Merge branch 'sokollip-data-stats-support-in-lime'

Author: marcotcr

lime/discretize.py

@@ -18,7 +18,8 @@ class BaseDiscretizer():
 
     __metaclass__ = ABCMeta  # abstract class
 
-    def __init__(self, data, categorical_features, feature_names, labels=None, random_state=None):
+    def __init__(self, data, categorical_features, feature_names, labels=None, random_state=None,
+                 data_stats=None):
         """Initializer
         Args:
             data: numpy 2d array
@@ -31,9 +32,12 @@ def __init__(self, data, categorical_features, feature_names, labels=None, rando
                 column x.
             feature_names: list of names (strings) corresponding to the columns
                 in the training data.
+            data_stats: must have 'means', 'stds', 'mins' and 'maxs', use this
+                if you don't want these values to be computed from data
         """
         self.to_discretize = ([x for x in range(data.shape[1])
-                              if x not in categorical_features])
+                               if x not in categorical_features])
+        self.data_stats = data_stats
         self.names = {}
         self.lambdas = {}
         self.means = {}
@@ -46,6 +50,13 @@ def __init__(self, data, categorical_features, feature_names, labels=None, rando
         bins = self.bins(data, labels)
         bins = [np.unique(x) for x in bins]
 
+        # Read the stats from data_stats if exists
+        if data_stats:
+            self.means = self.data_stats.get("means")
+            self.stds = self.data_stats.get("stds")
+            self.mins = self.data_stats.get("mins")
+            self.maxs = self.data_stats.get("maxs")
+
         for feature, qts in zip(self.to_discretize, bins):
             n_bins = qts.shape[0]  # Actually number of borders (= #bins-1)
             boundaries = np.min(data[:, feature]), np.max(data[:, feature])
@@ -60,6 +71,10 @@ def __init__(self, data, categorical_features, feature_names, labels=None, rando
             self.lambdas[feature] = lambda x, qts=qts: np.searchsorted(qts, x)
             discretized = self.lambdas[feature](data[:, feature])
 
+            # If data stats are provided no need to compute the below set of details
+            if data_stats:
+                continue
+
             self.means[feature] = []
             self.stds[feature] = []
             for x in range(n_bins + 1):
@@ -117,6 +132,31 @@ def get_inverse(q):
         return ret
 
 
+class StatsDiscretizer(BaseDiscretizer):
+    """
+        Class to be used to supply the data stats info when discretize_continuous is true
+    """
+
+    def __init__(self, data, categorical_features, feature_names, labels=None, random_state=None,
+                 data_stats=None):
+
+        BaseDiscretizer.__init__(self, data, categorical_features,
+                                 feature_names, labels=labels,
+                                 random_state=random_state,
+                                 data_stats=data_stats)
+
+    def bins(self, data, labels):
+        bins_from_stats = self.data_stats.get("bins")
+        bins = []
+        if bins_from_stats is not None:
+            for feature in self.to_discretize:
+                bins_from_stats_feature = bins_from_stats.get(feature)
+                if bins_from_stats_feature is not None:
+                    qts = np.array(bins_from_stats_feature)
+                    bins.append(qts)
+        return bins
+
+
 class QuartileDiscretizer(BaseDiscretizer):
     def __init__(self, data, categorical_features, feature_names, labels=None, random_state=None):
 

lime/lime_tabular.py

@@ -16,6 +16,7 @@
 from lime.discretize import DecileDiscretizer
 from lime.discretize import EntropyDiscretizer
 from lime.discretize import BaseDiscretizer
+from lime.discretize import StatsDiscretizer
 from . import explanation
 from . import lime_base
 
@@ -112,7 +113,8 @@ def __init__(self,
                  discretize_continuous=True,
                  discretizer='quartile',
                  sample_around_instance=False,
-                 random_state=None):
+                 random_state=None,
+                 training_data_stats=None):
         """Init function.
 
