diff --git a/research/GDP_2019/adult_tutorial.py b/research/GDP_2019/adult_tutorial.py
index 437770b..36741a0 100644
--- a/research/GDP_2019/adult_tutorial.py
+++ b/research/GDP_2019/adult_tutorial.py
@@ -26,6 +26,8 @@ import pandas as pd
 from sklearn.model_selection import KFold
 import tensorflow as tf
 
+from tensorflow import estimator as tf_estimator
+from tensorflow.compat.v1 import estimator as tf_compat_v1_estimator
 from tensorflow_privacy.privacy.analysis.gdp_accountant import compute_eps_poisson
 from tensorflow_privacy.privacy.analysis.gdp_accountant import compute_mu_poisson
 from tensorflow_privacy.privacy.optimizers import dp_optimizer
@@ -61,7 +63,7 @@ def nn_model_fn(features, labels, mode):
   scalar_loss = tf.reduce_mean(vector_loss)
 
   # Configure the training op (for TRAIN mode).
-  if mode == tf.estimator.ModeKeys.TRAIN:
+  if mode == tf_estimator.ModeKeys.TRAIN:
     if FLAGS.dpsgd:
       # Use DP version of GradientDescentOptimizer. Other optimizers are
       # available in dp_optimizer. Most optimizers inheriting from
@@ -83,17 +85,17 @@ def nn_model_fn(features, labels, mode):
     # the vector_loss because tf.estimator requires a scalar loss. This is only
     # used for evaluation and debugging by tf.estimator. The actual loss being
     # minimized is opt_loss defined above and passed to optimizer.minimize().
-    return tf.estimator.EstimatorSpec(
+    return tf_estimator.EstimatorSpec(
         mode=mode, loss=scalar_loss, train_op=train_op)
 
   # Add evaluation metrics (for EVAL mode).
-  if mode == tf.estimator.ModeKeys.EVAL:
+  if mode == tf_estimator.ModeKeys.EVAL:
     eval_metric_ops = {
         'accuracy':
             tf.compat.v1.metrics.accuracy(
                 labels=labels, predictions=tf.argmax(input=logits, axis=1))
     }
-    return tf.estimator.EstimatorSpec(
+    return tf_estimator.EstimatorSpec(
         mode=mode, loss=scalar_loss, eval_metric_ops=eval_metric_ops)
 
   return None
@@ -123,11 +125,11 @@ def main(unused_argv):
   train_data, train_labels, test_data, test_labels = load_adult()
 
   # Instantiate the tf.Estimator.
-  adult_classifier = tf.compat.v1.estimator.Estimator(
+  adult_classifier = tf_compat_v1_estimator.Estimator(
       model_fn=nn_model_fn, model_dir=FLAGS.model_dir)
 
   # Create tf.Estimator input functions for the training and test data.
-  eval_input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(
+  eval_input_fn = tf_compat_v1_estimator.inputs.numpy_input_fn(
       x={'x': test_data}, y=test_labels, num_epochs=1, shuffle=False)
 
   # Training loop.
@@ -141,7 +143,7 @@ def main(unused_argv):
       global microbatches
       microbatches = len(subsampling)
 
-      train_input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(
+      train_input_fn = tf_compat_v1_estimator.inputs.numpy_input_fn(
           x={'x': train_data[subsampling]},
           y=train_labels[subsampling],
           batch_size=len(subsampling),
diff --git a/research/GDP_2019/imdb_tutorial.py b/research/GDP_2019/imdb_tutorial.py
index 4281439..348321a 100644
--- a/research/GDP_2019/imdb_tutorial.py
+++ b/research/GDP_2019/imdb_tutorial.py
@@ -25,6 +25,8 @@ from keras.preprocessing import sequence
 import numpy as np
 import tensorflow as tf
 
+from tensorflow import estimator as tf_estimator
+from tensorflow.compat.v1 import estimator as tf_compat_v1_estimator
 from tensorflow_privacy.privacy.analysis.gdp_accountant import compute_eps_poisson
 from tensorflow_privacy.privacy.analysis.gdp_accountant import compute_mu_poisson
 from tensorflow_privacy.privacy.optimizers import dp_optimizer
@@ -65,7 +67,7 @@ def nn_model_fn(features, labels, mode):
   scalar_loss = tf.reduce_mean(vector_loss)
 
   # Configure the training op (for TRAIN mode).
-  if mode == tf.estimator.ModeKeys.TRAIN:
+  if mode == tf_estimator.ModeKeys.TRAIN:
     if FLAGS.dpsgd:
       # Use DP version of GradientDescentOptimizer. Other optimizers are
       # available in dp_optimizer. Most optimizers inheriting from
@@ -88,17 +90,17 @@ def nn_model_fn(features, labels, mode):
     # the vector_loss because tf.estimator requires a scalar loss. This is only
     # used for evaluation and debugging by tf.estimator. The actual loss being
     # minimized is opt_loss defined above and passed to optimizer.minimize().
-    return tf.estimator.EstimatorSpec(
+    return tf_estimator.EstimatorSpec(
         mode=mode, loss=scalar_loss, train_op=train_op)
 
   # Add evaluation metrics (for EVAL mode).
-  if mode == tf.estimator.ModeKeys.EVAL:
+  if mode == tf_estimator.ModeKeys.EVAL:
     eval_metric_ops = {
         'accuracy':
             tf.compat.v1.metrics.accuracy(
                 labels=labels, predictions=tf.argmax(input=logits, axis=1))
     }
-    return tf.estimator.EstimatorSpec(
+    return tf_estimator.EstimatorSpec(
         mode=mode, loss=scalar_loss, eval_metric_ops=eval_metric_ops)
   return None
 
@@ -122,11 +124,11 @@ def main(unused_argv):
   train_data, train_labels, test_data, test_labels = load_imdb()
 
   # Instantiate the tf.Estimator.
-  imdb_classifier = tf.estimator.Estimator(
+  imdb_classifier = tf_estimator.Estimator(
       model_fn=nn_model_fn, model_dir=FLAGS.model_dir)
 
   # Create tf.Estimator input functions for the training and test data.
-  eval_input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(
+  eval_input_fn = tf_compat_v1_estimator.inputs.numpy_input_fn(
       x={'x': test_data}, y=test_labels, num_epochs=1, shuffle=False)
 
   # Training loop.
@@ -141,7 +143,7 @@ def main(unused_argv):
       global microbatches
       microbatches = len(subsampling)
 
