arize 8.0.0a9__py3-none-any.whl → 8.0.0a10__py3-none-any.whl

arize/models/client.py

@@ -80,6 +80,11 @@ _BATCH_DEPS = (
     "tqdm",
 )
 _BATCH_EXTRA = "ml-batch"
+_MIMIC_DEPS = (
+    "interpret_community.mimic",
+    "sklearn.preprocessing",
+)
+_MIMIC_EXTRA = "mimic-explainer"
 
 
 class MLModelsClient:
@@ -116,7 +121,6 @@ class MLModelsClient:
         timeout: float | None = None,
     ) -> cf.Future:
         require(_STREAM_EXTRA, _STREAM_DEPS)
-
         from arize._generated.protocol.rec import public_pb2 as pb2
         from arize.utils.proto import (
             get_pb_dictionary,
@@ -545,17 +549,10 @@ class MLModelsClient:
             dataframe = dataframe.astype(cat_str_map)
 
         if surrogate_explainability:
-            logger.debug("Running surrogate_explainability.")
+            require(_MIMIC_EXTRA, _MIMIC_DEPS)
+            from arize.models.surrogate_explainer.mimic import Mimic
 
-            try:
-                # WARNING: MIMIC EXPLAINER IS NOT DONE
-                from arize.pandas.surrogate_explainer.mimic import Mimic
-            except ImportError:
-                raise ImportError(
-                    "To enable surrogate explainability, "
-                    "the arize module must be installed with the MimicExplainer option: pip "
-                    "install 'arize[MimicExplainer]'."
-                ) from None
+            logger.debug("Running surrogate_explainability.")
             if schema.shap_values_column_names:
                 logger.info(
                     "surrogate_explainability=True has no effect "

