better nuts

kormilitzin · Apr 5, 2018 · 98cfc6f · 98cfc6f
1 parent 704cc47
commit 98cfc6f
Show file tree

Hide file tree

Showing 2 changed files with 19 additions and 35 deletions.
diff --git a/solutions.ipynb b/solutions.ipynb
diff --git a/worksheet.ipynb b/worksheet.ipynb
@@ -406,7 +406,7 @@
     "        log_prob,\n",
     "        [m_tensor, b_tensor, log_s_tensor],\n",
     "        3500,          # The number of MCMC steps\n",
-    "        0.1,           # The integration step size\n",
+    "        1.0,           # The integration step size\n",
     "        metric=metric  # The scaling metric computed above\n",
     "    )\n",
     "nuts_time = time.time() - nuts_time"
@@ -457,7 +457,8 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "In this case, the NUTS sampler is actually less efficient (i.e. it requires more computation time per independent sample) than emcee (and we're not even considering the cost of tuning here), but this result will quickly be reversed for models with more parameters.\n",
+    "If everything went as planned, you should see that the NUTS sampler is substantially more efficient (i.e. it requires less computation time per independent sample) than emcee (but we're not considering the cost of tuning here).\n",
+    "This disparity will tend to increase for \n",
     "In general, sampling a density with tens of parameters using emcee will require substantially more patience and computation power than sampling the same problem with NUTS. "
    ]
   },