-      train_input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(
+      train_input_fn = tf_compat_v1_estimator.inputs.numpy_input_fn(
           x={'x': train_data[subsampling]},
           y=train_labels[subsampling],
           batch_size=len(subsampling),
diff --git a/tensorflow_privacy/privacy/estimators/v1/BUILD b/tensorflow_privacy/privacy/estimators/v1/BUILD
index 0522c75..3195cd8 100644
--- a/tensorflow_privacy/privacy/estimators/v1/BUILD
+++ b/tensorflow_privacy/privacy/estimators/v1/BUILD
@@ -23,7 +23,10 @@ py_library(
         "dnn.py",
     ],
     srcs_version = "PY3",
-    deps = [":head"],
+    deps = [
+        ":head",
+        "//third_party/py/tensorflow:tensorflow_estimator",
+    ],
 )
 
 py_test(
@@ -36,6 +39,7 @@ py_test(
         ":head",
         "//tensorflow_privacy/privacy/estimators:test_utils",
         "//tensorflow_privacy/privacy/optimizers:dp_optimizer",
+        "//third_party/py/tensorflow:tensorflow_estimator",
     ],
 )
 
diff --git a/tensorflow_privacy/privacy/estimators/v1/dnn.py b/tensorflow_privacy/privacy/estimators/v1/dnn.py
index 6cc09f6..143d793 100644
--- a/tensorflow_privacy/privacy/estimators/v1/dnn.py
+++ b/tensorflow_privacy/privacy/estimators/v1/dnn.py
@@ -16,12 +16,13 @@
 
 import tensorflow as tf
 
+from tensorflow import estimator as tf_estimator
 from tensorflow_privacy.privacy.estimators.v1 import head as head_lib
 from tensorflow_estimator.python.estimator import estimator
 from tensorflow_estimator.python.estimator.canned import dnn
 
 
-class DNNClassifier(tf.estimator.Estimator):
+class DNNClassifier(tf_estimator.Estimator):
   """DP version of `tf.compat.v1.estimator.DNNClassifier`."""
 
   def __init__(
diff --git a/tensorflow_privacy/privacy/estimators/v1/head_test.py b/tensorflow_privacy/privacy/estimators/v1/head_test.py
index 5238912..3536e09 100644
--- a/tensorflow_privacy/privacy/estimators/v1/head_test.py
+++ b/tensorflow_privacy/privacy/estimators/v1/head_test.py
@@ -15,6 +15,7 @@
 from absl.testing import parameterized
 
 import tensorflow as tf
+from tensorflow import estimator as tf_estimator
 from tensorflow_privacy.privacy.estimators import test_utils
 from tensorflow_privacy.privacy.estimators.v1 import head as head_lib
 from tensorflow_privacy.privacy.optimizers.dp_optimizer import DPGradientDescentGaussianOptimizer
@@ -69,7 +70,7 @@ class DPHeadTest(tf.test.TestCase, parameterized.TestCase):
         noise_multiplier=0.0,
         num_microbatches=2)
     model_fn = make_model_fn(head, optimizer, feature_columns)
-    classifier = tf.estimator.Estimator(model_fn=model_fn)
+    classifier = tf_estimator.Estimator(model_fn=model_fn)
 
     classifier.train(
         input_fn=test_utils.make_input_fn(train_features, train_labels, True),
diff --git a/tensorflow_privacy/privacy/optimizers/BUILD b/tensorflow_privacy/privacy/optimizers/BUILD
index e6ca5ce..17d6742 100644
--- a/tensorflow_privacy/privacy/optimizers/BUILD
+++ b/tensorflow_privacy/privacy/optimizers/BUILD
@@ -53,6 +53,8 @@ py_test(
     deps = [
         ":dp_optimizer",
         "//tensorflow_privacy/privacy/dp_query:gaussian_query",
+        "//third_party/py/tensorflow:tensorflow_compat_v1_estimator",
+        "//third_party/py/tensorflow:tensorflow_estimator",
     ],
 )
 
@@ -62,7 +64,11 @@ py_test(
     srcs = ["dp_optimizer_vectorized_test.py"],
     python_version = "PY3",
     srcs_version = "PY3",
-    deps = [":dp_optimizer_vectorized"],
+    deps = [
+        ":dp_optimizer_vectorized",
+        "//third_party/py/tensorflow:tensorflow_compat_v1_estimator",
+        "//third_party/py/tensorflow:tensorflow_estimator",
+    ],
 )
 
 py_test(
@@ -86,5 +92,6 @@ py_test(
     deps = [
         "//tensorflow_privacy/privacy/optimizers:dp_optimizer_keras",
         "//tensorflow_privacy/privacy/optimizers:dp_optimizer_keras_vectorized",
+        "//third_party/py/tensorflow:tensorflow_estimator",
     ],
 )
diff --git a/tensorflow_privacy/privacy/optimizers/dp_optimizer_keras_test.py b/tensorflow_privacy/privacy/optimizers/dp_optimizer_keras_test.py
index 718684b..8417750 100644
--- a/tensorflow_privacy/privacy/optimizers/dp_optimizer_keras_test.py
+++ b/tensorflow_privacy/privacy/optimizers/dp_optimizer_keras_test.py
@@ -15,6 +15,7 @@
 from absl.testing import parameterized
 import numpy as np
 import tensorflow as tf
+from tensorflow import estimator as tf_estimator
 from tensorflow_privacy.privacy.optimizers import dp_optimizer_keras
 from tensorflow_privacy.privacy.optimizers import dp_optimizer_keras_vectorized
 
@@ -227,7 +228,7 @@ class DPOptimizerGetGradientsTest(tf.test.TestCase, parameterized.TestCase):
       train_op = tf.group(
           optimizer.get_updates(loss=vector_loss, params=params),
           [tf.compat.v1.assign_add(global_step, 1)])
-      return tf.estimator.EstimatorSpec(
+      return tf_estimator.EstimatorSpec(
           mode=mode, loss=scalar_loss, train_op=train_op)
 
     return linear_model_fn
@@ -249,7 +250,7 @@ class DPOptimizerGetGradientsTest(tf.test.TestCase, parameterized.TestCase):
   def testBaseline(self, cls, num_microbatches):
     """Tests that DP optimizers work with tf.estimator."""
 
-    linear_regressor = tf.estimator.Estimator(
+    linear_regressor = tf_estimator.Estimator(
         model_fn=self._make_linear_model_fn(cls, 100.0, 0.0, num_microbatches,
                                             0.05))
 
@@ -293,7 +294,7 @@ class DPOptimizerGetGradientsTest(tf.test.TestCase, parameterized.TestCase):
       return tf.data.Dataset.from_tensor_slices(
           (train_data, train_labels)).batch(1)
 
-    unclipped_linear_regressor = tf.estimator.Estimator(
+    unclipped_linear_regressor = tf_estimator.Estimator(
         model_fn=self._make_linear_model_fn(cls, 1.0e9, 0.0, num_microbatches,
                                             1.0))
     unclipped_linear_regressor.train(input_fn=train_input_fn, steps=1)
@@ -302,7 +303,7 @@ class DPOptimizerGetGradientsTest(tf.test.TestCase, parameterized.TestCase):
     bias_value = unclipped_linear_regressor.get_variable_value('dense/bias')
     global_norm = np.linalg.norm(np.concatenate((kernel_value, [bias_value])))
 
-    clipped_linear_regressor = tf.estimator.Estimator(
+    clipped_linear_regressor = tf_estimator.Estimator(
         model_fn=self._make_linear_model_fn(cls, 1.0, 0.0, num_microbatches,
                                             1.0))
     clipped_linear_regressor.train(input_fn=train_input_fn, steps=1)
@@ -339,7 +340,7 @@ class DPOptimizerGetGradientsTest(tf.test.TestCase, parameterized.TestCase):
                           num_microbatches):
     """Tests that DP optimizers work with tf.estimator."""
 