arize/models/surrogate_explainer/mimic.py

@@ -0,0 +1,164 @@
+from __future__ import annotations
+
+import random
+import string
+from dataclasses import replace
+from typing import TYPE_CHECKING, Callable, Tuple
+
+import numpy as np
+import pandas as pd
+from interpret_community.mimic.mimic_explainer import (
+    LGBMExplainableModel,
+    MimicExplainer,
+)
+from sklearn.preprocessing import LabelEncoder
+
+from arize.types import (
+    CATEGORICAL_MODEL_TYPES,
+    NUMERIC_MODEL_TYPES,
+    ModelTypes,
+)
+
+if TYPE_CHECKING:
+    from arize.types import Schema
+
+
+class Mimic:
+    _testing = False
+
+    def __init__(self, X: pd.DataFrame, model_func: Callable):
+        self.explainer = MimicExplainer(
+            model_func,
+            X,
+            LGBMExplainableModel,
+            augment_data=False,
+            is_function=True,
+        )
+
+    def explain(self, X: pd.DataFrame) -> pd.DataFrame:
+        return pd.DataFrame(
+            self.explainer.explain_local(X).local_importance_values,
+            columns=X.columns,
+            index=X.index,
+        )
+
+    @staticmethod
+    def augment(
+        df: pd.DataFrame, schema: Schema, model_type: ModelTypes
+    ) -> Tuple[pd.DataFrame, Schema]:
+        features = schema.feature_column_names
+        X = df[features]
+
+        if X.shape[1] == 0:
+            return df, schema
+
+        if model_type in CATEGORICAL_MODEL_TYPES:
+            if not schema.prediction_score_column_name:
+                raise ValueError(
+                    "To calculate surrogate explainability, "
+                    f"prediction_score_column_name must be specified in schema for {model_type}."
+                )
+
+            y_col_name = schema.prediction_score_column_name
+            y = df[y_col_name].to_numpy()
+
+            _min, _max = np.min(y), np.max(y)
+            if not 0 <= _min <= 1 or not 0 <= _max <= 1:
+                raise ValueError(
+                    f"To calculate surrogate explainability for {model_type}, "
+                    f"prediction scores must be between 0 and 1, but current "
+                    f"prediction scores range from {_min} to {_max}."
+                )
+
+            # model func requires 1 positional argument
+            def model_func(_):  # type: ignore
+                return np.column_stack((1 - y, y))
+
+        elif model_type in NUMERIC_MODEL_TYPES:
+            y_col_name = schema.prediction_label_column_name
+            if schema.prediction_score_column_name is not None:
+                y_col_name = schema.prediction_score_column_name
+            y = df[y_col_name].to_numpy()
+
+            _finite_count = np.isfinite(y).sum()
+            if len(y) - _finite_count:
+                raise ValueError(
+                    f"To calculate surrogate explainability for {model_type}, "
+                    f"predictions must not contain NaN or infinite values, but "
+                    f"{len(y) - _finite_count} NaN or infinite value(s) are found in {y_col_name}."
+                )
+
+            # model func requires 1 positional argument
+            def model_func(_):  # type: ignore
+                return y
+
+        else:
+            raise ValueError(
+                "Surrogate explainability is not supported for the specified "
+                f"model type {model_type}."
+            )
+
+        # Column name mapping between features and feature importance values.
+        # This is used to augment the schema.
+        col_map = {
+            ft: f"{''.join(random.choices(string.ascii_letters, k=8))}"
+            for ft in features
+        }
+        aug_schema = replace(schema, shap_values_column_names=col_map)
+
+        # Limit the total number of "cells" to 20M, unless it results in too few or
+        # too many rows. This is done to keep the runtime low. Records not sampled
+        # have feature importance values set to 0.
+        samp_size = min(
+            len(X), min(100_000, max(1_000, 20_000_000 // X.shape[1]))
+        )
+
+        if samp_size < len(X):
+            _mask = np.zeros(len(X), dtype=int)
+            _mask[:samp_size] = 1
+            np.random.shuffle(_mask)
+            _mask = _mask.astype(bool)
+            X = X[_mask]
+            y = y[_mask]
+
+        # Replace all pd.NA values with np.nan values
+        for col in X.columns:
+            if X[col].isna().any():
+                X[col] = X[col].astype(object).where(~X[col].isna(), np.nan)
+
+        # Apply integer encoding to non-numeric columns.
+        # Currently training and explaining detasets are the same, but
+        # this can be changed in the future. The student model can be
+        # fitted on a much larger dataset since it takes a lot less time.
+        X = pd.concat(
+            [
+                X.select_dtypes(exclude=[object, "string"]),
+                pd.DataFrame(
+                    {
+                        name: LabelEncoder().fit_transform(data)
+                        for name, data in X.select_dtypes(
+                            include=[object, "string"]
+                        ).items()
+                    },
+                    index=X.index,
+                ),
+            ],
+            axis=1,
+        )
+
+        aug_df = pd.concat(
+            [
+                df,
+                Mimic(X, model_func).explain(X).rename(col_map, axis=1),
+            ],
+            axis=1,
+        )
+
+        # Fill null with zero so they're not counted as missing records by server
+        if not Mimic._testing:
+            aug_df.fillna({c: 0 for c in col_map.values()}, inplace=True)
+
+        return (
+            aug_df,
+            aug_schema,
+        )

arize/version.py

@@ -1 +1 @@
-__version__ = "8.0.0a9"
+__version__ = "8.0.0a10"

{arize-8.0.0a9.dist-info → arize-8.0.0a10.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: arize
-Version: 8.0.0a9
+Version: 8.0.0a10
 Summary: A helper library to interact with Arize AI APIs
 Project-URL: Homepage, https://arize.com
 Project-URL: Documentation, https://docs.arize.com/arize
@@ -30,6 +30,8 @@ Requires-Dist: numpy>=2.0.0
 Provides-Extra: dev
 Requires-Dist: pytest==8.4.2; extra == 'dev'
 Requires-Dist: ruff==0.13.2; extra == 'dev'
+Provides-Extra: mimic-explainer
+Requires-Dist: interpret-community[mimic]<1,>=0.22.0; extra == 'mimic-explainer'
 Provides-Extra: ml-batch
 Requires-Dist: pandas<3,>=1.0.0; extra == 'ml-batch'
 Requires-Dist: protobuf<6,>=4.21.0; extra == 'ml-batch'

