Add note to tutorial

Author: pchlenski

tutorial.ipynb

@@ -1388,7 +1388,12 @@
    "id": "e06bba03",
    "metadata": {},
    "source": [
-    "In fact, since we've written our classes to be `scikit-learn` compatible where possible, we can even use the `Pipeline` class to chain together embedders and predictors, just as we would in a regular Euclidean scikit-learn workflow. "
+    "In fact, since we've written our classes to be `scikit-learn` compatible where possible, we can even use the `Pipeline` class to chain together embedders and predictors, just as we would in a regular Euclidean scikit-learn workflow. \n",
+    "\n",
+    "(**NOTE:** This code will occasionally throw some warnings, particularly \"NaN in parallel transport,\" which causes\n",
+    "the product space VAE to skip the gradient update for that batch and keep going. It should train successfully, but if\n",
+    "it doesn't, you can try changing the signature or the learning rate. The warnings are not indicative of a deeper bug,\n",
+    "but rather the inherent numerical instability of working with Gaussian distributions on product manifolds.)"
    ]
   },
   {