         Args:
@@ -153,11 +155,21 @@ def __init__(self,
             random_state: an integer or numpy.RandomState that will be used to
                 generate random numbers. If None, the random state will be
                 initialized using the internal numpy seed.
+            training_data_stats: a dict object having the details of training data
+                statistics. If None, training data information will be used, only matters
+                if discretize_continuous is True. Must have the following keys:
+                means", "mins", "maxs", "stds", "feature_values",
+                "feature_frequencies"
         """
         self.random_state = check_random_state(random_state)
         self.mode = mode
         self.categorical_names = categorical_names or {}
         self.sample_around_instance = sample_around_instance
+        self.training_data_stats = training_data_stats
+
+        # Check and raise proper error in stats are supplied in non-descritized path
+        if self.training_data_stats:
+            self.validate_training_data_stats(self.training_data_stats)
 
         if categorical_features is None:
             categorical_features = []
@@ -169,6 +181,12 @@ def __init__(self,
 
         self.discretizer = None
         if discretize_continuous:
+            # Set the discretizer if training data stats are provided
+            if self.training_data_stats:
+                discretizer = StatsDiscretizer(training_data, self.categorical_features,
+                                               self.feature_names, labels=training_labels,
+                                               data_stats=self.training_data_stats)
+
             if discretizer == 'quartile':
                 self.discretizer = QuartileDiscretizer(
                         training_data, self.categorical_features,
@@ -188,7 +206,10 @@ def __init__(self,
                                  ''' 'decile', 'entropy' or a''' +
                                  ''' BaseDiscretizer instance''')
             self.categorical_features = list(range(training_data.shape[1]))
-            discretized_training_data = self.discretizer.discretize(
+
+            # Get the discretized_training_data when the stats are not provided
+            if(self.training_data_stats is None):
+                discretized_training_data = self.discretizer.discretize(
                     training_data)
 
         if kernel_width is None:
@@ -203,21 +224,27 @@ def kernel(d, kernel_width):
 
         self.feature_selection = feature_selection
         self.base = lime_base.LimeBase(kernel_fn, verbose, random_state=self.random_state)
-        self.scaler = None
         self.class_names = class_names
+
+        # Though set has no role to play if training data stats are provided
+        self.scaler = None
         self.scaler = sklearn.preprocessing.StandardScaler(with_mean=False)
         self.scaler.fit(training_data)
         self.feature_values = {}
         self.feature_frequencies = {}
 
         for feature in self.categorical_features:
-            if self.discretizer is not None:
-                column = discretized_training_data[:, feature]
-            else:
-                column = training_data[:, feature]
+            if training_data_stats is None:
+                if self.discretizer is not None:
+                    column = discretized_training_data[:, feature]
+                else:
+                    column = training_data[:, feature]
 
-            feature_count = collections.Counter(column)
-            values, frequencies = map(list, zip(*(sorted(feature_count.items()))))
+                feature_count = collections.Counter(column)
+                values, frequencies = map(list, zip(*(sorted(feature_count.items()))))
+            else:
+                values = training_data_stats["feature_values"][feature]
+                frequencies = training_data_stats["feature_frequencies"][feature]
 
             self.feature_values[feature] = values
             self.feature_frequencies[feature] = (np.array(frequencies) /
@@ -229,6 +256,17 @@ def kernel(d, kernel_width):
     def convert_and_round(values):
         return ['%.2f' % v for v in values]
 
+    @staticmethod
+    def validate_training_data_stats(training_data_stats):
+        """
+            Method to validate the structure of training data stats
+        """
+        stat_keys = list(training_data_stats.keys())
+        valid_stat_keys = ["means", "mins", "maxs", "stds", "feature_values", "feature_frequencies"]
+        missing_keys = list(set(valid_stat_keys) - set(stat_keys))
+        if len(missing_keys) > 0:
+            raise Exception("Missing keys in training_data_stats. Details:" % (missing_keys))
+
     def explain_instance(self,
                          data_row,
                          predict_fn,
@@ -414,8 +452,8 @@ def __data_inverse(self,
         categorical_features = range(data_row.shape[0])
         if self.discretizer is None:
             data = self.random_state.normal(
-                    0, 1, num_samples * data_row.shape[0]).reshape(
-                    num_samples, data_row.shape[0])
+                0, 1, num_samples * data_row.shape[0]).reshape(
+                num_samples, data_row.shape[0])
             if self.sample_around_instance:
                 data = data * self.scaler.scale_ + data_row
             else:

