Support modern Keras 3.x

CHANGES.rst

@@ -1,6 +1,12 @@
 Changelog
 =========
 
+0.16.0 (2024-04-??)
+-------------------
+
+* support explain_prediction with Grad-CAM for Keras 3.x, TF 2.x image classifiers
+  (use versions prior to 0.14 for TF 1.x support)
+
 0.15.0 (2025-04-06)
 -------------------
 

README.rst

@@ -25,6 +25,9 @@ classifiers and explain their predictions.
 .. image:: https://raw.githubusercontent.com/eli5-org/eli5/refs/heads/master/docs/source/static/readme-show-prediction.png
    :alt: explain_prediction for text data
 
+.. image:: https://raw.githubusercontent.com/eli5-org/eli5/refs/heads/master/docs/source/static/gradcam-catdog.png
+   :alt: explain_prediction for image data
+
 .. image:: https://raw.githubusercontent.com/eli5-org/eli5/refs/heads/master/docs/source/static/readme-show-weights.png
    :alt: explain_weights for text data
 
@@ -39,6 +42,8 @@ It provides support for the following machine learning frameworks and packages:
   It also allows to debug scikit-learn pipelines which contain
   HashingVectorizer, by undoing hashing.
 
+* Keras_ - explain predictions of image classifiers via Grad-CAM visualizations.
+
 * xgboost_ - show feature importances and explain predictions of XGBClassifier,
   XGBRegressor and xgboost.Booster.
 
@@ -80,6 +85,7 @@ and formatting on a client.
 .. _xgboost: https://github.com/dmlc/xgboost
 .. _LightGBM: https://github.com/Microsoft/LightGBM
 .. _Catboost: https://github.com/catboost/catboost
+.. _Keras: https://keras.io/
 .. _Permutation importance: https://eli5.readthedocs.io/en/latest/blackbox/permutation_importance.html
 .. _Inspecting Black-Box Estimators: https://eli5.readthedocs.io/en/latest/blackbox/index.html
 .. _OpenAI: https://github.com/openai/openai-python

docs/source/libraries/keras.rst

@@ -3,13 +3,11 @@
 Keras
 =====
 
-Keras_ is "a high-level neural networks API, written in Python and capable of running on top of TensorFlow, CNTK, or Theano". 
+Keras_ is a high-level neural networks API, written in Python and capable of running on top of TensorFlow and other backends.
 
 Keras can be used for many Machine Learning tasks, and it has support for both popular
 and experimental neural network architectures.
 
-Note: only TensorFlow 1.x is supported, recommended Keras version is 2.3.1 or earlier, and eli5 version 0.13 or earlier, as you can't install TensorFlow 1.x on Python 3.9+ which is required for eli5 0.14+
-
 .. _Keras: https://keras.io/
 
 

docs/update-notebooks.sh

@@ -77,7 +77,7 @@ mv ../notebooks/keras-image-classifiers.rst \
 rm -r source/_notebooks/keras-image-classifiers_files
 mv ../notebooks/keras-image-classifiers_files/ \
     source/_notebooks/
-sed -i 's&.. image:: keras-image-classifiers_files/&.. image:: ../_notebooks/keras-image-classifiers_files/&g' \
+sed -i '' 's/image:: keras-image-classifiers_files/image:: ..\/_notebooks\/keras-image-classifiers_files/g' \
     source/_notebooks/keras-image-classifiers.rst
 
 

eli5/keras/explain_prediction.py

@@ -1,12 +1,8 @@
-# -*- coding: utf-8 -*-
-from __future__ import absolute_import
-from typing import Union, Optional, Callable, Tuple, List, TYPE_CHECKING
+from typing import Union, Optional, Callable, TYPE_CHECKING
 if TYPE_CHECKING:
     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 (
@@ -30,13 +26,12 @@
 
 # note that keras.models.Sequential subclasses keras.models.Model
 @explain_prediction.register(Model)
-def explain_prediction_keras(model, # type: Model
-                             doc, # type: np.ndarray
-                             targets=None, # type: Optional[list]
-                             layer=None, # type: Optional[Union[int, str, Layer]]
+def explain_prediction_keras(model: Model,
+                             doc: np.ndarray,
+                             targets: Optional[list] = None,
+                             layer: Optional[Union[int, str, Layer]] = None,
                              image=None,
-                             ):
-    # type: (...) -> Explanation
+                             ) -> Explanation:
     """
     Explain the prediction of a Keras classifier with the Grad-CAM technique.
 