-    linear_regressor = tf.estimator.Estimator(
+    linear_regressor = tf_estimator.Estimator(
         model_fn=self._make_linear_model_fn(
             cls,
             l2_norm_clip,
diff --git a/tensorflow_privacy/privacy/optimizers/dp_optimizer_test.py b/tensorflow_privacy/privacy/optimizers/dp_optimizer_test.py
index cfa9b89..3b1c04f 100644
--- a/tensorflow_privacy/privacy/optimizers/dp_optimizer_test.py
+++ b/tensorflow_privacy/privacy/optimizers/dp_optimizer_test.py
@@ -18,6 +18,8 @@ import unittest
 from absl.testing import parameterized
 import numpy as np
 import tensorflow as tf
+from tensorflow import estimator as tf_estimator
+from tensorflow.compat.v1 import estimator as tf_compat_v1_estimator
 from tensorflow_privacy.privacy.dp_query import gaussian_query
 from tensorflow_privacy.privacy.optimizers import dp_optimizer
 
@@ -205,10 +207,10 @@ class DPOptimizerTest(tf.test.TestCase, parameterized.TestCase):
           dp_sum_query, num_microbatches=1, learning_rate=1.0)
       global_step = tf.compat.v1.train.get_global_step()
       train_op = optimizer.minimize(loss=vector_loss, global_step=global_step)
-      return tf.estimator.EstimatorSpec(
+      return tf_estimator.EstimatorSpec(
           mode=mode, loss=scalar_loss, train_op=train_op)
 
-    linear_regressor = tf.estimator.Estimator(model_fn=linear_model_fn)
+    linear_regressor = tf_estimator.Estimator(model_fn=linear_model_fn)
     true_weights = np.array([[-5], [4], [3], [2]]).astype(np.float32)
     true_bias = 6.0
     train_data = np.random.normal(scale=3.0, size=(200, 4)).astype(np.float32)
@@ -217,7 +219,7 @@ class DPOptimizerTest(tf.test.TestCase, parameterized.TestCase):
                              true_weights) + true_bias + np.random.normal(
                                  scale=0.1, size=(200, 1)).astype(np.float32)
 
-    train_input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(
+    train_input_fn = tf_compat_v1_estimator.inputs.numpy_input_fn(
         x={'x': train_data},
         y=train_labels,
         batch_size=20,
diff --git a/tensorflow_privacy/privacy/optimizers/dp_optimizer_vectorized_test.py b/tensorflow_privacy/privacy/optimizers/dp_optimizer_vectorized_test.py
index 1e6ab9e..17119f7 100644
--- a/tensorflow_privacy/privacy/optimizers/dp_optimizer_vectorized_test.py
+++ b/tensorflow_privacy/privacy/optimizers/dp_optimizer_vectorized_test.py
@@ -17,6 +17,8 @@ import unittest
 from absl.testing import parameterized
 import numpy as np
 import tensorflow as tf
+from tensorflow import estimator as tf_estimator
+from tensorflow.compat.v1 import estimator as tf_compat_v1_estimator
 from tensorflow_privacy.privacy.optimizers import dp_optimizer_vectorized
 from tensorflow_privacy.privacy.optimizers.dp_optimizer_vectorized import VectorizedDPAdagrad
 from tensorflow_privacy.privacy.optimizers.dp_optimizer_vectorized import VectorizedDPAdam
@@ -144,10 +146,10 @@ class DPOptimizerTest(tf.test.TestCase, parameterized.TestCase):
           learning_rate=1.0)
       global_step = tf.compat.v1.train.get_global_step()
       train_op = optimizer.minimize(loss=vector_loss, global_step=global_step)
-      return tf.estimator.EstimatorSpec(
+      return tf_estimator.EstimatorSpec(
           mode=mode, loss=scalar_loss, train_op=train_op)
 
-    linear_regressor = tf.estimator.Estimator(model_fn=linear_model_fn)
+    linear_regressor = tf_estimator.Estimator(model_fn=linear_model_fn)
     true_weights = np.array([[-5], [4], [3], [2]]).astype(np.float32)
     true_bias = 6.0
     train_data = np.random.normal(scale=3.0, size=(200, 4)).astype(np.float32)
@@ -156,7 +158,7 @@ class DPOptimizerTest(tf.test.TestCase, parameterized.TestCase):
                              true_weights) + true_bias + np.random.normal(
                                  scale=0.1, size=(200, 1)).astype(np.float32)
 
-    train_input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(
+    train_input_fn = tf_compat_v1_estimator.inputs.numpy_input_fn(
         x={'x': train_data},
         y=train_labels,
         batch_size=20,
diff --git a/tensorflow_privacy/privacy/privacy_tests/membership_inference_attack/BUILD b/tensorflow_privacy/privacy/privacy_tests/membership_inference_attack/BUILD
index 1044742..6f4ffbf 100644
--- a/tensorflow_privacy/privacy/privacy_tests/membership_inference_attack/BUILD
+++ b/tensorflow_privacy/privacy/privacy_tests/membership_inference_attack/BUILD
@@ -113,6 +113,8 @@ py_test(
     deps = [
         ":membership_inference_attack",
         ":tf_estimator_evaluation",
+        "//third_party/py/tensorflow:tensorflow_compat_v1_estimator",
+        "//third_party/py/tensorflow:tensorflow_estimator",
     ],
 )
 
@@ -124,6 +126,7 @@ py_library(
         ":membership_inference_attack",
         ":utils",
         ":utils_tensorboard",
+        "//third_party/py/tensorflow:tensorflow_estimator",
     ],
 )
 
