Support for TensorFlow Keras

eli5/formatters/image.py

@@ -10,12 +10,12 @@
 
 
 def format_as_image(expl, # type: Explanation
-    resampling_filter=Image.LANCZOS, # type: int
+    resampling_filter=Image.Resampling.LANCZOS, # type: int
     colormap=matplotlib.cm.viridis, # type: Callable[[np.ndarray], np.ndarray]
     alpha_limit=0.65, # type: Optional[Union[float, int]]
     ):
     # type: (...) -> Image
-    """format_as_image(expl, resampling_filter=Image.LANCZOS, colormap=matplotlib.cm.viridis, alpha_limit=0.65)
+    """format_as_image(expl, resampling_filter=Image.Resampling.LANCZOS, colormap=matplotlib.cm.viridis, alpha_limit=0.65)
 
     Format a :class:`eli5.base.Explanation` object as an image.
 
@@ -50,7 +50,7 @@ def format_as_image(expl, # type: Explanation
 
         *Note that these attributes are integer values*.
 
-        Default is ``PIL.Image.LANCZOS``.
+        Default is ``PIL.Image.Resampling.LANCZOS``.
     :type resampling_filter: int, optional
 
     :param colormap:
@@ -239,7 +239,7 @@ def _cap_alpha(alpha_arr, alpha_limit):
                         'got: {}'.format(alpha_limit))
 
 
-def expand_heatmap(heatmap, image, resampling_filter=Image.LANCZOS):
+def expand_heatmap(heatmap, image, resampling_filter=Image.Resampling.LANCZOS):
     # type: (np.ndarray, Image, Union[None, int]) -> Image
     """
     Resize the ``heatmap`` image array to fit over the original ``image``,
@@ -286,4 +286,4 @@ def _overlay_heatmap(heatmap, image):
     """
     # note that the order of alpha_composite arguments matters
     overlayed_image = Image.alpha_composite(image, heatmap)
-    return overlayed_image
+    return overlayed_image

eli5/keras/explain_prediction.py

@@ -5,18 +5,28 @@
     import PIL
 
 import numpy as np
-import keras
-import keras.backend as K
-from keras.models import Model
-from keras.layers import Layer
-from keras.layers import (
-    Conv2D,
-    MaxPooling2D,
-    AveragePooling2D,
-    GlobalMaxPooling2D,
-    GlobalAveragePooling2D,
-)
-from keras.preprocessing.image import array_to_img
+import os
+if'TF_KERAS' in os.environ and os.environ['TF_KERAS'] == '1':
+    from tensorflow import keras
+else:
+    import keras
+K = keras.backend
+Model = keras.models.Model
+Layer = keras.layers.Layer
+Conv2D = keras.layers.Conv2D
+MaxPooling2D = keras.layers.MaxPooling2D
+AveragePooling2D = keras.layers.AveragePooling2D
+GlobalMaxPooling2D = keras.layers.GlobalMaxPooling2D
+GlobalAveragePooling2D = keras.layers.GlobalAveragePooling2D
+
+try:
+    # tensorflow<2.9
+    array_to_img = keras.preprocessing.image.array_to_img
+except:
+    # tensorflow>=2.9
+    # reference: https://www.tensorflow.org/api_docs/python/tf/keras/utils/array_to_img
+    from tensorflow.keras.utils import array_to_img
+
 
 from eli5.base import Explanation, TargetExplanation
 from eli5.explain import explain_prediction

eli5/keras/gradcam.py

@@ -3,10 +3,14 @@
 from typing import Union, Optional, Tuple, List
 
 import numpy as np
-import keras
-import keras.backend as K
-from keras.models import Model
-from keras.layers import Layer
+import os
+if 'TF_KERAS' in os.environ and os.environ['TF_KERAS'] == '1':
+    from tensorflow import keras
+else:
+    import keras
+K = keras.backend
+Model = keras.models.Model
+Layer = keras.layers.Layer
 
