Merge pull request cleverhans-lab#356 from goodfeli/dtype

support float64

cleverhans/attacks.py

@@ -529,7 +529,7 @@ def generate(self, x, **kwargs):
 
             # Normalize current gradient and add it to the accumulated gradient
             red_ind = list(xrange(1, len(grad.get_shape())))
-            avoid_zero_div = 1e-12
+            avoid_zero_div = tf.cast(1e-12, grad.dtype)
             grad = grad / tf.maximum(avoid_zero_div,
                                      tf.reduce_mean(tf.abs(grad),
                                                     red_ind,

tests_tf/test_attacks.py

@@ -29,9 +29,10 @@ class SimpleModel(Model):
     """
 
     def get_logits(self, x):
-        W1 = tf.constant([[1.5, .3], [-2, 0.3]], dtype=tf.float32)
+        W1 = tf.constant([[1.5, .3], [-2, 0.3]], dtype=x.dtype)
         h1 = tf.nn.sigmoid(tf.matmul(x, W1))
-        W2 = tf.constant([[-2.4, 1.2], [0.5, -2.3]], dtype=tf.float32)
+        W2 = tf.constant([[-2.4, 1.2], [0.5, -2.3]], dtype=x.dtype)
+
         res = tf.matmul(h1, W2)
         return res
 
@@ -174,6 +175,11 @@ def test_generate_np_gives_adversarial_example_l1(self):
     def test_generate_np_gives_adversarial_example_l2(self):
         self.help_generate_np_gives_adversarial_example(2)
 
+    def test_generate_respects_dtype(self):
+        x = tf.placeholder(dtype=tf.float64, shape=(100, 2))
+        x_adv = self.attack.generate(x)
+        self.assertEqual(x_adv.dtype, tf.float64)
+
     def test_targeted_generate_np_gives_adversarial_example(self):
         x_val = np.random.rand(100, 2)
         x_val = np.array(x_val, dtype=np.float32)