added cntk to tested keras backends (#50)

* added cntk to tested keras backends

* added travis_wait to cntk install

* sudo required

* added cntk compatibility to most of keras model and keras model tests

* assertions

* CNTK keras backend is not yet complete, removed from tests for now

"Currently not supported: Gradient as symbolic ops"
source: https://docs.microsoft.com/en-us/cognitive-toolkit/Using-CNTK-with-Keras

* added no-coverage annotation for cntk-specific code

.travis.yml

@@ -1,4 +1,4 @@
-sudo: false
+sudo: required
 dist: trusty
 language: python
 python:
@@ -25,8 +25,19 @@ before_install:
     - if [[ $TRAVIS_PYTHON_VERSION == 2.7 ]]; then pip install http://download.pytorch.org/whl/cu75/torch-0.1.12.post2-cp27-none-linux_x86_64.whl; fi
     - if [[ $TRAVIS_PYTHON_VERSION == 3.5 ]]; then pip install http://download.pytorch.org/whl/cu75/torch-0.1.12.post2-cp35-cp35m-linux_x86_64.whl; fi
     - 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
+    - if [[ $TRAVIS_PYTHON_VERSION == 2.7 ]]; then travis_wait travis_retry pip install https://cntk.ai/PythonWheel/CPU-Only/cntk-2.0-cp27-cp27mu-linux_x86_64.whl; fi
+    - if [[ $TRAVIS_PYTHON_VERSION == 3.5 ]]; then travis_wait travis_retry pip install https://cntk.ai/PythonWheel/CPU-Only/cntk-2.0-cp35-cp35m-linux_x86_64.whl; fi
+    - if [[ $TRAVIS_PYTHON_VERSION == 3.6 ]]; then travis_wait travis_retry pip install https://cntk.ai/PythonWheel/CPU-Only/cntk-2.0-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 open mpi for cntk
+    - mkdir /tmp/mpi
+    - pushd /tmp/mpi
+    - wget http://cntk.ai/PythonWheel/ForKeras/depends/openmpi_1.10-3.zip
+    - unzip ./openmpi_1.10-3.zip
+    - sudo dpkg -i openmpi_1.10-3.deb
+    - popd
 install:
     - pip install -e .
 script:

foolbox/models/keras.py

@@ -1,5 +1,4 @@
 from __future__ import absolute_import
-import warnings
 import numpy as np
 
 from .base import DifferentiableModel
@@ -41,9 +40,6 @@ def __init__(
 
         from keras import backend as K
 
-        if K.backend() != 'tensorflow':  # pragma: no cover
-            warnings.warn('Your keras backend might not be supported.')
-
         if predicts == 'probs':
             predicts = 'probabilities'
         assert predicts in ['probabilities', 'logits']
@@ -79,7 +75,13 @@ def __init__(
             # theano always returns the gradient itself (and requires
             # that loss is a single scalar tensor)
             assert isinstance(grads, list)
+            assert len(grads) == 1
+            grad = grads[0]
+        elif K.backend() == 'cntk':  # pragma: no cover
+            assert isinstance(grads, list)
+            assert len(grads) == 1
             grad = grads[0]
+            grad = K.reshape(grad, (1,) + grad.shape)
         else:
             assert not isinstance(grads, list)
             grad = grads

foolbox/tests/test_models_keras.py

@@ -1,4 +1,5 @@
 import pytest
+import warnings
 
 import numpy as np
 from keras.layers import GlobalAveragePooling2D
@@ -18,13 +19,15 @@ def test_keras_model(num_classes):
     channels = num_classes
 
     model = Sequential()
-    model.add(GlobalAveragePooling2D(
-        data_format='channels_last', input_shape=(5, 5, channels)))
+    with warnings.catch_warnings():
+        warnings.filterwarnings("ignore", category=DeprecationWarning)
+        model.add(GlobalAveragePooling2D(
+            data_format='channels_last', input_shape=(5, 5, channels)))
 
-    model = KerasModel(
-        model,
-        bounds=bounds,
-        predicts='logits')
+        model = KerasModel(
+            model,
+            bounds=bounds,
+            predicts='logits')
 
     test_images = np.random.rand(2, 5, 5, channels).astype(np.float32)
     test_label = 7
@@ -53,25 +56,27 @@ def test_keras_model_probs(num_classes):
     bounds = (0, 255)
     channels = num_classes
 
-    inputs = Input(shape=(5, 5, channels))
-    logits = GlobalAveragePooling2D(
-        data_format='channels_last')(inputs)
-    probs = Activation(softmax)(logits)
+    with warnings.catch_warnings():
+        warnings.filterwarnings("ignore", category=DeprecationWarning)
+        inputs = Input(shape=(5, 5, channels))
+        logits = GlobalAveragePooling2D(
+            data_format='channels_last')(inputs)
+        probs = Activation(softmax)(logits)
 
