Use learn.DNNClassifier for Iris examples. Use default graph and existing global step (if there is one) in Estimator, because they may have been created earlier when constructing an optimizer, as in the example iris_custom_decay_dnn.

Change: 126369635

6 files changed, 79 insertions, 48 deletions

diff --git a/tensorflow/examples/skflow/BUILD b/tensorflow/examples/skflow/BUILD
index 5d6eae8745..7cac13df98 100644
--- a/tensorflow/examples/skflow/BUILD
+++ b/tensorflow/examples/skflow/BUILD
@@ -231,7 +231,11 @@ sh_test(
     data = [
         ":boston",
         ":iris",
+        ":iris_custom_decay_dnn",
         ":iris_custom_model",
+        ":iris_run_config",
+        ":iris_val_based_early_stopping",
+        ":iris_with_pipeline",
         ":text_classification",
         ":text_classification_builtin_rnn_model",
         ":text_classification_character_cnn",
diff --git a/tensorflow/examples/skflow/examples_test.sh b/tensorflow/examples/skflow/examples_test.sh
index da6b35c9bb..f4010c915e 100755
--- a/tensorflow/examples/skflow/examples_test.sh
+++ b/tensorflow/examples/skflow/examples_test.sh
@@ -49,6 +49,10 @@ function test() {
 test boston
 test iris
 test iris_custom_model
+test iris_custom_decay_dnn
+test iris_run_config
+test iris_val_based_early_stopping
+test iris_with_pipeline
 test text_classification --test_with_fake_data
 test text_classification_builtin_rnn_model --test_with_fake_data
 test text_classification_cnn --test_with_fake_data
diff --git a/tensorflow/examples/skflow/iris_custom_decay_dnn.py b/tensorflow/examples/skflow/iris_custom_decay_dnn.py
index c1e7d22d53..1ce6a830e4 100644
--- a/tensorflow/examples/skflow/iris_custom_decay_dnn.py
+++ b/tensorflow/examples/skflow/iris_custom_decay_dnn.py
@@ -17,24 +17,29 @@ from __future__ import print_function
 
 from sklearn import datasets, metrics
 from sklearn.cross_validation import train_test_split
-
 import tensorflow as tf
 
-iris = datasets.load_iris()
-X_train, X_test, y_train, y_test = train_test_split(iris.data,
-                                                    iris.target,
-                                                    test_size=0.2,
-                                                    random_state=42)
-# setup exponential decay function
-def exp_decay(global_step):
-    return tf.train.exponential_decay(
-        learning_rate=0.1, global_step=global_step,
-        decay_steps=100, decay_rate=0.001)
-
-# use customized decay function in learning_rate
-optimizer = tf.train.AdagradOptimizer(learning_rate=exp_decay)
-classifier = tf.contrib.learn.DNNClassifier(hidden_units=[10, 20, 10],
-                                            n_classes=3,
-                                            optimizer=optimizer)
-classifier.fit(X_train, y_train, steps=800)
-score = metrics.accuracy_score(y_test, classifier.predict(X_test))
+
+def optimizer_exp_decay():
+  global_step = tf.contrib.framework.get_or_create_global_step()
+  learning_rate = tf.train.exponential_decay(
+      learning_rate=0.1, global_step=global_step,
+      decay_steps=100, decay_rate=0.001)
+  return tf.train.AdagradOptimizer(learning_rate=learning_rate)
+
+def main(unused_argv):
+  iris = datasets.load_iris()
+  x_train, x_test, y_train, y_test = train_test_split(
+      iris.data, iris.target, test_size=0.2, random_state=42)
+
+  classifier = tf.contrib.learn.DNNClassifier(hidden_units=[10, 20, 10],
+                                              n_classes=3,
+                                              optimizer=optimizer_exp_decay)
+
+  classifier.fit(x_train, y_train, steps=800)
+  score = metrics.accuracy_score(y_test, classifier.predict(x_test))
+  print('Accuracy: {0:f}'.format(score))
+
+
+if __name__ == '__main__':
+  tf.app.run()
diff --git a/tensorflow/examples/skflow/iris_run_config.py b/tensorflow/examples/skflow/iris_run_config.py
index dff0daf9e8..c678c7c738 100644
--- a/tensorflow/examples/skflow/iris_run_config.py
+++ b/tensorflow/examples/skflow/iris_run_config.py
@@ -16,24 +16,31 @@ from __future__ import division
 from __future__ import print_function
 
 from sklearn import datasets, metrics, cross_validation
+import tensorflow as tf
 
-from tensorflow.contrib import learn
 
+def main(unused_argv):
+  # Load dataset.
+  iris = datasets.load_iris()
+  x_train, x_test, y_train, y_test = cross_validation.train_test_split(
+      iris.data, iris.target, test_size=0.2, random_state=42)
 
-# Load dataset.
-iris = datasets.load_iris()
-X_train, X_test, y_train, y_test = cross_validation.train_test_split(iris.data, iris.target,
-    test_size=0.2, random_state=42)
+  # You can define you configurations by providing a RunConfig object to
+  # estimator to control session configurations, e.g. num_cores
+  # and gpu_memory_fraction
+  run_config = tf.contrib.learn.estimators.RunConfig(
+      num_cores=3, gpu_memory_fraction=0.6)
 
