = ...
        return outputs
    ...

Example predict() implementation

This example uses Pandas dataframes as inputs and outputs, alongside a parameter.

Copied!
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
class ExampleModelAdapter(pm.ModelAdapter):
    ...

    @classmethod
    def api(cls):
        inputs = {
            "input_dataframe": pm.Pandas(columns=[("input_feature", float)]),
            "input_parameter": pm.Parameter(float, default=1.0)
        }
        outputs = {
            "output_dataframe": pm.Pandas(columns=[("input_feature", float), ("prediction", float)])
        }
        return inputs, outputs

    def predict(self, input_dataframe, input_parameter):
        outputs["prediction"] = self.model.predict(input_dataframe) * input_parameter
        return outputs
    ...

• For tabular outputs, the list of columns specified in the API should only contain the required columns. At runtime, columns that were not specified in the API that are passed to the dataframes will be available in the predict method's input and can be returned as outputs. In the example above, passing a dataframe with columns input_feature and extra_column will return a dataframe with columns input_feature, extra_column and prediction.
• Some models require that only columns used during training are passed to their prediction method. Therefore, we recommend only extracting the feature columns to pass to the model.
• Some models require the ordering of columns to be preserved. When inputs are passed via REST API requests as JSON objects, the ordering of columns is not necessarily preserved. Therefore, we recommend reordering the columns before passing them to the model's inference method.

API definition requirements

Requirements for Objectives batch deployment

Direct setup of batch deployment and automatic model evaluation in the Modeling Objectives application is only compatible with models that have a single tabular dataset input. If your model adapter requires several inputs, you can set up batch inference in a Python transform.

Requirements for Objectives evaluation

If using Evaluation in Modeling Objectives, the output should always contain the model prediction and the label and ensure that the label or other required columns are not dropped in the predict method.

Requirements for Direct Function Publishing and Model use in Functions

In order for the model to support being published as a function, the following must be true:

• The Model API can only contain Parameter, Tabular, Object, or ObjectSet inputs.
• The API must not contain the Typing.Any type.
• For tabular inputs, the API must specify the required input or output columns. Simply using pm.Pandas() or pm.Spark() is not allowed as it would be implicitly interpreted a TypeScript Array<any>.
• Any collection type such as list or dict must specify element types (for example, using list[str] or dict[str, str]). Element types are otherwise interpreted as being of the any type.