Update SGPR notebook (#2303)

Author: gpleiss

examples/02_Scalable_Exact_GPs/SGPR_Regression_CUDA.ipynb

@@ -20,6 +20,7 @@
     "import math\n",
     "import torch\n",
     "import gpytorch\n",
+    "import tqdm.notebook as tqdm\n",
     "from matplotlib import pyplot as plt\n",
     "\n",
     "# Make plots inline\n",
@@ -121,7 +122,7 @@
     "        super(GPRegressionModel, self).__init__(train_x, train_y, likelihood)\n",
     "        self.mean_module = ConstantMean()\n",
     "        self.base_covar_module = ScaleKernel(RBFKernel())\n",
-    "        self.covar_module = InducingPointKernel(self.base_covar_module, inducing_points=train_x[:500, :], likelihood=likelihood)\n",
+    "        self.covar_module = InducingPointKernel(self.base_covar_module, inducing_points=train_x[:500, :].clone(), likelihood=likelihood)\n",
     "\n",
     "    def forward(self, x):\n",
     "        mean_x = self.mean_module(x)\n",
@@ -154,70 +155,34 @@
    "cell_type": "code",
    "execution_count": 6,
    "metadata": {
-    "scrolled": false
+    "scrolled": true
    },
    "outputs": [
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "90181179167b4cf3949a43c21dd8f10c",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "Train:   0%|          | 0/100 [00:00<?, ?it/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
     {
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "Iter 1/50 - Loss: 0.794\n",
-      "Iter 2/50 - Loss: 0.782\n",
-      "Iter 3/50 - Loss: 0.770\n",
-      "Iter 4/50 - Loss: 0.758\n",
-      "Iter 5/50 - Loss: 0.746\n",
-      "Iter 6/50 - Loss: 0.734\n",
-      "Iter 7/50 - Loss: 0.721\n",
-      "Iter 8/50 - Loss: 0.708\n",
-      "Iter 9/50 - Loss: 0.695\n",
-      "Iter 10/50 - Loss: 0.681\n",
-      "Iter 11/50 - Loss: 0.667\n",
-      "Iter 12/50 - Loss: 0.654\n",
-      "Iter 13/50 - Loss: 0.641\n",
-      "Iter 14/50 - Loss: 0.626\n",
-      "Iter 15/50 - Loss: 0.613\n",
-      "Iter 16/50 - Loss: 0.598\n",
-      "Iter 17/50 - Loss: 0.584\n",
-      "Iter 18/50 - Loss: 0.571\n",
-      "Iter 19/50 - Loss: 0.555\n",
-      "Iter 20/50 - Loss: 0.541\n",
-      "Iter 21/50 - Loss: 0.526\n",
-      "Iter 22/50 - Loss: 0.510\n",
-      "Iter 23/50 - Loss: 0.495\n",
-      "Iter 24/50 - Loss: 0.481\n",
-      "Iter 25/50 - Loss: 0.465\n",
-      "Iter 26/50 - Loss: 0.449\n",
-      "Iter 27/50 - Loss: 0.435\n",
-      "Iter 28/50 - Loss: 0.417\n",
-      "Iter 29/50 - Loss: 0.401\n",
-      "Iter 30/50 - Loss: 0.384\n",
-      "Iter 31/50 - Loss: 0.369\n",
-      "Iter 32/50 - Loss: 0.351\n",
-      "Iter 33/50 - Loss: 0.336\n",
-      "Iter 34/50 - Loss: 0.319\n",
-      "Iter 35/50 - Loss: 0.303\n",
-      "Iter 36/50 - Loss: 0.286\n",
-      "Iter 37/50 - Loss: 0.269\n",
-      "Iter 38/50 - Loss: 0.253\n",
-      "Iter 39/50 - Loss: 0.236\n",
-      "Iter 40/50 - Loss: 0.217\n",
-      "Iter 41/50 - Loss: 0.200\n",
-      "Iter 42/50 - Loss: 0.181\n",
-      "Iter 43/50 - Loss: 0.167\n",
-      "Iter 44/50 - Loss: 0.149\n",
-      "Iter 45/50 - Loss: 0.132\n",
-      "Iter 46/50 - Loss: 0.112\n",
-      "Iter 47/50 - Loss: 0.096\n",
-      "Iter 48/50 - Loss: 0.078\n",
-      "Iter 49/50 - Loss: 0.061\n",
-      "Iter 50/50 - Loss: 0.044\n",