-# You can define you configurations by providing a RunConfig object to
-# estimator to control session configurations, e.g. num_cores and gpu_memory_fraction
-run_config = learn.estimators.RunConfig(num_cores=3, gpu_memory_fraction=0.6)
+  # Build 3 layer DNN with 10, 20, 10 units respectively.
+  classifier = tf.contrib.learn.DNNClassifier(hidden_units=[10, 20, 10],
+                                              n_classes=3,
+                                              config=run_config)
 
-# Build 3 layer DNN with 10, 20, 10 units respectively.
-classifier = learn.TensorFlowDNNClassifier(hidden_units=[10, 20, 10],
-    n_classes=3, steps=200, config=run_config)
+  # Fit and predict.
+  classifier.fit(x_train, y_train, steps=200)
+  score = metrics.accuracy_score(y_test, classifier.predict(x_test))
+  print('Accuracy: {0:f}'.format(score))
 
-# Fit and predict.
-classifier.fit(X_train, y_train)
-score = metrics.accuracy_score(y_test, classifier.predict(X_test))
-print('Accuracy: {0:f}'.format(score))
+
+if __name__ == '__main__':
+  tf.app.run()
diff --git a/tensorflow/examples/skflow/iris_val_based_early_stopping.py b/tensorflow/examples/skflow/iris_val_based_early_stopping.py
index 72e0595544..05dfa96a07 100644
--- a/tensorflow/examples/skflow/iris_val_based_early_stopping.py
+++ b/tensorflow/examples/skflow/iris_val_based_early_stopping.py
@@ -34,21 +34,23 @@ def main(unused_argv):
       x_val, y_val, early_stopping_rounds=200)
 
   # classifier with early stopping on training data
-  classifier1 = learn.TensorFlowDNNClassifier(
+  classifier1 = learn.DNNClassifier(
       hidden_units=[10, 20, 10], n_classes=3, model_dir='/tmp/iris_model/')
   classifier1.fit(x=x_train, y=y_train, steps=2000)
   score1 = metrics.accuracy_score(y_test, classifier1.predict(x_test))
 
   # classifier with early stopping on validation data, save frequently for
   # monitor to pick up new checkpoints.
-  classifier2 = learn.TensorFlowDNNClassifier(
+  classifier2 = learn.DNNClassifier(
       hidden_units=[10, 20, 10], n_classes=3, model_dir='/tmp/iris_model_val/',
       config=tf.contrib.learn.RunConfig(save_checkpoints_secs=1))
   classifier2.fit(x=x_train, y=y_train, steps=2000, monitors=[val_monitor])
   score2 = metrics.accuracy_score(y_test, classifier2.predict(x_test))
 
   # In many applications, the score is improved by using early stopping
-  print(score2 > score1)
+  print('score1: ', score1)
+  print('score2: ', score2)
+  print('score2 > score1: ', score2 > score1)
 
 
 if __name__ == '__main__':
diff --git a/tensorflow/examples/skflow/iris_with_pipeline.py b/tensorflow/examples/skflow/iris_with_pipeline.py
index 3ba5739250..5535cd9e3b 100644
--- a/tensorflow/examples/skflow/iris_with_pipeline.py
+++ b/tensorflow/examples/skflow/iris_with_pipeline.py
@@ -20,22 +20,31 @@ from sklearn.datasets import load_iris
 from sklearn import cross_validation
 from sklearn.preprocessing import StandardScaler
 from sklearn.metrics import accuracy_score
+import tensorflow as tf
+
 from tensorflow.contrib import learn
 
-iris = load_iris()
-X_train, X_test, y_train, y_test = cross_validation.train_test_split(iris.data, iris.target,
-    test_size=0.2, random_state=42)
 
-# It's useful to scale to ensure Stochastic Gradient Descent will do the right thing
-scaler = StandardScaler()
+def main(unused_argv):
+  iris = load_iris()
+  x_train, x_test, y_train, y_test = cross_validation.train_test_split(
+      iris.data, iris.target, test_size=0.2, random_state=42)
+
+  # It's useful to scale to ensure Stochastic Gradient Descent
+  # will do the right thing.
+  scaler = StandardScaler()
+
+  # DNN classifier
+  classifier = learn.DNNClassifier(hidden_units=[10, 20, 10], n_classes=3)
 
-# DNN classifier
-DNNclassifier = learn.TensorFlowDNNClassifier(hidden_units=[10, 20, 10], n_classes=3, steps=200)
+  pipeline = Pipeline([('scaler', scaler),
+                       ('DNNclassifier', classifier)])
 
-pipeline = Pipeline([('scaler', scaler), ('DNNclassifier', DNNclassifier)])
+  pipeline.fit(x_train, y_train, DNNclassifier__steps=200)
 
-pipeline.fit(X_train, y_train)
+  score = accuracy_score(y_test, pipeline.predict(x_test))
+  print('Accuracy: {0:f}'.format(score))
 
-score = accuracy_score(y_test, pipeline.predict(X_test))
 
-print('Accuracy: {0:f}'.format(score))
+if __name__ == '__main__':
+  tf.app.run()