 
 def gradcam(weights, activations):

eli5/lime/lime.py

@@ -27,6 +27,8 @@
 )
 from eli5.lime._vectorizer import SingleDocumentVectorizer
 
+from packaging.version import parse 
+
 
 class TextExplainer(BaseEstimator):
     """
@@ -320,12 +322,12 @@ def _fix_target_names(self, kwargs):
 
     def _default_clf(self):
         kwargs = dict(
-            loss='log',
+            loss='log_loss',
             penalty='elasticnet',
             alpha=1e-3,
             random_state=self.rng_
         )
-        if sklearn_version() >= '0.19':
+        if sklearn_version() >= parse('0.19'):
             kwargs['tol'] = 1e-3
         return SGDClassifier(**kwargs)
 

eli5/lime/samplers.py

@@ -6,7 +6,6 @@
 import six
 
 import numpy as np
-from scipy.stats import itemfreq
 from sklearn.base import BaseEstimator, clone
 from sklearn.neighbors import KernelDensity
 from sklearn.metrics import pairwise_distances
@@ -188,7 +187,8 @@ def _sampler_n_samples(self, n_samples):
                                            p=self.weights)
         return [
             (self.samplers[idx], freq)
-            for idx, freq in itemfreq(sampler_indices)
+            # use np.unique due to removal of scipy.stats.itemfreq
+            for idx, freq in np.vstack( np.unique(sampler_indices, return_counts=True) ).transpose()
         ]
 
 

eli5/lime/textutils.py

@@ -14,7 +14,7 @@
 
 
 # the same as scikit-learn token pattern, but allows single-char tokens
-DEFAULT_TOKEN_PATTERN = r'(?u)\b\w+\b'
+DEFAULT_TOKEN_PATTERN = r'\b\w+\b'
 
 # non-whitespace chars
 CHAR_TOKEN_PATTERN = r'[^\s]'
@@ -183,7 +183,7 @@ def __init__(self, parts):
     def fromtext(cls, text, token_pattern=DEFAULT_TOKEN_PATTERN):
         # type: (str, str) -> SplitResult
         token_pattern = u"(%s)" % token_pattern
-        parts = re.split(token_pattern, text)
+        parts = re.split(token_pattern, text, flags=re.UNICODE)
         return cls(parts)
 
     @property

eli5/lime/utils.py

@@ -5,13 +5,16 @@
 import numpy as np
 from scipy.stats import entropy
 from sklearn.pipeline import Pipeline
-from sklearn.utils import check_random_state, issparse
-from sklearn.utils.metaestimators import if_delegate_has_method
+from sklearn.utils import check_random_state
+from scipy.sparse import issparse
+from sklearn.utils.metaestimators import available_if
 from sklearn.utils import shuffle as _shuffle
 
 from eli5.utils import vstack
 from eli5.sklearn.utils import sklearn_version
 
+from packaging.version import parse
+
 
 def fit_proba(clf, X, y_proba, expand_factor=10, sample_weight=None,
               shuffle=True, random_state=None,
@@ -73,7 +76,15 @@ def fix_multiclass_predict_proba(y_proba,          # type: np.ndarray
 class _PipelinePatched(Pipeline):
     # Patch from https://github.com/scikit-learn/scikit-learn/pull/7723;
     # only needed for scikit-learn < 0.19.
-    @if_delegate_has_method(delegate='_final_estimator')
+
+    # Reference: https://github.com/scikit-learn/scikit-learn/issues/20506
+    def _estimator_has(attr):
+        def check(self):
+            return hasattr(self.estimator, attr)
+
+        return check
+
+    @available_if(_estimator_has('_final_estimator'))
     def score(self, X, y=None, **score_params):
         Xt = X
         for name, transform in self.steps[:-1]:
@@ -83,7 +94,7 @@ def score(self, X, y=None, **score_params):
 
