more verbose description for iris replacements

Author: ferrine

docs/source/notebooks/variational_api_quickstart.ipynb

@@ -1240,7 +1240,9 @@
   {
    "cell_type": "code",
    "execution_count": 52,
-   "metadata": {},
+   "metadata": {
+    "collapsed": true
+   },
    "outputs": [],
    "source": [
     "# First of all\n",
@@ -1269,7 +1271,9 @@
   {
    "cell_type": "code",
    "execution_count": 53,
-   "metadata": {},
+   "metadata": {
+    "collapsed": true
+   },
    "outputs": [],
    "source": [
     "Xt = theano.shared(X_train)\n",
@@ -1285,7 +1289,15 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "To get accuracy under approximate posterior we need to use `apply_replacements` or `sample_node`. It's better to get the whole distribution at each step to draw neat plot so `sample_node` is our choice."
+    "### Applying replacements in practice\n",
+    "Models defined with pymc3 have symbolic inputs for latent variables. To evaluate an espression that requires knoledge about latents one needs to pass values for latent variables. In VI setup we have an approximation for these variables ant it happens to be very usefull in practice. Having a simple functions that removes all symbolic dependencies is valuable. These functions are `sample_node` and `apply_replacements` described above. Before we did not use the full power of replacements. There is a usefull shortcut for applying even more replacements at once. \n",
+    "\n",
+    "To get accuracy under approximate posterior we need to use `apply_replacements` or `sample_node`. It's better to get the whole distribution at each step to draw neat plot so `sample_node` is our choice. As mentioned above we can apply more replacements in single function call. It can be done with readable kwarg `more_replacements` in both replacement functions.\n",
+    "\n",
+    "**HINT:** You can use `more_replacements` argument when calling `fit` too\n",
+    "\n",
+    "   * `pm.fit(more_replacements={full_data: minibatch_data})`\n",
+    "   * `inference.fit(more_replacements={full_data: minibatch_data})`"
    ]
   },
   {
@@ -1319,7 +1331,9 @@
   {
    "cell_type": "code",
    "execution_count": 55,
-   "metadata": {},
+   "metadata": {
+    "collapsed": true
+   },
    "outputs": [],
    "source": [
     "test_ok = tt.eq(test_probs.argmax(-1), y_test)\n",
@@ -1349,7 +1363,9 @@
   {
    "cell_type": "code",
    "execution_count": 57,
-   "metadata": {},
+   "metadata": {
+    "collapsed": true
+   },
    "outputs": [],
    "source": [
     "eval_tracker = pm.callbacks.Tracker(\n",