Clarify expected warnings in adversarial tutorial.

PiperOrigin-RevId: 290347223

Author: csferng

g3doc/tutorials/adversarial_keras_cnn_mnist.ipynb

@@ -442,7 +442,7 @@
       "source": [
         "results = base_model.evaluate(test_dataset)\n",
         "named_results = dict(zip(base_model.metrics_names, results))\n",
-        "print('accuracy:', named_results['accuracy'])"
+        "print('\\naccuracy:', named_results['accuracy'])"
       ]
     },
     {
@@ -548,7 +548,11 @@
         "id": "aKTQWzfj7JvL"
       },
       "source": [
-        "Next we compile, train, and evaluate the adversarial-regularized model."
+        "Next we compile, train, and evaluate the\n",
+        "adversarial-regularized model. There might be warnings like\n",
+        "\"Output missing from loss dictionary,\" which is fine because\n",
+        "the `adv_model` doesn't rely on the base implementation to\n",
+        "calculate the total loss."
       ]
     },
     {
@@ -578,7 +582,7 @@
       "source": [
         "results = adv_model.evaluate(test_set_for_adv_model)\n",
         "named_results = dict(zip(adv_model.metrics_names, results))\n",
-        "print('accuracy:', named_results['sparse_categorical_accuracy'])"
+        "print('\\naccuracy:', named_results['sparse_categorical_accuracy'])"
       ]
     },
     {