-    model1 = KerasModel(
-        Model(inputs=inputs, outputs=logits),
-        bounds=bounds,
-        predicts='logits')
+        model1 = KerasModel(
+            Model(inputs=inputs, outputs=logits),
+            bounds=bounds,
+            predicts='logits')
 
-    model2 = KerasModel(
-        Model(inputs=inputs, outputs=probs),
-        bounds=bounds,
-        predicts='probabilities')
+        model2 = KerasModel(
+            Model(inputs=inputs, outputs=probs),
+            bounds=bounds,
+            predicts='probabilities')
 
-    model3 = KerasModel(
-        Model(inputs=inputs, outputs=probs),
-        bounds=bounds,
-        predicts='probs')
+        model3 = KerasModel(
+            Model(inputs=inputs, outputs=probs),
+            bounds=bounds,
+            predicts='probs')
 
     np.random.seed(22)
     test_images = np.random.rand(2, 5, 5, channels).astype(np.float32)
@@ -98,28 +103,30 @@ def test_keras_model_preprocess():
     bounds = (0, 255)
     channels = num_classes
 
-    inputs = Input(shape=(5, 5, channels))
-    logits = GlobalAveragePooling2D(
-        data_format='channels_last')(inputs)
+    with warnings.catch_warnings():
+        warnings.filterwarnings("ignore", category=DeprecationWarning)
+        inputs = Input(shape=(5, 5, channels))
+        logits = GlobalAveragePooling2D(
+            data_format='channels_last')(inputs)
 
-    preprocessing = (np.arange(num_classes)[None, None],
-                     np.random.uniform(size=(5, 5, channels)) + 1)
+        preprocessing = (np.arange(num_classes)[None, None],
+                         np.random.uniform(size=(5, 5, channels)) + 1)
 
-    model1 = KerasModel(
-        Model(inputs=inputs, outputs=logits),
-        bounds=bounds,
-        predicts='logits')
+        model1 = KerasModel(
+            Model(inputs=inputs, outputs=logits),
+            bounds=bounds,
+            predicts='logits')
 
-    model2 = KerasModel(
-        Model(inputs=inputs, outputs=logits),
-        bounds=bounds,
-        predicts='logits',
-        preprocessing=preprocessing)
+        model2 = KerasModel(
+            Model(inputs=inputs, outputs=logits),
+            bounds=bounds,
+            predicts='logits',
+            preprocessing=preprocessing)
 
-    model3 = KerasModel(
-        Model(inputs=inputs, outputs=logits),
-        bounds=bounds,
-        predicts='logits')
+        model3 = KerasModel(
+            Model(inputs=inputs, outputs=logits),
+            bounds=bounds,
+            predicts='logits')
 
     np.random.seed(22)
     test_images = np.random.rand(2, 5, 5, channels).astype(np.float32)
@@ -147,18 +154,20 @@ def test_keras_model_gradients():
     bounds = (0, 255)
     channels = num_classes
 
-    inputs = Input(shape=(5, 5, channels))
-    logits = GlobalAveragePooling2D(
-        data_format='channels_last')(inputs)
+    with warnings.catch_warnings():
+        warnings.filterwarnings("ignore", category=DeprecationWarning)
+        inputs = Input(shape=(5, 5, channels))
+        logits = GlobalAveragePooling2D(
+            data_format='channels_last')(inputs)
 
-    preprocessing = (np.arange(num_classes)[None, None],
-                     np.random.uniform(size=(5, 5, channels)) + 1)
+        preprocessing = (np.arange(num_classes)[None, None],
+                         np.random.uniform(size=(5, 5, channels)) + 1)
 
-    model = KerasModel(
-        Model(inputs=inputs, outputs=logits),
-        bounds=bounds,
-        predicts='logits',
-        preprocessing=preprocessing)
+        model = KerasModel(
+            Model(inputs=inputs, outputs=logits),
+            bounds=bounds,
+            predicts='logits',
+            preprocessing=preprocessing)
 
     eps = 1e-3
 
@@ -168,8 +177,9 @@ 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]
-    l2 = model._loss_fn([test_image[None] + eps / 2 * g1, [test_label]])[0]
+    test_label_array = np.array([test_label])
+    l1 = model._loss_fn([test_image[None] - eps / 2 * g1, test_label_array])[0]
+    l2 = model._loss_fn([test_image[None] + eps / 2 * g1, test_label_array])[0]
 
     assert 1e5 * (l2 - l1) > 1