@@ -135,6 +138,8 @@ py_binary(
     deps = [
         ":membership_inference_attack",
         ":tf_estimator_evaluation",
+        "//third_party/py/tensorflow:tensorflow_compat_v1_estimator",
+        "//third_party/py/tensorflow:tensorflow_estimator",
     ],
 )
 
diff --git a/tensorflow_privacy/privacy/privacy_tests/membership_inference_attack/tf_estimator_evaluation.py b/tensorflow_privacy/privacy/privacy_tests/membership_inference_attack/tf_estimator_evaluation.py
index 1b72086..bf13558 100644
--- a/tensorflow_privacy/privacy/privacy_tests/membership_inference_attack/tf_estimator_evaluation.py
+++ b/tensorflow_privacy/privacy/privacy_tests/membership_inference_attack/tf_estimator_evaluation.py
@@ -19,6 +19,7 @@ from typing import Iterable
 from absl import logging
 import numpy as np
 import tensorflow as tf
+from tensorflow import estimator as tf_estimator
 from tensorflow_privacy.privacy.privacy_tests.membership_inference_attack import data_structures
 from tensorflow_privacy.privacy.privacy_tests.membership_inference_attack import membership_inference_attack as mia
 from tensorflow_privacy.privacy.privacy_tests.membership_inference_attack import utils
@@ -47,7 +48,7 @@ def calculate_losses(estimator, input_fn, labels):
   return pred, loss
 
 
-class MembershipInferenceTrainingHook(tf.estimator.SessionRunHook):
+class MembershipInferenceTrainingHook(tf_estimator.SessionRunHook):
   """Training hook to perform membership inference attack on epoch end."""
 
   def __init__(self,
diff --git a/tensorflow_privacy/privacy/privacy_tests/membership_inference_attack/tf_estimator_evaluation_example.py b/tensorflow_privacy/privacy/privacy_tests/membership_inference_attack/tf_estimator_evaluation_example.py
index 0763e91..1481af5 100644
--- a/tensorflow_privacy/privacy/privacy_tests/membership_inference_attack/tf_estimator_evaluation_example.py
+++ b/tensorflow_privacy/privacy/privacy_tests/membership_inference_attack/tf_estimator_evaluation_example.py
@@ -18,6 +18,8 @@ from absl import flags
 from absl import logging
 import numpy as np
 import tensorflow as tf
+from tensorflow import estimator as tf_estimator
+from tensorflow.compat.v1 import estimator as tf_compat_v1_estimator
 from tensorflow_privacy.privacy.privacy_tests.membership_inference_attack.data_structures import AttackType
 from tensorflow_privacy.privacy.privacy_tests.membership_inference_attack.data_structures import get_flattened_attack_metrics
 from tensorflow_privacy.privacy.privacy_tests.membership_inference_attack.data_structures import SlicingSpec
@@ -46,34 +48,34 @@ def small_cnn_fn(features, labels, mode):
   y = tf.keras.layers.Dense(64, activation='relu')(y)
   logits = tf.keras.layers.Dense(10)(y)
 
-  if mode != tf.estimator.ModeKeys.PREDICT:
+  if mode != tf_estimator.ModeKeys.PREDICT:
     vector_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
         labels=labels, logits=logits)
     scalar_loss = tf.reduce_mean(input_tensor=vector_loss)
 
   # Configure the training op (for TRAIN mode).
-  if mode == tf.estimator.ModeKeys.TRAIN:
+  if mode == tf_estimator.ModeKeys.TRAIN:
     optimizer = tf.train.MomentumOptimizer(
         learning_rate=FLAGS.learning_rate, momentum=0.9)
     global_step = tf.compat.v1.train.get_global_step()
     train_op = optimizer.minimize(loss=scalar_loss, global_step=global_step)
-    return tf.estimator.EstimatorSpec(
+    return tf_estimator.EstimatorSpec(
         mode=mode, loss=scalar_loss, train_op=train_op)
 
   # Add evaluation metrics (for EVAL mode).
-  elif mode == tf.estimator.ModeKeys.EVAL:
+  elif mode == tf_estimator.ModeKeys.EVAL:
     eval_metric_ops = {
         'accuracy':
             tf.metrics.accuracy(
                 labels=labels, predictions=tf.argmax(input=logits, axis=1))
     }
-    return tf.estimator.EstimatorSpec(
+    return tf_estimator.EstimatorSpec(
         mode=mode, loss=scalar_loss, eval_metric_ops=eval_metric_ops)
 
   # Output the prediction probability (for PREDICT mode).
-  elif mode == tf.estimator.ModeKeys.PREDICT:
+  elif mode == tf_estimator.ModeKeys.PREDICT:
     predictions = tf.nn.softmax(logits)
-    return tf.estimator.EstimatorSpec(mode=mode, predictions=predictions)
+    return tf_estimator.EstimatorSpec(mode=mode, predictions=predictions)
 
 
 def load_cifar10():
@@ -97,13 +99,13 @@ def main(unused_argv):
   x_train, y_train, x_test, y_test = load_cifar10()
 
   # Instantiate the tf.Estimator.
-  classifier = tf.estimator.Estimator(
+  classifier = tf_estimator.Estimator(
       model_fn=small_cnn_fn, model_dir=FLAGS.model_dir)
 
   # A function to construct input_fn given (data, label), to be used by the
   # membership inference training hook.
   def input_fn_constructor(x, y):
-    return tf.compat.v1.estimator.inputs.numpy_input_fn(
+    return tf_compat_v1_estimator.inputs.numpy_input_fn(
         x={'x': x}, y=y, shuffle=False)
 
   # Get hook for membership inference attack.
@@ -118,13 +120,13 @@ def main(unused_argv):
       tensorboard_merge_classifiers=FLAGS.tensorboard_merge_classifiers)
 
   # Create tf.Estimator input functions for the training and test data.
-  train_input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(
+  train_input_fn = tf_compat_v1_estimator.inputs.numpy_input_fn(
       x={'x': x_train},
       y=y_train,
       batch_size=FLAGS.batch_size,
       num_epochs=FLAGS.epochs,
       shuffle=True)
-  eval_input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(
+  eval_input_fn = tf_compat_v1_estimator.inputs.numpy_input_fn(
       x={'x': x_test}, y=y_test, num_epochs=1, shuffle=False)
 
   # Training loop.
diff --git a/tensorflow_privacy/privacy/privacy_tests/membership_inference_attack/tf_estimator_evaluation_test.py b/tensorflow_privacy/privacy/privacy_tests/membership_inference_attack/tf_estimator_evaluation_test.py
index 515ba12..66d337e 100644
--- a/tensorflow_privacy/privacy/privacy_tests/membership_inference_attack/tf_estimator_evaluation_test.py
+++ b/tensorflow_privacy/privacy/privacy_tests/membership_inference_attack/tf_estimator_evaluation_test.py
@@ -15,6 +15,8 @@
 from absl.testing import absltest
 import numpy as np
 import tensorflow as tf
+from tensorflow import estimator as tf_estimator
+from tensorflow.compat.v1 import estimator as tf_compat_v1_estimator
 from tensorflow_privacy.privacy.privacy_tests.membership_inference_attack import data_structures
 from tensorflow_privacy.privacy.privacy_tests.membership_inference_attack import tf_estimator_evaluation
 
@@ -44,18 +46,18 @@ class UtilsTest(absltest.TestCase):
       logits = tf.keras.layers.Dense(self.nclass)(input_layer)
 
