specified keras backends in travis config (#49)

* specified keras backends in travis config

* squeeze singleton dimension

* handle keras backend differences between tensorflow and theano

* dtype fixes for lasagne model

* corrected test name

* set theano.config.floatX to float32

* fixed travis env config

* fixed dtypes in preprocessing

.travis.yml

@@ -5,12 +5,20 @@ python:
     - 2.7
     - 3.5
     - 3.6
+env:
+    global:
+        - THEANO_FLAGS='floatX=float32'
+    matrix:
+        - KERAS_BACKEND=tensorflow
+        - KERAS_BACKEND=theano
 before_install:
     - pip install -U pip
-install:
     - travis_wait travis_retry pip install --upgrade numpy
     - travis_wait travis_retry pip install --upgrade scipy
-    - if [[ $TRAVIS_PYTHON_VERSION == 2.7 ]]; then travis_wait travis_retry pip install -r requirements-dev-py2.txt; else travis_wait travis_retry pip install -r requirements-dev.txt; fi
+    - travis_wait travis_retry pip install -r requirements-dev.txt
+    - if [[ $TRAVIS_PYTHON_VERSION == 2.7 ]]; then pip install -r requirements-dev-py2.txt; fi
+    - if [[ $TRAVIS_PYTHON_VERSION == 3.5 ]]; then pip install -r requirements-dev-py3.txt; fi
+    - if [[ $TRAVIS_PYTHON_VERSION == 3.6 ]]; then pip install -r requirements-dev-py3.txt; fi
     - travis_wait travis_retry pip install --upgrade tensorflow
     - travis_wait travis_retry pip install --upgrade theano
     - travis_wait travis_retry pip install --upgrade https://github.com/Lasagne/Lasagne/archive/master.zip
@@ -19,6 +27,7 @@ install:
     - if [[ $TRAVIS_PYTHON_VERSION == 3.6 ]]; then pip install http://download.pytorch.org/whl/cu75/torch-0.1.12.post2-cp36-cp36m-linux_x86_64.whl; fi
     - travis_wait travis_retry pip install --upgrade keras
     - travis_wait travis_retry pip install --upgrade mxnet==0.10.0
+install:
     - pip install -e .
 script:
     - pytest

foolbox/models/base.py

@@ -38,6 +38,7 @@ def __init__(self, bounds, channel_axis, preprocessing=(0, 1)):
         self._bounds = bounds
         assert channel_axis in [0, 1, 2, 3]
         self._channel_axis = channel_axis
+        assert len(preprocessing) == 2
         self._preprocessing = preprocessing
 
     def __enter__(self):
@@ -52,11 +53,17 @@ def bounds(self):
     def channel_axis(self):
         return self._channel_axis
 
-    def _process_input(self, input):
-        return (input - self._preprocessing[0]) / self._preprocessing[1]
+    def _process_input(self, input_):
+        result = (input_ - self._preprocessing[0]) / self._preprocessing[1]
+        result = result.astype(input_.dtype, copy=False)
+        assert result.dtype == input_.dtype
+        return result
 
     def _process_gradient(self, gradient):
-        return gradient / self._preprocessing[1]
+        result = gradient / self._preprocessing[1]
+        result = result.astype(gradient.dtype, copy=False)
+        assert result.dtype == gradient.dtype
+        return result
 
     @abstractmethod
     def batch_predictions(self, images):

foolbox/models/keras.py

@@ -67,8 +67,22 @@ def __init__(
         loss = K.sparse_categorical_crossentropy(
             predictions, label_input, from_logits=predictions_are_logits)
 
+        # sparse_categorical_crossentropy returns 1-dim tensor,
+        # gradients wants 0-dim tensor (for some backends)
+        loss = K.squeeze(loss, axis=0)
+
         grads = K.gradients(loss, images_input)
-        grad = grads[0]
+        if K.backend() == 'tensorflow':
+            # tensorflow backend returns a list with the gradient
+            # as the only element, even if loss is a single scalar
+            # tensor;
+            # theano always returns the gradient itself (and requires
+            # that loss is a single scalar tensor)
+            assert isinstance(grads, list)
+            grad = grads[0]
+        else:
+            assert not isinstance(grads, list)
+            grad = grads
 
