Fix bug in v1 estimators that was preventing use of microbatches.

PiperOrigin-RevId: 560765153

tensorflow_privacy/privacy/estimators/v1/dnn_test.py

@@ -20,16 +20,20 @@ from tensorflow_privacy.privacy.estimators import test_utils
 from tensorflow_privacy.privacy.estimators.v1 import dnn
 from tensorflow_privacy.privacy.optimizers.dp_optimizer import DPGradientDescentGaussianOptimizer
 
+# pylint: disable=g-deprecated-tf-checker
+
 
 class DPDNNClassifierTest(tf.test.TestCase, parameterized.TestCase):
   """Tests for DP-enabled DNNClassifier."""
 
   @parameterized.named_parameters(
-      ('BinaryClassDNN', 2),
+      ('BinaryClassDNN', 2, 1),
-      ('MultiClassDNN 3', 3),
+      ('BinaryClassDNN 4', 2, 4),
-      ('MultiClassDNN 4', 4),
+      ('MultiClassDNN 3', 3, 1),
+      ('MultiClassDNN 4', 4, 1),
+      ('MultiClassDNN 4 4', 4, 4),
   )
-  def testDNN(self, n_classes):
+  def testDNN(self, n_classes, num_microbatches):
     train_features, train_labels = test_utils.make_input_data(256, n_classes)
     feature_columns = []
     for key in train_features:
@@ -40,7 +44,8 @@ class DPDNNClassifierTest(tf.test.TestCase, parameterized.TestCase):
         learning_rate=0.5,
         l2_norm_clip=1.0,
         noise_multiplier=0.0,
-        num_microbatches=1)
+        num_microbatches=num_microbatches,
+    )
 
     classifier = dnn.DNNClassifier(
         hidden_units=[10],

tensorflow_privacy/privacy/estimators/v1/head.py

@@ -16,6 +16,7 @@
 import tensorflow as tf
 
 from tensorflow.python.ops import lookup_ops  # pylint: disable=g-direct-tensorflow-import
+# pylint: disable=g-deprecated-tf-checker
 from tensorflow_estimator.python.estimator import model_fn
 from tensorflow_estimator.python.estimator.canned import head as head_lib
 from tensorflow_estimator.python.estimator.canned import metric_keys
@@ -23,6 +24,7 @@ from tensorflow_estimator.python.estimator.canned import prediction_keys
 from tensorflow_estimator.python.estimator.export import export_output
 from tensorflow_estimator.python.estimator.mode_keys import ModeKeys
 
+
 # Collect together all protected access items needed from base head.
 # pylint: disable=protected-access
 _DEFAULT_SERVING_KEY = head_lib._DEFAULT_SERVING_KEY
@@ -39,8 +41,12 @@ _create_eval_metrics_tuple = head_lib._create_eval_metrics_tuple
 _summary_key = head_lib._summary_key
 _validate_loss_fn_args = head_lib._validate_loss_fn_args
 
-_BaseBinaryLogisticHeadWithSigmoidCrossEntropyLoss = head_lib._BinaryLogisticHeadWithSigmoidCrossEntropyLoss
+_BaseBinaryLogisticHeadWithSigmoidCrossEntropyLoss = (
-_BaseMultiClassHeadWithSoftmaxCrossEntropyLoss = head_lib._MultiClassHeadWithSoftmaxCrossEntropyLoss
+    head_lib._BinaryLogisticHeadWithSigmoidCrossEntropyLoss
+)
+_BaseMultiClassHeadWithSoftmaxCrossEntropyLoss = (
+    head_lib._MultiClassHeadWithSoftmaxCrossEntropyLoss
+)
 # pylint: enable=protected-access
 
 
@@ -146,25 +152,33 @@ class _MultiClassHeadWithSoftmaxCrossEntropyLoss(
         classifier_output = _classification_output(
             scores=probabilities,
             n_classes=self._n_classes,
-            label_vocabulary=self._label_vocabulary)
+            label_vocabulary=self._label_vocabulary,
+        )
         return model_fn._TPUEstimatorSpec(  # pylint: disable=protected-access
             mode=ModeKeys.PREDICT,
             predictions=predictions,
             export_outputs={
                 _DEFAULT_SERVING_KEY: classifier_output,
                 _CLASSIFY_SERVING_KEY: classifier_output,
-                _PREDICT_SERVING_KEY: export_output.PredictOutput(predictions)
+                _PREDICT_SERVING_KEY: export_output.PredictOutput(predictions),
-            })
+            },
+        )
 
