Merge pull request #16 from keisen/issues/#15

Fixes Issues #15 and #8

.gitignore

@@ -107,6 +107,7 @@ venv.bak/
 *.swp
 examples/core*
 examples/sandbox.ipynb
+examples/workspace.ipynb
 /*.ipynb
 .node-version
 *.nbconvert.ipynb

.travis.yml

@@ -17,7 +17,7 @@ jobs:
       env: TF_VERSION=2.2.0 TF_KERAS_VIS_MAX_STEPS=2
 install:
   - pip install -U pip
-  - pip install -U --force-reinstall -e .[development,examples] tensorflow==$TF_VERSION
+  - pip install -U --force-reinstall -e .[develop,examples] tensorflow==$TF_VERSION
 script:
   - PYTHONPATH=$PWD:$PYTHONPATH py.test
   - jupyter-nbconvert --ExecutePreprocessor.timeout=600 --to notebook --execute examples/attentions.ipynb

README.md

@@ -4,14 +4,7 @@
 [![Build Status](https://travis-ci.org/keisen/tf-keras-vis.svg?branch=master)](https://travis-ci.org/keisen/tf-keras-vis)
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
 
-tf-keras-vis is a visualization toolkit for debugging `tf.keras` models in Tensorflow2.
-
-The features of tf-keras-vis are based on [keras-vis](https://github.com/raghakot/keras-vis), but tf-keras-vis's APIs doesn't have compatibility with keras-vis's, because we prioritized to get following features for our expriments.
-
-- Support processing multiple images at a time as a batch
-- Support tf.keras.Model that has multiple inputs (and, of course, multiple outpus too)
-- Allow to use optimizers that embeded in tf.keras
-
+tf-keras-vis is a visualization toolkit for debugging `tf.keras` models in Tensorflow2.0+.
 
 ## Visualizations
 
@@ -25,7 +18,14 @@ The features of tf-keras-vis are based on [keras-vis](https://github.com/raghako
 
 ### Saliency Map and GradCAM
 
-<img src='https://github.com/keisen/tf-keras-vis/raw/master/examples/images/gradcam.png' width='400px' />
+<img src='https://github.com/keisen/tf-keras-vis/raw/master/examples/images/gradcam.png' width='600px' />
+
+
+These features are based on ones of [keras-vis](https://github.com/raghakot/keras-vis), but tf-keras-vis APIs doesn't have compatibility with keras-vis, because we prioritized to get following features for our expriments.
+
+- Support processing multiple images at a time as a batch
+- Support tf.keras.Model that has multiple inputs (and, of course, multiple outpus too)
+- Allow to use optimizers that embeded in tf.keras
 
 
 ## Requirements
@@ -65,15 +65,16 @@ Please see [examples/attentions.ipynb](https://github.com/keisen/tf-keras-vis/bl
 ## Known Issues
 
 * With InceptionV3, ActivationMaximization doesn't work well, that's, it might generate meanninglessly bulr image.
-* With cascading model, Gradcam doesn't work well, that's, it might occur some error.
-* Unsupported `channels-first` models and datas.
+* With cascading model, Gradcam and Gradcam++ don't work well, that's, it might occur some error.
+* Unsupport `channels-first` models and datas.
 
 
 ## ToDo
 * API documentations
 * We're going to add some algorisms such as below.
-   - [GradCAM++](https://arxiv.org/abs/1710.11063)
    - [SmoothGrad: removing noise by adding noise](https://arxiv.org/pdf/1706.03825.pdf) (DONE)
+   - [GradCAM++](https://arxiv.org/abs/1710.11063)
+   - [ScoreCAM](https://arxiv.org/pdf/1910.01279.pdf)
    - Deep Dream
    - Style transfer
 

setup.py

@@ -5,7 +5,7 @@
 
 setup(
     name="tf-keras-vis",
-    version="0.3.1",
+    version="0.3.2",
     author="keisen",
     author_email="[email protected]",
     description="Neural network visualization toolkit for tf.keras",
@@ -21,8 +21,7 @@
     python_requires='>=3.5, <3.9',
     install_requires=['numpy', 'scipy', 'imageio', 'pillow'],
     extras_require={
-        'development':
-        ['flake8', 'isort', 'yapf', 'pytest', 'pytest-pep8', 'pytest-xdist', 'pytest-cov'],
+        'develop': ['flake8', 'isort', 'yapf', 'pytest', 'pytest-pep8', 'pytest-cov'],
         'examples': ['jupyterlab', 'matplotlib'],
     },
     include_package_data=True,

tf_keras_vis/gradcam.py

@@ -15,7 +15,8 @@ def __call__(self,
                  penultimate_layer=-1,
                  seek_penultimate_conv_layer=True,
                  activation_modifier=lambda cam: K.relu(cam),
-                 normalize_gradient=True):
+                 normalize_gradient=True,
+                 expand_cam=True):
         """Generate a gradient based class activation map (CAM) by using positive gradient of
             penultimate_layer with respect to loss.
 
@@ -34,6 +35,10 @@ def __call__(self,
                 If False, the penultimate layer is that was elected by penultimate_layer index.
             normalize_gradient: True to normalize gradients.
             activation_modifier: A function to modify gradients.
+            expand_cam: True to expand cam to same as input image size.
+                ![Note] Even if the model has multiple inputs, this function return only one cam
+                value (That's, when `expand_cam` is True, multiple cam images are generated from
+                a model that has multiple inputs).
         # Returns
             The heatmap image or a list of their images that indicate the `seed_input` regions
                 whose change would most contribute  the loss value,
@@ -60,6 +65,10 @@ def __call__(self,
         cam = np.asarray([np.sum(o * w, axis=-1) for o, w in zip(penultimate_output, weights)])
         if activation_modifier is not None:
             cam = activation_modifier(cam)
+
+        if not expand_cam:
+            return cam
+
         # Visualizing
         cam = self._zoom_for_visualizing(seed_inputs, cam)
         if len(self.model.inputs) == 1 and not isinstance(seed_input, list):

tf_keras_vis/utils/__init__.py

@@ -57,8 +57,8 @@ def listify(value, return_empty_list_if_none=True, convert_tuple_to_list=True):
 
 
 def normalize(array, value_range=(1., 0.)):
-    max_value = np.max(array)
-    min_value = np.min(array)
+    max_value = np.max(array, axis=tuple(np.arange(len(array.shape))[1:]), keepdims=True)
+    min_value = np.min(array, axis=tuple(np.arange(len(array.shape))[1:]), keepdims=True)
     normalized_array = (array - min_value) / (max_value - min_value + K.epsilon())
     if value_range is None:
         return normalized_array