         self._loss_fn = K.function(
             [images_input, label_input],

foolbox/models/lasagne.py

@@ -77,13 +77,15 @@ def predictions_and_gradient(self, image, label):
         gradient = np.squeeze(gradient, axis=0)
         assert predictions.shape == (self.num_classes(),)
         assert gradient.shape == image.shape
+        gradient = gradient.astype(image.dtype, copy=False)
         return predictions, gradient
 
     def gradient(self, image, label):
         label = np.array(label, dtype=np.int32)
         gradient = self._gradient_fn(image[np.newaxis], label[np.newaxis])
         gradient = np.squeeze(gradient, axis=0)
         assert gradient.shape == image.shape
+        gradient = gradient.astype(image.dtype, copy=False)
         return gradient
 
     def num_classes(self):

foolbox/models/theano.py

@@ -74,6 +74,7 @@ def predictions_and_gradient(self, image, label):
         gradient = np.squeeze(gradient, axis=0)
         assert predictions.shape == (self.num_classes(),)
         assert gradient.shape == image.shape
+        assert gradient.dtype == image.dtype
         return predictions, gradient
 
     def gradient(self, image, label):
@@ -83,6 +84,7 @@ def gradient(self, image, label):
         gradient = self._process_gradient(gradient)
         gradient = np.squeeze(gradient, axis=0)
         assert gradient.shape == image.shape
+        assert gradient.dtype == image.dtype
         return gradient
 
     def num_classes(self):

foolbox/tests/test_models_keras.py

@@ -168,8 +168,8 @@ def test_keras_model_gradients():
 
     _, g1 = model.predictions_and_gradient(test_image, test_label)
 
-    l1 = model._loss_fn([test_image[None] - eps / 2 * g1, [test_label]])[0][0]
-    l2 = model._loss_fn([test_image[None] + eps / 2 * g1, [test_label]])[0][0]
+    l1 = model._loss_fn([test_image[None] - eps / 2 * g1, [test_label]])[0]
+    l2 = model._loss_fn([test_image[None] + eps / 2 * g1, [test_label]])[0]
 
     assert 1e5 * (l2 - l1) > 1
 

foolbox/tests/test_models_lasagne.py

@@ -16,7 +16,7 @@ def mean_brightness_net(images):
         logits = GlobalPoolLayer(images)
         return logits
 
-    images_var = T.tensor4('images')
+    images_var = T.tensor4('images', dtype='float32')
     images = InputLayer((None, channels, 5, 5), images_var)
     logits = mean_brightness_net(images)
 
@@ -56,7 +56,7 @@ def mean_brightness_net(images):
         logits = GlobalPoolLayer(images)
         return logits
 
-    images_var = T.tensor4('images')
+    images_var = T.tensor4('images', dtype='float32')
     images = InputLayer((None, channels, 5, 5), images_var)
     logits = mean_brightness_net(images)
 

foolbox/tests/test_models_theano.py

@@ -46,7 +46,7 @@ def mean_brightness_net(images):
 
 
 @pytest.mark.parametrize('num_classes', [10, 1000])
-def test_lasagne_gradient(num_classes):
+def test_theano_gradient(num_classes):
     bounds = (0, 255)
     channels = num_classes
 

requirements-dev-py2.txt

@@ -1,12 +1,2 @@
-numpydoc >= 0.6.0
-sphinx >= 1.6.2
-sphinx-autobuild >= 0.6.0
-sphinx_rtd_theme >= 0.2.4
-twine >= 1.9.1
-pytest >= 3.1.0
-pytest-cov >= 2.5.1
-flake8 >= 3.3.0
-python-coveralls >= 2.9.1
-pillow >= 4.1.1
-flake8 >= 3.3.0
+-r requirements-dev.txt
 mock >= 2.0.0

requirements-dev-py3.txt

@@ -0,0 +1,2 @@
+-r requirements-dev.txt
+mypy >= 0.511

requirements-dev.txt

@@ -6,7 +6,5 @@ twine >= 1.9.1
 pytest >= 3.1.0
 pytest-cov >= 2.5.1
 flake8 >= 3.3.0
-mypy >= 0.511
 python-coveralls >= 2.9.1
 pillow >= 4.1.1
-flake8 >= 3.3.0