       # Define the PREDICT mode becasue we only need that
-      if mode == tf.estimator.ModeKeys.PREDICT:
+      if mode == tf_estimator.ModeKeys.PREDICT:
         predictions = tf.nn.softmax(logits)
-        return tf.estimator.EstimatorSpec(mode=mode, predictions=predictions)
+        return tf_estimator.EstimatorSpec(mode=mode, predictions=predictions)
 
     # Define the classifier, input_fn for training and test data
-    self.classifier = tf.estimator.Estimator(model_fn=model_fn)
-    self.input_fn_train = tf.compat.v1.estimator.inputs.numpy_input_fn(
+    self.classifier = tf_estimator.Estimator(model_fn=model_fn)
+    self.input_fn_train = tf_compat_v1_estimator.inputs.numpy_input_fn(
         x={'x': self.train_data},
         y=self.train_labels,
         num_epochs=1,
         shuffle=False)
-    self.input_fn_test = tf.compat.v1.estimator.inputs.numpy_input_fn(
+    self.input_fn_test = tf_compat_v1_estimator.inputs.numpy_input_fn(
         x={'x': self.test_data},
         y=self.test_labels,
         num_epochs=1,
@@ -94,7 +96,7 @@ class UtilsTest(absltest.TestCase):
     """Test the attack on the final models."""
 
     def input_fn_constructor(x, y):
-      return tf.compat.v1.estimator.inputs.numpy_input_fn(
+      return tf_compat_v1_estimator.inputs.numpy_input_fn(
           x={'x': x}, y=y, shuffle=False)
 
     results = tf_estimator_evaluation.run_attack_on_tf_estimator_model(
diff --git a/tutorials/BUILD b/tutorials/BUILD
index ae8ff2f..4ab0c98 100644
--- a/tutorials/BUILD
+++ b/tutorials/BUILD
@@ -17,6 +17,7 @@ py_binary(
         ":mnist_dpsgd_tutorial_common",
         "//tensorflow_privacy/privacy/analysis:compute_dp_sgd_privacy_lib",
         "//tensorflow_privacy/privacy/optimizers:dp_optimizer",
+        "//third_party/py/tensorflow:tensorflow_estimator",
     ],
 )
 
@@ -61,6 +62,8 @@ py_binary(
     deps = [
         "//tensorflow_privacy/privacy/analysis:rdp_accountant",
         "//tensorflow_privacy/privacy/optimizers:dp_optimizer_vectorized",
+        "//third_party/py/tensorflow:tensorflow_compat_v1_estimator",
+        "//third_party/py/tensorflow:tensorflow_estimator",
     ],
 )
 
@@ -73,6 +76,7 @@ py_binary(
         ":mnist_dpsgd_tutorial_common",
         "//tensorflow_privacy/privacy/analysis:compute_dp_sgd_privacy_lib",
         "//tensorflow_privacy/privacy/optimizers:dp_optimizer",
+        "//third_party/py/tensorflow:tensorflow_estimator",
     ],
 )
 
@@ -84,6 +88,8 @@ py_binary(
     deps = [
         "//tensorflow_privacy/privacy/analysis:rdp_accountant",
         "//tensorflow_privacy/privacy/optimizers:dp_optimizer",
+        "//third_party/py/tensorflow:tensorflow_compat_v1_estimator",
+        "//third_party/py/tensorflow:tensorflow_estimator",
     ],
 )
 
@@ -95,6 +101,8 @@ py_binary(
     deps = [
         "//tensorflow_privacy/privacy/analysis:rdp_accountant",
         "//tensorflow_privacy/privacy/optimizers:dp_optimizer",
+        "//third_party/py/tensorflow:tensorflow_compat_v1_estimator",
+        "//third_party/py/tensorflow:tensorflow_estimator",
     ],
 )
 
@@ -104,6 +112,8 @@ py_binary(
     deps = [
         "//tensorflow_privacy/privacy/analysis:gdp_accountant",
         "//tensorflow_privacy/privacy/optimizers:dp_optimizer",
+        "//third_party/py/tensorflow:tensorflow_compat_v1_estimator",
+        "//third_party/py/tensorflow:tensorflow_estimator",
     ],
 )
 
diff --git a/tutorials/lm_dpsgd_tutorial.py b/tutorials/lm_dpsgd_tutorial.py
index 448c70e..038b42b 100644
--- a/tutorials/lm_dpsgd_tutorial.py
+++ b/tutorials/lm_dpsgd_tutorial.py
@@ -37,6 +37,8 @@ from absl import flags
 from absl import logging
 import numpy as np
 import tensorflow as tf
+from tensorflow import estimator as tf_estimator
+from tensorflow.compat.v1 import estimator as tf_compat_v1_estimator
 import tensorflow_datasets as tfds
 from tensorflow_privacy.privacy.analysis.rdp_accountant import compute_rdp
 from tensorflow_privacy.privacy.analysis.rdp_accountant import get_privacy_spent
@@ -82,7 +84,7 @@ def rnn_model_fn(features, labels, mode):  # pylint: disable=unused-argument
   scalar_loss = tf.reduce_mean(vector_loss)
 
   # Configure the training op (for TRAIN mode).
-  if mode == tf.estimator.ModeKeys.TRAIN:
+  if mode == tf_estimator.ModeKeys.TRAIN:
     if FLAGS.dpsgd:
 
       optimizer = dp_optimizer.DPAdamGaussianOptimizer(
@@ -98,18 +100,18 @@ def rnn_model_fn(features, labels, mode):  # pylint: disable=unused-argument
       opt_loss = scalar_loss
     global_step = tf.compat.v1.train.get_global_step()
     train_op = optimizer.minimize(loss=opt_loss, global_step=global_step)
-    return tf.estimator.EstimatorSpec(
+    return tf_estimator.EstimatorSpec(
         mode=mode, loss=scalar_loss, train_op=train_op)
 
   # Add evaluation metrics (for EVAL mode).
-  elif mode == tf.estimator.ModeKeys.EVAL:
+  elif mode == tf_estimator.ModeKeys.EVAL:
     eval_metric_ops = {
         'accuracy':
             tf.metrics.accuracy(
                 labels=tf.cast(x[:, 1:], dtype=tf.int32),
                 predictions=tf.argmax(input=logits, axis=2))
     }
-    return tf.estimator.EstimatorSpec(
+    return tf_estimator.EstimatorSpec(
         mode=mode, loss=scalar_loss, eval_metric_ops=eval_metric_ops)
 
 
@@ -168,20 +170,20 @@ def main(unused_argv):
   train_data, test_data = load_data()
 
   # Instantiate the tf.Estimator.
-  conf = tf.estimator.RunConfig(save_summary_steps=1000)
-  lm_classifier = tf.estimator.Estimator(
+  conf = tf_estimator.RunConfig(save_summary_steps=1000)
+  lm_classifier = tf_estimator.Estimator(
       model_fn=rnn_model_fn, model_dir=FLAGS.model_dir, config=conf)
 
