Implement Counterplots

Signed-off-by: rmazzine <[email protected]>

dice_ml/counterfactual_explanations.py

@@ -2,9 +2,12 @@
 import os
 
 import jsonschema
+import numpy as np
+import pandas as pd
+from counterplots import CreatePlot
 from raiutils.exceptions import UserConfigValidationException
 
-from dice_ml.constants import _SchemaVersions
+from dice_ml.constants import _SchemaVersions, BackEndTypes
 from dice_ml.diverse_counterfactuals import (CounterfactualExamples,
                                              _DiverseCFV2SchemaConstants)
 
@@ -111,6 +114,54 @@ def visualize_as_list(self, display_sparse_df=True,
                     display_sparse_df=display_sparse_df,
                     show_only_changes=show_only_changes)
 
+    def plot_counterplots(self, dice_model):
+        """Plot counterfactual with CounterPlots package.
+
+        :param dice_model: DiCE's model object.
+        """
+        counterplots_out = []
+        for cf_examples in self.cf_examples_list:
+            features_names = list(cf_examples.test_instance_df.columns)[:-1]
+            features_dtypes = list(cf_examples.test_instance_df.dtypes)[:-1]
+            factual_instance = cf_examples.test_instance_df.to_numpy()[0][:-1]
+
+            def convert_data(x):
+                df_x = pd.DataFrame(data=x, columns=features_names)
+                # Transform each dtype according to features_dtypes
+                for feature_name, f_dtype in zip(features_names, features_dtypes):
+                    df_x[feature_name] = df_x[feature_name].astype(f_dtype)
+
+                return df_x
+
+            if dice_model.backend == BackEndTypes.Sklearn:
+                factual_class_idx = np.argmax(
+                    dice_model.model.predict_proba(convert_data([factual_instance])))
+                def model_pred(x):
+                    # Use one against all strategy
+                    pred_prob = dice_model.model.predict_proba(convert_data(x))
+                    class_f_proba = pred_prob[:, factual_class_idx]
+
+                    # Probability for all other classes (excluding class 0)
+                    not_class_f_proba = 1 - class_f_proba
+
+                    # Normalize to sum to 1
+                    class_f_proba = class_f_proba / (class_f_proba + not_class_f_proba)
+
+                    return class_f_proba
+            else:
+                def model_pred(x):
+                    return dice_model.model.predict(dice_model.transformer.transform(convert_data(x)))
+
+            for cf_instance in cf_examples.final_cfs_df.to_numpy():
+                counterplots_out.append(
+                    CreatePlot(
+                        factual=factual_instance,
+                        cf=cf_instance[:-1],
+                        model_pred=model_pred,
+                        feature_names=features_names,
+                    ))
+        return counterplots_out
+
     @staticmethod
     def _check_cf_exp_output_against_json_schema(
             cf_dict, version):

requirements.txt

@@ -4,3 +4,4 @@ pandas<2.0.0
 scikit-learn
 tqdm
 raiutils>=0.4.0
+counterplots>=0.0.7

tests/test_counterfactual_explanations.py

@@ -1,8 +1,12 @@
 import json
 
 import pytest
+import unittest
+from unittest.mock import patch, Mock
 from raiutils.exceptions import UserConfigValidationException
 
+import pandas as pd
+import numpy as np
 from dice_ml.counterfactual_explanations import CounterfactualExplanations
 
 
@@ -319,3 +323,78 @@ def test_unsupported_versions_to_json(self, unsupported_version):
             counterfactual_explanations.to_json()
 
         assert "Unsupported serialization version {}".format(unsupported_version) in str(ucve)
+
+
+class TestCounterfactualExplanations(unittest.TestCase):
+
+    @patch('dice_ml.counterfactual_explanations.CreatePlot', return_value="dummy_plot")
+    def test_plot_counterplots_sklearn(self, mock_create_plot):
+        # Dummy DiCE's model object with a Sklearn backend
+        dummy_model = Mock()
+        dummy_model.backend = "sklearn"
+        dummy_model.model.predict_proba = Mock(return_value=np.array([[0.4, 0.6], [0.2, 0.8]]))
+
+        # Sample cf_examples to test with
+        cf_examples_mock = Mock()
+        cf_examples_mock.test_instance_df = pd.DataFrame({
+            'feature1': [1],
+            'feature2': [2],
+            'target': [0]
+        })
+        cf_examples_mock.final_cfs_df = pd.DataFrame({
+            'feature1': [1.1, 1.2],
+            'feature2': [2.1, 2.2],
+            'target': [1, 1]
+        })
+
+        counterfact = CounterfactualExplanations(
+            cf_examples_list=[cf_examples_mock],
+            local_importance=None,
+            summary_importance=None,
+            version=None)
+
+        # Call function
+        result = counterfact.plot_counterplots(dummy_model)
+
+        # Assert the CreatePlot was called twice (as there are 2 counterfactual instances)
+        self.assertEqual(mock_create_plot.call_count, 2)
+
+        # Assert that the result is as expected
+        self.assertEqual(result, ["dummy_plot", "dummy_plot"])
+
+    @patch('dice_ml.counterfactual_explanations.CreatePlot', return_value="dummy_plot")
+    def test_plot_counterplots_non_sklearn(self, mock_create_plot):
+        # Sample Non-Sklearn backend
+        dummy_model = Mock()
+        dummy_model.backend = "NonSklearn"
+        dummy_model.model.predict = Mock(return_value=np.array([0, 1]))
+        dummy_model.transformer = Mock()
+        dummy_model.transformer.transform = Mock(return_value=np.array([[1, 2], [1.1, 2.1]]))
+
+        # Sample cf_examples to test with
+        cf_examples_mock = Mock()
+        cf_examples_mock.test_instance_df = pd.DataFrame({
+            'feature1': [1],
+            'feature2': [2],
+            'target': [0]
+        })
+        cf_examples_mock.final_cfs_df = pd.DataFrame({
+            'feature1': [1.1, 1.2],
+            'feature2': [2.1, 2.2],
+            'target': [1, 1]
+        })
+
+        counterfact = CounterfactualExplanations(
+            cf_examples_list=[cf_examples_mock],
+            local_importance=None,
+            summary_importance=None,
+            version=None)
+
+        # Call function
+        result = counterfact.plot_counterplots(dummy_model)
+
+        # Assert the CreatePlot was called twice (as there are 2 counterfactual instances)
+        self.assertEqual(mock_create_plot.call_count, 2)
+
+        # Assert that the result is as expected
+        self.assertEqual(result, ["dummy_plot", "dummy_plot"])