@@ -133,7 +128,7 @@ def explain_prediction_keras_not_supported(model, doc):
 
 def explain_prediction_keras_image(model,
                                    doc,
-                                   image=None, # type: Optional['PIL.Image.Image']
+                                   image: Optional['PIL.Image.Image'] = None,
                                    targets=None,
                                    layer=None,
                                    ):
@@ -204,23 +199,20 @@ def explain_prediction_keras_image(model,
     )
 
 
-def _maybe_image(model, doc):
-    # type: (Model, np.ndarray) -> bool
+def _maybe_image(model: Model, doc: np.ndarray) -> bool:
     """Decide whether we are dealing with a image-based explanation 
     based on heuristics on ``model`` and ``doc``."""
     return _maybe_image_input(doc) and _maybe_image_model(model)
 
 
-def _maybe_image_input(doc):
-    # type: (np.ndarray) -> bool
+def _maybe_image_input(doc: np.ndarray) -> bool:
     """Decide whether ``doc`` represents an image input."""
     rank = len(doc.shape)
     # image with channels or without (spatial only)
     return rank == 4 or rank == 3
 
 
-def _maybe_image_model(model):
-    # type: (Model) -> bool
+def _maybe_image_model(model: Model) -> bool:
     """Decide whether ``model`` is used for images."""
     # FIXME: replace try-except with something else
     try:
@@ -239,22 +231,19 @@ def _maybe_image_model(model):
                       )
 
 
-def _is_possible_image_model_layer(model, layer):
-    # type: (Model, Layer) -> bool
+def _is_possible_image_model_layer(model: Model, layer: Layer) -> bool:
     """Check that the given ``layer`` is usually used for images."""
     return isinstance(layer, image_model_layers)
 
 
-def _extract_image(doc):
-    # type: (np.ndarray) -> 'PIL.Image.Image'
+def _extract_image(doc: np.ndarray) -> 'PIL.Image.Image':
     """Convert ``doc`` tensor to image."""
     im_arr, = doc  # rank 4 batch -> rank 3 single image
     image = array_to_img(im_arr)
     return image
 
 
-def _validate_doc(model, doc):
-    # type: (Model, np.ndarray) -> None
+def _validate_doc(model: Model, doc: np.ndarray) -> None:
     """
     Check that the input ``doc`` is suitable for ``model``.
     """
@@ -277,8 +266,7 @@ def _validate_doc(model, doc):
                              'input: {}, doc: {}'.format(input_sh, doc_sh))
 
 
-def _get_activation_layer(model, layer):
-    # type: (Model, Union[None, int, str, Layer]) -> Layer
+def _get_activation_layer(model: Model, layer: Union[None, int, str, Layer]) -> Layer:
     """
     Get an instance of the desired activation layer in ``model``,
     as specified by ``layer``.
@@ -306,8 +294,7 @@ def _get_activation_layer(model, layer):
         raise ValueError('Can not perform Grad-CAM on the retrieved activation layer')
 
 
-def _search_layer_backwards(model, condition):
-    # type: (Model, Callable[[Model, Layer], bool]) -> Layer
+def _search_layer_backwards(model: Model, condition: Callable[[Model, Layer], bool]) -> Layer:
     """
     Search for a layer in ``model``, backwards (starting from the output layer),
     checking if the layer is suitable with the callable ``condition``,
@@ -321,8 +308,7 @@ def _search_layer_backwards(model, condition):
     raise ValueError('Could not find a suitable target layer automatically.')        
 
 
-def _is_suitable_activation_layer(model, layer):
-    # type: (Model, Layer) -> bool
+def _is_suitable_activation_layer(model: Model, layer: Layer) -> bool:
     """
     Check whether the layer ``layer`` matches what is required 
     by ``model`` to do Grad-CAM on ``layer``.
@@ -337,6 +323,8 @@ def _is_suitable_activation_layer(model, layer):
     # check layer name
 
     # a check that asks "can we resize this activation layer over the image?"
-    rank = len(layer.output_shape)
+    # Use the tensor shape of the layer's output
+    output_shape = layer.output.shape
+    rank = len(output_shape)
     required_rank = len(model.input_shape)
     return rank == required_rank