{arize-8.0.0a9.dist-info → arize-8.0.0a10.dist-info}/RECORD

@@ -4,7 +4,7 @@ arize/client.py,sha256=0LtZU3WeEatGd1QgQsMrJOuI-tFmzM3y1AfO74BLJys,5716
 arize/config.py,sha256=iynVEZhrOPdTNJTQ_KQmwKOPiwL0LfEP8AUIDYW86Xw,5801
 arize/logging.py,sha256=2vwdta2-kR78GeBFGK2vpk51rQ2d06HoKzuARI9qFQk,7317
 arize/types.py,sha256=z1yg5-brmTD4kVHDmmTVkYke53JpusXXeOOpdQw7rYg,69508
-arize/version.py,sha256=BDT_fj8aJ4pvQMjTUrdbEd6vY71GkAguJScCzty97gA,24
+arize/version.py,sha256=Wv8B6KxzS2ThGtkzs_13OkvwSugf5HITHYMQsGk1gjg,25
 arize/_exporter/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 arize/_exporter/client.py,sha256=eAxJX1sUfdpLrtaQ0ynMTd5jI37JOp9fbl3NWp4WFEA,15216
 arize/_exporter/validation.py,sha256=6ROu5p7uaolxQ93lO_Eiwv9NVw_uyi3E5T--C5Klo5Q,1021
@@ -71,11 +71,13 @@ arize/experiments/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU
 arize/experiments/client.py,sha256=2fDq0fr_h6Knn_9zgDAlAhSUCKUrKozGLOQRTInCr4c,344
 arize/models/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 arize/models/bounded_executor.py,sha256=o-PJsDAXQdiJ9dc-jzGCHMhT0-QBY9bvl4Ckn1017Eo,1131
-arize/models/client.py,sha256=aYXPv5Pq2Va2_aEEptw6-iD5zDEFV4UJz2bPnXvvIHw,31419
+arize/models/client.py,sha256=ZHxGYmCKP5ZX001qVNQc96QoclP4jvYVkLW11Xfqo2M,31199
 arize/models/stream_validation.py,sha256=PtmqWgRdCxVtTNkHHEHIM1S6ECbYLA1vuQQFBw_t3Lw,7118
 arize/models/batch_validation/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 arize/models/batch_validation/errors.py,sha256=__I8l25zf4kGv6qgiwEm9LzGNgqmMSM8Fb88pBtyMxE,39990
 arize/models/batch_validation/validator.py,sha256=acnGcMt-pETmPJUfYj5tIzIBvmBhWoXoWmDYi_Gkq6Y,146910
+arize/models/surrogate_explainer/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+arize/models/surrogate_explainer/mimic.py,sha256=MsMfhU9IhQJWm0kK6jpFkcTW6kw5IGJE3Kv94oOzMo0,5517
 arize/spans/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 arize/spans/client.py,sha256=5yODUaSqxH-dLAenjRZBKbpsK7XgewZKwJpXzHWPNf0,47248
 arize/spans/columns.py,sha256=BbB11jF4YHYfjrKbSd1r3K2F0AGA8KULTj1W3e2rwhM,12912
@@ -107,7 +109,7 @@ arize/utils/arrow.py,sha256=4In1gQc0i4Rb8zuwI0w-Hv-10wiItu5opqqGrJ8tSzo,5277
 arize/utils/casting.py,sha256=KUrPUQN6qJEVe39nxbr0T-0GjAJLHjf4xWuzV71QezI,12468
 arize/utils/dataframe.py,sha256=I0FloPgNiqlKga32tMOvTE70598QA8Hhrgf-6zjYMAM,1120
 arize/utils/proto.py,sha256=9vLo53INYjdF78ffjm3E48jFwK6LbPD2FfKei7VaDy8,35477
-arize-8.0.0a9.dist-info/METADATA,sha256=WGsj9jvMjlNY9Xy4c9GpIT6mkJ1551ITCUYEhhQESck,12453
-arize-8.0.0a9.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-arize-8.0.0a9.dist-info/licenses/LICENSE.md,sha256=8vLN8Gms62NCBorxIv9MUvuK7myueb6_-dhXHPmm4H0,1479
-arize-8.0.0a9.dist-info/RECORD,,
+arize-8.0.0a10.dist-info/METADATA,sha256=9u9UPm9jOeZp9pxLo9R5mDYvrACrOzbPET51mNyyXQU,12567
+arize-8.0.0a10.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+arize-8.0.0a10.dist-info/licenses/LICENSE.md,sha256=8vLN8Gms62NCBorxIv9MUvuK7myueb6_-dhXHPmm4H0,1479
+arize-8.0.0a10.dist-info/RECORD,,