Merge Clever fix from dev

Author: Irina Nicolae

src/attacks/fast_gradient.py

@@ -211,7 +211,8 @@ def set_params(self, **kwargs):
         if self.ord not in [np.inf, int(1), int(2)]:
             raise ValueError("Norm order must be either np.inf, 1, or 2.")
 
-        if self.eps <= self.clip_min or self.eps > self.clip_max:
-            raise ValueError('The amount of perturbation has to be in the data range.')
+        if self.clip_min is not None and self.clip_max is not None:
+            if self.eps <= self.clip_min or self.eps > self.clip_max:
+                raise ValueError('The amount of perturbation has to be in the data range.')
 
         return True

src/classifiers/utils.py

@@ -7,9 +7,6 @@
 from keras.models import model_from_json
 from keras.optimizers import SGD
 
-from src.classifiers.cnn import CNN
-from src.classifiers.mlp import MLP
-from src.classifiers.resnet import ResNet
 from src.layers.activations import BoundedReLU
 from src.utils import make_directory
 
@@ -55,6 +52,11 @@ def load_classifier(file_path, weights_name="weights.h5"):
     :param weights_name: name of the file containing the weights
     :return: Classifier
     """
+    from src.classifiers.bnn import BNN
+    from src.classifiers.cnn import CNN
+    from src.classifiers.mlp import MLP
+    from src.classifiers.resnet import ResNet
+
     # Load json and create model
     with open(os.path.join(file_path, "model.json"), "r") as json_file:
         model_json = json_file.read()

src/metrics.py

@@ -2,6 +2,7 @@
 Module implementing varying metrics for assessing model robustness. These fall mainly under two categories:
 attack-dependent and attack-independent.
 """
+from __future__ import absolute_import, division, print_function, unicode_literals
 import config
 
 import numpy as np
@@ -170,10 +171,13 @@ def loss_sensitivity(x, classifier, sess):
     :return: The average loss sensitivity of the model
     :rtype: `float`
     """
+    from src.attacks.attack import class_derivative
+
     x_op = tf.placeholder(dtype=tf.float32, shape=list(x.shape))
     y_pred = classifier.predict(x)
     indices = np.argmax(y_pred, axis=1)
-    grads = [tf.gradients(classifier.model(x_op)[:, i], x_op) for i in range(10)]
+    grads = class_derivative(classifier._get_predictions(x_op, log=True), x_op,
+                             classifier.model.get_output_shape_at(0)[1])
     res = sess.run(grads, feed_dict={x_op: x})
     res = np.asarray([r[0] for r in res])[indices, list(range(x.shape[0]))]
     res = la.norm(res.reshape(res.shape[0], -1), ord=2, axis=1)

src/metrics_unittest.py

@@ -1,7 +1,8 @@
-# from config import config_dict
+from __future__ import absolute_import, division, print_function, unicode_literals
+
 import unittest
 
-import keras.backend as K
+import keras.backend as k
 import tensorflow as tf
 import numpy as np
 
@@ -36,7 +37,7 @@ class TestMinimalPerturbation(unittest.TestCase):
 
     def test_emp_robustness_mnist(self):
         session = tf.Session()
-        K.set_session(session)
+        k.set_session(session)
 
         comp_params = {"loss": 'categorical_crossentropy',
                        "optimizer": 'adam',
@@ -108,7 +109,7 @@ def test_clever_t_unit(self):
         print("Unit test for the targeted version with simplified data.")
         # Define session & params
         session = tf.Session()
-        K.set_session(session)
+        k.set_session(session)
 
         # Get classifier
         classifier = TestClassifier()
@@ -129,7 +130,7 @@ def test_clever_u_unit(self):
         print("Unit test for the untargeted version with simplified data.")
         # Define session & params
         session = tf.Session()
-        K.set_session(session)
+        k.set_session(session)
 
         # Get classifier
         classifier = TestClassifier()
@@ -150,7 +151,7 @@ def test_clever_t(self):
         print("Test if the targeted version works on a true classifier/data")
         # Define session & params
         session = tf.Session()
-        K.set_session(session)
+        k.set_session(session)
 
         comp_params = {"loss": 'categorical_crossentropy', "optimizer": 'adam',
                        "metrics": ['accuracy']}
@@ -178,7 +179,7 @@ def test_clever_u(self):
         print("Test if the untargeted version works on a true classifier/data")
         # Define session & params
         session = tf.Session()
-        K.set_session(session)
+        k.set_session(session)
 
         comp_params = {"loss": 'categorical_crossentropy', "optimizer": 'adam',
                        "metrics": ['accuracy']}

src/utils.py

@@ -1,7 +1,7 @@
 """
 Module providing convenience functions.
 """
-from __future__ import absolute_import, division, print_function
+from __future__ import absolute_import, division, print_function, unicode_literals
 
 import argparse
 import json