   # Create tf.Estimator input functions for the training and test data.
   batch_len = FLAGS.batch_size * SEQ_LEN
   train_data_end = len(train_data) - len(train_data) % batch_len
   test_data_end = len(test_data) - len(test_data) % batch_len
-  train_input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(
+  train_input_fn = tf_compat_v1_estimator.inputs.numpy_input_fn(
       x={'x': train_data[:train_data_end]},
       batch_size=batch_len,
       num_epochs=FLAGS.epochs,
       shuffle=False)
-  eval_input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(
+  eval_input_fn = tf_compat_v1_estimator.inputs.numpy_input_fn(
       x={'x': test_data[:test_data_end]},
       batch_size=batch_len,
       num_epochs=1,
diff --git a/tutorials/mnist_dpsgd_tutorial.py b/tutorials/mnist_dpsgd_tutorial.py
index 31ab5fa..3ba706c 100644
--- a/tutorials/mnist_dpsgd_tutorial.py
+++ b/tutorials/mnist_dpsgd_tutorial.py
@@ -19,6 +19,7 @@ from absl import app
 from absl import flags
 from absl import logging
 import tensorflow as tf
+from tensorflow import estimator as tf_estimator
 from tensorflow_privacy.privacy.analysis import compute_dp_sgd_privacy_lib
 from tensorflow_privacy.privacy.optimizers import dp_optimizer
 import mnist_dpsgd_tutorial_common as common
@@ -53,7 +54,7 @@ def cnn_model_fn(features, labels, mode, params):  # pylint: disable=unused-argu
   scalar_loss = tf.reduce_mean(input_tensor=vector_loss)
 
   # Configure the training op (for TRAIN mode).
-  if mode == tf.estimator.ModeKeys.TRAIN:
+  if mode == tf_estimator.ModeKeys.TRAIN:
     if FLAGS.dpsgd:
       # Use DP version of GradientDescentOptimizer. Other optimizers are
       # available in dp_optimizer. Most optimizers inheriting from
@@ -77,17 +78,17 @@ def cnn_model_fn(features, labels, mode, params):  # pylint: disable=unused-argu
     # the vector_loss because tf.estimator requires a scalar loss. This is only
     # used for evaluation and debugging by tf.estimator. The actual loss being
     # minimized is opt_loss defined above and passed to optimizer.minimize().
-    return tf.estimator.EstimatorSpec(
+    return tf_estimator.EstimatorSpec(
         mode=mode, loss=scalar_loss, train_op=train_op)
 
   # Add evaluation metrics (for EVAL mode).
-  elif mode == tf.estimator.ModeKeys.EVAL:
+  elif mode == tf_estimator.ModeKeys.EVAL:
     eval_metric_ops = {
         'accuracy':
             tf.metrics.accuracy(
                 labels=labels, predictions=tf.argmax(input=logits, axis=1))
     }
-    return tf.estimator.EstimatorSpec(
+    return tf_estimator.EstimatorSpec(
         mode=mode, loss=scalar_loss, eval_metric_ops=eval_metric_ops)
 
 
@@ -97,7 +98,7 @@ def main(unused_argv):
     raise ValueError('Number of microbatches should divide evenly batch_size')
 
   # Instantiate the tf.Estimator.
-  mnist_classifier = tf.estimator.Estimator(
+  mnist_classifier = tf_estimator.Estimator(
       model_fn=cnn_model_fn, model_dir=FLAGS.model_dir)
 
   # Training loop.
diff --git a/tutorials/mnist_dpsgd_tutorial_tpu.py b/tutorials/mnist_dpsgd_tutorial_tpu.py
index 5a5a0c6..74d2cd0 100644
--- a/tutorials/mnist_dpsgd_tutorial_tpu.py
+++ b/tutorials/mnist_dpsgd_tutorial_tpu.py
@@ -20,6 +20,7 @@ from absl import app
 from absl import flags
 from absl import logging
 import tensorflow as tf
+from tensorflow import estimator as tf_estimator
 from tensorflow_privacy.privacy.analysis import compute_dp_sgd_privacy_lib
 from tensorflow_privacy.privacy.optimizers import dp_optimizer
 import mnist_dpsgd_tutorial_common as common
@@ -56,7 +57,7 @@ def cnn_model_fn(features, labels, mode, params):  # pylint: disable=unused-argu
   scalar_loss = tf.reduce_mean(input_tensor=vector_loss)
 
   # Configure the training op (for TRAIN mode).
-  if mode == tf.estimator.ModeKeys.TRAIN:
+  if mode == tf_estimator.ModeKeys.TRAIN:
     if FLAGS.dpsgd:
       # Use DP version of GradientDescentOptimizer. Other optimizers are
       # available in dp_optimizer. Most optimizers inheriting from
@@ -84,11 +85,11 @@ def cnn_model_fn(features, labels, mode, params):  # pylint: disable=unused-argu
     # the vector_loss because tf.estimator requires a scalar loss. This is only
     # used for evaluation and debugging by tf.estimator. The actual loss being
     # minimized is opt_loss defined above and passed to optimizer.minimize().
-    return tf.estimator.tpu.TPUEstimatorSpec(
+    return tf_estimator.tpu.TPUEstimatorSpec(
         mode=mode, loss=scalar_loss, train_op=train_op)
 
   # Add evaluation metrics (for EVAL mode).
-  elif mode == tf.estimator.ModeKeys.EVAL:
+  elif mode == tf_estimator.ModeKeys.EVAL:
 
     def metric_fn(labels, logits):
       predictions = tf.argmax(logits, 1)
@@ -97,7 +98,7 @@ def cnn_model_fn(features, labels, mode, params):  # pylint: disable=unused-argu
               tf.metrics.accuracy(labels=labels, predictions=predictions),
       }
 
-    return tf.estimator.tpu.TPUEstimatorSpec(
+    return tf_estimator.tpu.TPUEstimatorSpec(
         mode=mode,
         loss=scalar_loss,
         eval_metrics=(metric_fn, {
@@ -112,8 +113,8 @@ def main(unused_argv):
     raise ValueError('Number of microbatches should divide evenly batch_size')
 