       training_loss, unreduced_loss, weights, label_ids = self.create_loss(
-          features=features, mode=mode, logits=logits, labels=labels)
+          features=features, mode=mode, logits=logits, labels=labels
+      )
       if regularization_losses:
         regularization_loss = tf.math.add_n(regularization_losses)
         regularized_training_loss = tf.math.add_n(
-            [training_loss, regularization_loss])
+            [training_loss, regularization_loss]
+        )
+        unreduced_regularized_training_loss = tf.math.add(
+            unreduced_loss, regularization_loss
+        )
       else:
         regularization_loss = None
         regularized_training_loss = training_loss
+        unreduced_regularized_training_loss = unreduced_loss
 
       if self._loss_reduction == tf.compat.v1.losses.Reduction.NONE:
         scalar_loss = tf.reduce_mean(regularized_training_loss)
@@ -191,8 +205,10 @@ class _MultiClassHeadWithSoftmaxCrossEntropyLoss(
         if train_op_fn is not None:
           raise ValueError('train_op_fn and optimizer cannot both be set.')
         train_op = optimizer.minimize(
-            regularized_training_loss,
+            # regularized_training_loss,
-            global_step=tf.compat.v1.train.get_global_step())
+            unreduced_regularized_training_loss,
+            global_step=tf.compat.v1.train.get_global_step(),
+        )
       elif train_op_fn is not None:
         train_op = train_op_fn(regularized_training_loss)
       else:
@@ -352,9 +368,13 @@ class _BinaryLogisticHeadWithSigmoidCrossEntropyLoss(
         regularization_loss = tf.math.add_n(regularization_losses)
         regularized_training_loss = tf.math.add_n(
             [training_loss, regularization_loss])
+        unreduced_regularized_training_loss = tf.math.add(
+            unreduced_loss, regularization_loss
+        )
       else:
         regularization_loss = None
         regularized_training_loss = training_loss
+        unreduced_regularized_training_loss = unreduced_loss
 
       if self._loss_reduction == tf.compat.v1.losses.Reduction.NONE:
         scalar_loss = tf.reduce_mean(regularized_training_loss)
@@ -382,8 +402,9 @@ class _BinaryLogisticHeadWithSigmoidCrossEntropyLoss(
         if train_op_fn is not None:
           raise ValueError('train_op_fn and optimizer cannot both be set.')
         train_op = optimizer.minimize(
-            regularized_training_loss,
+            unreduced_regularized_training_loss,
-            global_step=tf.compat.v1.train.get_global_step())
+            global_step=tf.compat.v1.train.get_global_step(),
+        )
       elif train_op_fn is not None:
         train_op = train_op_fn(regularized_training_loss)
       else:

tensorflow_privacy/privacy/estimators/v1/linear_test.py

@@ -21,6 +21,8 @@ from tensorflow_privacy.privacy.estimators import test_utils
 from tensorflow_privacy.privacy.estimators.v1 import linear
 from tensorflow_privacy.privacy.optimizers.dp_optimizer import DPGradientDescentGaussianOptimizer
 
+# pylint: disable=g-deprecated-tf-checker
+
 
 class DPLinearClassifierClassifierTest(
     tf.test.TestCase, parameterized.TestCase
@@ -28,11 +30,13 @@ class DPLinearClassifierClassifierTest(
   """Tests for DP-enabled LinearClassifier."""
 
   @parameterized.named_parameters(
-      ('BinaryClassLinear', 2),
+      ('BinaryClassLinear 1', 2, 1),
-      ('MultiClassLinear 3', 3),
+      ('BinaryClassLinear 4', 2, 4),
-      ('MultiClassLinear 4', 4),
+      ('MultiClassLinear 3', 3, 1),
+      ('MultiClassLinear 4', 4, 1),
+      ('MultiClassLinear 4 1', 4, 2),
   )
-  def testLinearClassifier(self, n_classes):
+  def testRunsWithoutErrors(self, n_classes, num_microbatches):
     train_features, train_labels = test_utils.make_input_data(256, n_classes)
     feature_columns = []
     for key in train_features:
@@ -43,7 +47,7 @@ class DPLinearClassifierClassifierTest(
         learning_rate=0.5,
         l2_norm_clip=1.0,
         noise_multiplier=0.0,
-        num_microbatches=1,
+        num_microbatches=num_microbatches,
     )
 
     classifier = linear.LinearClassifier(