Ensure no execution of slow notebooks.

Fixed dpsgd notebook.

PiperOrigin-RevId: 605543405

docs/conf.py

@@ -222,8 +222,10 @@ def new_process_docstring(app, what, name, obj, options, lines):
 nb_execution_allow_errors = False
 nb_execution_excludepatterns = [
     # slow examples
-    '_collections/examples/cifar10_resnet.ipynb'
-    '_collections/examples/adversarial_training.ipynb'
+    'cifar10_resnet.ipynb',
+    'adversarial_training.ipynb',
+    'reduce_on_plateau.ipynb',
+    'differentially_private_sgd.ipynb'
 ]
 
 # -- Options for katex ------------------------------------------------------

examples/contrib/differentially_private_sgd.ipynb

@@ -99,17 +99,17 @@
       "outputs": [],
       "source": [
         "# @markdown Whether to use DP-SGD or vanilla SGD:\n",
-        "DPSGD = False  # @param{type:\"boolean\"}\n",
+        "DPSGD = True  # @param{type:\"boolean\"}\n",
         "# @markdown Learning rate for the optimizer:\n",
-        "LEARNING_RATE = 0.1  # @param{type:\"number\"}\n",
+        "LEARNING_RATE = 0.25  # @param{type:\"number\"}\n",
         "# @markdown Noise multiplier for DP-SGD optimizer:\n",
-        "NOISE_MULTIPLIER = 1.1  # @param{type:\"number\"}\n",
+        "NOISE_MULTIPLIER = 1.3  # @param{type:\"number\"}\n",
         "# @markdown L2 norm clip:\n",
-        "L2_NORM_CLIP = 1.0  # @param{type:\"number\"}\n",
+        "L2_NORM_CLIP = 1.5  # @param{type:\"number\"}\n",
         "# @markdown Number of samples in each batch:\n",
         "BATCH_SIZE = 256  # @param{type:\"integer\"}\n",
         "# @markdown Number of epochs:\n",
-        "NUM_EPOCHS = 20  # @param{type:\"integer\"}\n",
+        "NUM_EPOCHS = 15  # @param{type:\"integer\"}\n",
         "# @markdown Probability of information leakage:\n",
         "DELTA = 1e-5  # @param{type:\"number\"}"
       ]
@@ -351,6 +351,7 @@
         "  test_loss, test_acc = test_step(params, test_batch)\n",
         "  accuracy.append(test_acc)\n",
         "  loss.append(test_loss)\n",
+        "  print(f\"Epoch {epoch + 1}/{NUM_EPOCHS}, test accuracy: {test_acc}\")\n",
         "\n",
         "  #\n",
         "  if DPSGD:\n",
@@ -397,7 +398,10 @@
         }
       ],
       "source": [
-        "_, axs = plt.subplots(ncols=3 if DPSGD else 2)\n",
+        "if DPSGD:\n",
+        "  _, axs = plt.subplots(ncols=3, figsize=(9, 3))\n",
+        "else:\n",
+        "  _, axs = plt.subplots(ncols=2, figsize=(6, 3))\n",
         "\n",
         "axs[0].plot(accuracy)\n",
         "axs[0].set_title(\"Test accuracy\")\n",
@@ -406,7 +410,9 @@
         "\n",
         "if DPSGD:\n",
         "  axs[2].plot(epsilon)\n",
-        "  axs[2].set_title(\"Epsilon\")"
+        "  axs[2].set_title(\"Epsilon\")\n",
+        "\n",
+        "plt.tight_layout()"
       ]
     },
     {
@@ -439,7 +445,7 @@
         }
       ],
       "source": [
-        "accuracy[-1]"
+        "print(f'Final accuracy: {accuracy[-1]}')"
       ]
     }
   ],

pyproject.toml

@@ -50,7 +50,8 @@ test = [
 
 examples = [
   "tensorflow-datasets>=4.2.0",
-  "tensorflow>=2.4.0"
+  "tensorflow>=2.4.0",
+  "dp_accounting>=0.4"
 ]
 
 docs = [