Closes cleverhans-lab#744 (PGD wrapper issue)

tests_tf/test_projected_gradient_descent.py

@@ -0,0 +1,28 @@
+"""Tests for the ProjectGradientDescent attack
+"""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from nose.tools import assert_raises
+import tensorflow as tf
+
+from cleverhans.attacks import ProjectedGradientDescent
+
+def test_callable_no_softmax():
+  batch_size = 2
+  nb_classes = 3
+  def model(x):
+    return tf.ones((batch_size, nb_classes)) / nb_classes
+  sess = tf.Session()
+  attack = ProjectedGradientDescent(model, sess=sess)
+  x = tf.ones((batch_size, 3))
+  # Currently ProjectedGradientDescent treats the output of a callable
+  # as probs rather than logits.
+  # Since our callable does not use a softmax, it is impossible to get
+  # the logits back. The test confirms that this causes an error.
+  assert_raises(TypeError, attack.generate, x)
+
+if __name__ == "__main__":
+  test_callable()