lime/tests/test_lime_tabular.py

@@ -1,17 +1,19 @@
 import unittest
 
 import numpy as np
-import sklearn # noqa
+import collections
+import sklearn  # noqa
 import sklearn.datasets
 import sklearn.ensemble
-import sklearn.linear_model # noqa
+import sklearn.linear_model  # noqa
 from numpy.testing import assert_array_equal
 from sklearn.datasets import load_iris, make_classification
 from sklearn.ensemble import RandomForestClassifier
 from sklearn.linear_model import Lasso
 from sklearn.linear_model import LinearRegression
 from lime.discretize import QuartileDiscretizer, DecileDiscretizer, EntropyDiscretizer
 
+
 try:
     from sklearn.model_selection import train_test_split
 except ImportError:
@@ -577,6 +579,72 @@ def testFeatureValues(self):
         assert_array_equal(explainer.feature_frequencies[1], np.array([.25, .25, .25, .25]))
         assert_array_equal(explainer.feature_frequencies[2], np.array([.5, .5]))
 
+    def test_lime_explainer_with_data_stats(self):
+        np.random.seed(1)
+
+        rf = RandomForestClassifier(n_estimators=500)
+        rf.fit(self.train, self.labels_train)
+        i = np.random.randint(0, self.test.shape[0])
+
+        # Generate stats using a quartile descritizer
+        descritizer = QuartileDiscretizer(self.train, [], self.feature_names, self.target_names,
+                                          random_state=20)
+
+        d_means = descritizer.means
+        d_stds = descritizer.stds
+        d_mins = descritizer.mins
+        d_maxs = descritizer.maxs
+        d_bins = descritizer.bins(self.train, self.target_names)
+
+        # Compute feature values and frequencies of all columns
+        cat_features = np.arange(self.train.shape[1])
+        discretized_training_data = descritizer.discretize(self.train)
+
+        feature_values = {}
+        feature_frequencies = {}
+        for feature in cat_features:
+            column = discretized_training_data[:, feature]
+            feature_count = collections.Counter(column)
+            values, frequencies = map(list, zip(*(feature_count.items())))
+            feature_values[feature] = values
+            feature_frequencies[feature] = frequencies
+
+        # Convert bins to list from array
+        d_bins_revised = {}
+        index = 0
+        for bin in d_bins:
+            d_bins_revised[index] = bin.tolist()
+            index = index+1
+
+        # Descritized stats
+        data_stats = {}
+        data_stats["means"] = d_means
+        data_stats["stds"] = d_stds
+        data_stats["maxs"] = d_maxs
+        data_stats["mins"] = d_mins
+        data_stats["bins"] = d_bins_revised
+        data_stats["feature_values"] = feature_values
+        data_stats["feature_frequencies"] = feature_frequencies
+
+        data = np.zeros((2, len(self.feature_names)))
+        explainer = LimeTabularExplainer(
+            data, feature_names=self.feature_names, random_state=10,
+            training_data_stats=data_stats, training_labels=self.target_names)
+
+        exp = explainer.explain_instance(self.test[i],
+                                         rf.predict_proba,
+                                         num_features=2,
+                                         model_regressor=LinearRegression())
+
+        self.assertIsNotNone(exp)
+        keys = [x[0] for x in exp.as_list()]
+        self.assertEqual(1,
+                         sum([1 if 'petal width' in x else 0 for x in keys]),
+                         "Petal Width is a major feature")
+        self.assertEqual(1,
+                         sum([1 if 'petal length' in x else 0 for x in keys]),
+                         "Petal Length is a major feature")
+
 
 if __name__ == '__main__':
     unittest.main()