   # Instantiate the tf.Estimator.
-  run_config = tf.estimator.tpu.RunConfig(master=FLAGS.master)
-  mnist_classifier = tf.estimator.tpu.TPUEstimator(
+  run_config = tf_estimator.tpu.RunConfig(master=FLAGS.master)
+  mnist_classifier = tf_estimator.tpu.TPUEstimator(
       train_batch_size=FLAGS.batch_size,
       eval_batch_size=FLAGS.batch_size,
       model_fn=cnn_model_fn,
diff --git a/tutorials/mnist_dpsgd_tutorial_vectorized.py b/tutorials/mnist_dpsgd_tutorial_vectorized.py
index a12c7c5..cc15d82 100644
--- a/tutorials/mnist_dpsgd_tutorial_vectorized.py
+++ b/tutorials/mnist_dpsgd_tutorial_vectorized.py
@@ -18,6 +18,8 @@ from absl import flags
 from absl import logging
 import numpy as np
 import tensorflow as tf
+from tensorflow import estimator as tf_estimator
+from tensorflow.compat.v1 import estimator as tf_compat_v1_estimator
 from tensorflow_privacy.privacy.analysis.rdp_accountant import compute_rdp
 from tensorflow_privacy.privacy.analysis.rdp_accountant import get_privacy_spent
 from tensorflow_privacy.privacy.optimizers import dp_optimizer_vectorized
@@ -80,7 +82,7 @@ def cnn_model_fn(features, labels, mode):
   scalar_loss = tf.reduce_mean(input_tensor=vector_loss)
 
   # Configure the training op (for TRAIN mode).
-  if mode == tf.estimator.ModeKeys.TRAIN:
+  if mode == tf_estimator.ModeKeys.TRAIN:
 
     if FLAGS.dpsgd:
       # Use DP version of GradientDescentOptimizer. Other optimizers are
@@ -102,18 +104,18 @@ def cnn_model_fn(features, labels, mode):
     # the vector_loss because tf.estimator requires a scalar loss. This is only
     # used for evaluation and debugging by tf.estimator. The actual loss being
     # minimized is opt_loss defined above and passed to optimizer.minimize().
-    return tf.estimator.EstimatorSpec(
+    return tf_estimator.EstimatorSpec(
         mode=mode, loss=scalar_loss, train_op=train_op)
 
   # Add evaluation metrics (for EVAL mode).
-  elif mode == tf.estimator.ModeKeys.EVAL:
+  elif mode == tf_estimator.ModeKeys.EVAL:
     eval_metric_ops = {
         'accuracy':
             tf.metrics.accuracy(
                 labels=labels, predictions=tf.argmax(input=logits, axis=1))
     }
 
-    return tf.estimator.EstimatorSpec(
+    return tf_estimator.EstimatorSpec(
         mode=mode, loss=scalar_loss, eval_metric_ops=eval_metric_ops)
 
 
@@ -150,17 +152,17 @@ def main(unused_argv):
   train_data, train_labels, test_data, test_labels = load_mnist()
 
   # Instantiate the tf.Estimator.
-  mnist_classifier = tf.estimator.Estimator(
+  mnist_classifier = tf_estimator.Estimator(
       model_fn=cnn_model_fn, model_dir=FLAGS.model_dir)
 
   # Create tf.Estimator input functions for the training and test data.
-  train_input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(
+  train_input_fn = tf_compat_v1_estimator.inputs.numpy_input_fn(
       x={'x': train_data},
       y=train_labels,
       batch_size=FLAGS.batch_size,
       num_epochs=FLAGS.epochs,
       shuffle=True)
-  eval_input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(
+  eval_input_fn = tf_compat_v1_estimator.inputs.numpy_input_fn(
       x={'x': test_data}, y=test_labels, num_epochs=1, shuffle=False)
 
   # Training loop.
diff --git a/tutorials/mnist_lr_tutorial.py b/tutorials/mnist_lr_tutorial.py
index fff9718..557bf21 100644
--- a/tutorials/mnist_lr_tutorial.py
+++ b/tutorials/mnist_lr_tutorial.py
@@ -28,6 +28,8 @@ from absl import flags
 from absl import logging
 import numpy as np
 import tensorflow as tf
+from tensorflow import estimator as tf_estimator
+from tensorflow.compat.v1 import estimator as tf_compat_v1_estimator
 from tensorflow_privacy.privacy.analysis.rdp_accountant import compute_rdp
 from tensorflow_privacy.privacy.analysis.rdp_accountant import get_privacy_spent
 from tensorflow_privacy.privacy.optimizers import dp_optimizer
@@ -65,7 +67,7 @@ def lr_model_fn(features, labels, mode, nclasses, dim):
   scalar_loss = tf.reduce_mean(vector_loss)
 
   # Configure the training op (for TRAIN mode).
-  if mode == tf.estimator.ModeKeys.TRAIN:
+  if mode == tf_estimator.ModeKeys.TRAIN:
     if FLAGS.dpsgd:
       # The loss function is L-Lipschitz with L = sqrt(2*(||x||^2 + 1)) where
       # ||x|| is the norm of the data.
@@ -86,17 +88,17 @@ def lr_model_fn(features, labels, mode, nclasses, dim):
     # the vector_loss because tf.estimator requires a scalar loss. This is only
     # used for evaluation and debugging by tf.estimator. The actual loss being
     # minimized is opt_loss defined above and passed to optimizer.minimize().
-    return tf.estimator.EstimatorSpec(
+    return tf_estimator.EstimatorSpec(
         mode=mode, loss=scalar_loss, train_op=train_op)
 
   # Add evaluation metrics (for EVAL mode).
-  elif mode == tf.estimator.ModeKeys.EVAL:
+  elif mode == tf_estimator.ModeKeys.EVAL:
     eval_metric_ops = {
         'accuracy':
             tf.metrics.accuracy(
                 labels=labels, predictions=tf.argmax(input=logits, axis=1))
     }
-    return tf.estimator.EstimatorSpec(
+    return tf_estimator.EstimatorSpec(
         mode=mode, loss=scalar_loss, eval_metric_ops=eval_metric_ops)
 
 
@@ -199,19 +201,19 @@ def main(unused_argv):
   # pylint: disable=g-long-lambda
   model_fn = lambda features, labels, mode: lr_model_fn(
       features, labels, mode, nclasses=10, dim=train_data.shape[1:])
-  mnist_classifier = tf.estimator.Estimator(
+  mnist_classifier = tf_estimator.Estimator(
       model_fn=model_fn, model_dir=FLAGS.model_dir)
 
   # Create tf.Estimator input functions for the training and test data.
   # To analyze the per-user privacy loss, we keep the same orders of samples in
   # each epoch by setting shuffle=False.
-  train_input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(
+  train_input_fn = tf_compat_v1_estimator.inputs.numpy_input_fn(
       x={'x': train_data},
       y=train_labels,
       batch_size=FLAGS.batch_size,
       num_epochs=FLAGS.epochs,
       shuffle=False)
-  eval_input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(
+  eval_input_fn = tf_compat_v1_estimator.inputs.numpy_input_fn(
       x={'x': test_data}, y=test_labels, num_epochs=1, shuffle=False)
 