 
 def score_with_sample_weight(estimator, X, y=None, sample_weight=None):
-    if sklearn_version() < '0.19':
+    if sklearn_version() < parse('0.19'):
         if isinstance(estimator, Pipeline) and sample_weight is not None:
             estimator = _PipelinePatched(estimator.steps)
     if sample_weight is None:

eli5/sklearn/permutation_importance.py

@@ -4,7 +4,7 @@
 
 import numpy as np
 from sklearn.model_selection import check_cv
-from sklearn.utils.metaestimators import if_delegate_has_method
+from sklearn.utils.metaestimators import available_if
 from sklearn.utils import check_array, check_random_state
 from sklearn.base import (
     BaseEstimator,
@@ -247,23 +247,30 @@ def caveats_(self):
 
     # ============= Exposed methods of a wrapped estimator:
 
-    @if_delegate_has_method(delegate='wrapped_estimator_')
+    # Reference: https://github.com/scikit-learn/scikit-learn/issues/20506
+    def _estimator_has(attr):
+        def check(self):
+            return hasattr(self.estimator, attr)
+
+        return check
+
+    @available_if(_estimator_has('wrapped_estimator_'))
     def score(self, X, y=None, *args, **kwargs):
         return self.wrapped_estimator_.score(X, y, *args, **kwargs)
 
-    @if_delegate_has_method(delegate='wrapped_estimator_')
+    @available_if(_estimator_has('wrapped_estimator_'))
     def predict(self, X):
         return self.wrapped_estimator_.predict(X)
 
-    @if_delegate_has_method(delegate='wrapped_estimator_')
+    @available_if(_estimator_has('wrapped_estimator_'))
     def predict_proba(self, X):
         return self.wrapped_estimator_.predict_proba(X)
 
-    @if_delegate_has_method(delegate='wrapped_estimator_')
+    @available_if(_estimator_has('wrapped_estimator_'))
     def predict_log_proba(self, X):
         return self.wrapped_estimator_.predict_log_proba(X)
 
-    @if_delegate_has_method(delegate='wrapped_estimator_')
+    @available_if(_estimator_has('wrapped_estimator_'))
     def decision_function(self, X):
         return self.wrapped_estimator_.decision_function(X)
 

eli5/sklearn/utils.py

@@ -1,6 +1,7 @@
 # -*- coding: utf-8 -*-
 from __future__ import absolute_import
-from distutils.version import LooseVersion
+#from distutils.version import LooseVersion # deprecated
+from packaging.version import parse as LooseVersion
 from typing import Any, Optional, List, Tuple
 
 import numpy as np
@@ -80,7 +81,9 @@ def get_feature_names(clf, vec=None, bias_name='<BIAS>', feature_names=None,
         bias_name = None
 
     if feature_names is None:
-        if vec and hasattr(vec, 'get_feature_names'):
+        if vec and hasattr(vec, 'get_feature_names_out'):
+            return FeatureNames(vec.get_feature_names_out(), bias_name=bias_name)
+        elif vec and hasattr(vec, 'get_feature_names'):
             return FeatureNames(vec.get_feature_names(), bias_name=bias_name)
         else:
             if estimator_feature_names is None:

tests/test_keras_integration.py

@@ -81,7 +81,7 @@ def assert_attention_over_area(expl, area):
     heatmap = expl.targets[0].heatmap
 
     # fit heatmap over image
-    heatmap = expand_heatmap(heatmap, image, Image.LANCZOS)
+    heatmap = expand_heatmap(heatmap, image, Image.Resampling.LANCZOS)
     heatmap = np.array(heatmap)
 
     # get a slice of the area
@@ -159,4 +159,4 @@ def test_show_prediction_nodeps(show_nodeps, keras_clf, cat_dog_image):
 ])
 def test_explain_prediction_not_supported(model, doc):
     res = eli5.explain_prediction(model, doc)
-    assert 'supported' in res.error
+    assert 'supported' in res.error