-      "CPU times: user 2min 47s, sys: 7.87 s, total: 2min 55s\n",
-      "Wall time: 34.6 s\n"
+      "CPU times: user 2.7 s, sys: 852 ms, total: 3.55 s\n",
+      "Wall time: 3.58 s\n"
      ]
     }
    ],
    "source": [
-    "training_iterations = 2 if smoke_test else 50\n",
+    "training_iterations = 2 if smoke_test else 100\n",
     "\n",
     "# Find optimal model hyperparameters\n",
     "model.train()\n",
@@ -230,58 +195,51 @@
     "mll = gpytorch.mlls.ExactMarginalLogLikelihood(likelihood, model)\n",
     "\n",
     "def train():\n",
-    "    for i in range(training_iterations):\n",
+    "    iterator = tqdm.tqdm(range(training_iterations), desc=\"Train\")\n",
+    "\n",
+    "    for i in iterator:\n",
     "        # Zero backprop gradients\n",
     "        optimizer.zero_grad()\n",
     "        # Get output from model\n",
     "        output = model(train_x)\n",
     "        # Calc loss and backprop derivatives\n",
     "        loss = -mll(output, train_y)\n",
     "        loss.backward()\n",
-    "        print('Iter %d/%d - Loss: %.3f' % (i + 1, training_iterations, loss.item()))\n",
+    "        iterator.set_postfix(loss=loss.item())\n",
     "        optimizer.step()\n",
     "        torch.cuda.empty_cache()\n",
     "        \n",
-    "# See dkl_mnist.ipynb for explanation of this flag\n",
     "%time train()"
    ]
   },
   {
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "### Making Predictions\n",
-    "\n",
-    "The next cell makes predictions with SKIP. We use the same max_root_decomposition size, and we also demonstrate increasing the max preconditioner size. Increasing the preconditioner size on this dataset is **not** necessary, but can make a big difference in final test performance, and is often preferable to increasing the number of CG iterations if you can afford the space."
+    "### Making Predictions"
    ]
   },
   {
    "cell_type": "code",
    "execution_count": 7,
    "metadata": {},
-   "outputs": [],
-   "source": [
-    "model.eval()\n",
-    "likelihood.eval()\n",
-    "with gpytorch.settings.max_preconditioner_size(10), torch.no_grad():\n",
-    "    preds = model(test_x)"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 8,
-   "metadata": {},
    "outputs": [
     {
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "Test MAE: 0.07271435856819153\n"
+      "Test MAE: 0.07258129864931107\n",
+      "Test NLL: 0.3463870584964752\n"
      ]
     }
    ],
    "source": [
-    "print('Test MAE: {}'.format(torch.mean(torch.abs(preds.mean - test_y))))"
+    "model.eval()\n",
+    "likelihood.eval()\n",
+    "with torch.no_grad():\n",
+    "    preds = model.likelihood(model(test_x))\n",
+    "    print('Test MAE: {}'.format(torch.mean(torch.abs(preds.mean - test_y))))\n",
+    "    print('Test NLL: {}'.format(-preds.to_data_independent_dist().log_prob(test_y).mean().item()))"
    ]
   },
   {
@@ -295,7 +253,7 @@
  "metadata": {
   "anaconda-cloud": {},
   "kernelspec": {
-   "display_name": "Python 3",
+   "display_name": "Python 3 (ipykernel)",
    "language": "python",
    "name": "python3"
   },
@@ -309,7 +267,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.7.3"
+   "version": "3.8.0"
   }
  },
  "nbformat": 4,