   # Train the model.
diff --git a/tutorials/movielens_tutorial.py b/tutorials/movielens_tutorial.py
index 2321abc..f39ec50 100644
--- a/tutorials/movielens_tutorial.py
+++ b/tutorials/movielens_tutorial.py
@@ -21,6 +21,8 @@ import pandas as pd
 from scipy import stats
 from sklearn.model_selection import train_test_split
 import tensorflow as tf
+from tensorflow import estimator as tf_estimator
+from tensorflow.compat.v1 import estimator as tf_compat_v1_estimator
 from tensorflow_privacy.privacy.analysis.gdp_accountant import compute_eps_poisson
 from tensorflow_privacy.privacy.analysis.gdp_accountant import compute_mu_poisson
 from tensorflow_privacy.privacy.optimizers import dp_optimizer
@@ -87,7 +89,7 @@ def nn_model_fn(features, labels, mode):
   scalar_loss = tf.reduce_mean(vector_loss)
 
   # Configure the training op (for TRAIN mode).
-  if mode == tf.estimator.ModeKeys.TRAIN:
+  if mode == tf_estimator.ModeKeys.TRAIN:
     if FLAGS.dpsgd:
       # Use DP version of GradientDescentOptimizer. Other optimizers are
       # available in dp_optimizer. Most optimizers inheriting from
@@ -110,11 +112,11 @@ def nn_model_fn(features, labels, mode):
     # the vector_loss because tf.estimator requires a scalar loss. This is only
     # used for evaluation and debugging by tf.estimator. The actual loss being
     # minimized is opt_loss defined above and passed to optimizer.minimize().
-    return tf.estimator.EstimatorSpec(
+    return tf_estimator.EstimatorSpec(
         mode=mode, loss=scalar_loss, train_op=train_op)
 
   # Add evaluation metrics (for EVAL mode).
-  if mode == tf.estimator.ModeKeys.EVAL:
+  if mode == tf_estimator.ModeKeys.EVAL:
     eval_metric_ops = {
         'rmse':
             tf.compat.v1.metrics.root_mean_squared_error(
@@ -124,7 +126,7 @@ def nn_model_fn(features, labels, mode):
                     b=tf.constant(np.array([0, 1, 2, 3, 4]), dtype=tf.float32),
                     axes=1))
     }
-    return tf.estimator.EstimatorSpec(
+    return tf_estimator.EstimatorSpec(
         mode=mode, loss=scalar_loss, eval_metric_ops=eval_metric_ops)
   return None
 
@@ -161,11 +163,11 @@ def main(unused_argv):
   train_data, test_data, _ = load_movielens()
 
   # Instantiate the tf.Estimator.
-  ml_classifier = tf.estimator.Estimator(
+  ml_classifier = tf_estimator.Estimator(
       model_fn=nn_model_fn, model_dir=FLAGS.model_dir)
 
   # Create tf.Estimator input functions for the training and test data.
-  eval_input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(
+  eval_input_fn = tf_compat_v1_estimator.inputs.numpy_input_fn(
       x={
           'user': test_data[:, 0],
           'movie': test_data[:, 4]
@@ -185,7 +187,7 @@ def main(unused_argv):
       global microbatches
       microbatches = len(subsampling)
 
-      train_input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(
+      train_input_fn = tf_compat_v1_estimator.inputs.numpy_input_fn(
           x={
               'user': train_data[subsampling, 0],
               'movie': train_data[subsampling, 4]
diff --git a/tutorials/walkthrough/BUILD b/tutorials/walkthrough/BUILD
index e1a6791..2d6e382 100644
--- a/tutorials/walkthrough/BUILD
+++ b/tutorials/walkthrough/BUILD
@@ -7,4 +7,8 @@ py_binary(
     srcs = ["mnist_scratch.py"],
     python_version = "PY3",
     srcs_version = "PY3",
+    deps = [
+        "//third_party/py/tensorflow:tensorflow_compat_v1_estimator",
+        "//third_party/py/tensorflow:tensorflow_estimator",
+    ],
 )
diff --git a/tutorials/walkthrough/mnist_scratch.py b/tutorials/walkthrough/mnist_scratch.py
index f387ddc..9cf8341 100644
--- a/tutorials/walkthrough/mnist_scratch.py
+++ b/tutorials/walkthrough/mnist_scratch.py
@@ -16,6 +16,8 @@
 from absl import logging
 import numpy as np
 import tensorflow as tf
+from tensorflow import estimator as tf_estimator
+from tensorflow.compat.v1 import estimator as tf_compat_v1_estimator
 
 tf.flags.DEFINE_float('learning_rate', .15, 'Learning rate for training')
 tf.flags.DEFINE_integer('batch_size', 256, 'Batch size')
@@ -45,22 +47,22 @@ def cnn_model_fn(features, labels, mode):
   scalar_loss = tf.reduce_mean(vector_loss)
 
   # Configure the training op (for TRAIN mode).
-  if mode == tf.estimator.ModeKeys.TRAIN:
+  if mode == tf_estimator.ModeKeys.TRAIN:
     optimizer = tf.compat.v1.train.GradientDescentOptimizer(FLAGS.learning_rate)
     opt_loss = scalar_loss
     global_step = tf.compat.v1.train.get_global_step()
     train_op = optimizer.minimize(loss=opt_loss, global_step=global_step)
-    return tf.estimator.EstimatorSpec(
+    return tf_estimator.EstimatorSpec(
         mode=mode, loss=scalar_loss, train_op=train_op)
 
   # Add evaluation metrics (for EVAL mode).
-  elif mode == tf.estimator.ModeKeys.EVAL:
+  elif mode == tf_estimator.ModeKeys.EVAL:
     eval_metric_ops = {
         'accuracy':
             tf.metrics.accuracy(
                 labels=labels, predictions=tf.argmax(input=logits, axis=1))
     }
-    return tf.estimator.EstimatorSpec(
+    return tf_estimator.EstimatorSpec(
         mode=mode, loss=scalar_loss, eval_metric_ops=eval_metric_ops)
 
 
@@ -94,16 +96,16 @@ def main(unused_argv):
   train_data, train_labels, test_data, test_labels = load_mnist()
 
   # Instantiate the tf.Estimator.
-  mnist_classifier = tf.estimator.Estimator(model_fn=cnn_model_fn)
+  mnist_classifier = tf_estimator.Estimator(model_fn=cnn_model_fn)
 
   # Create tf.Estimator input functions for the training and test data.
-  train_input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(
+  train_input_fn = tf_compat_v1_estimator.inputs.numpy_input_fn(
       x={'x': train_data},
       y=train_labels,
       batch_size=FLAGS.batch_size,
       num_epochs=FLAGS.epochs,
       shuffle=True)
-  eval_input_fn = tf.compat.v1.estimator.inputs.numpy_input_fn(
+  eval_input_fn = tf_compat_v1_estimator.inputs.numpy_input_fn(
       x={'x': test_data}, y=test_labels, num_epochs=1, shuffle=False)
 
   # Training loop.