1 # Copyright 2017 The TensorFlow Authors. All Rights Reserved. 2 # 3 # Licensed under the Apache License, Version 2.0 (the "License"); 4 # you may not use this file except in compliance with the License. 5 # You may obtain a copy of the License at 6 # 7 # http://www.apache.org/licenses/LICENSE-2.0 8 # 9 # Unless required by applicable law or agreed to in writing, software 10 # distributed under the License is distributed on an "AS IS" BASIS, 11 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 # See the License for the specific language governing permissions and 13 # limitations under the License. 14 # ============================================================================== 15 r"""Regression on Boston housing data using DNNBoostedTreeCombinedRegressor. 16 17 Example Usage: 18 19 python tensorflow/contrib/boosted_trees/examples/boston_combined.py \ 20 --batch_size=404 --output_dir="/tmp/boston" \ 21 --dnn_hidden_units="8,4" --dnn_steps_to_train=1000 \ 22 --tree_depth=4 --tree_learning_rate=0.1 \ 23 --num_trees=100 --tree_l2=0.001 --num_eval_steps=1 \ 24 --vmodule=training_ops=1 25 26 When training is done, mean squared error on eval data is reported. 27 Point tensorboard to the directory for the run to see how the training 28 progresses: 29 30 tensorboard --logdir=/tmp/boston 31 32 """ 33 from __future__ import absolute_import 34 from __future__ import division 35 from __future__ import print_function 36 37 import argparse 38 import sys 39 import tensorflow as tf 40 41 from tensorflow.contrib.boosted_trees.estimator_batch.dnn_tree_combined_estimator import DNNBoostedTreeCombinedRegressor 42 from tensorflow.contrib.boosted_trees.proto import learner_pb2 43 from tensorflow.contrib.layers.python.layers import feature_column 44 from tensorflow.contrib.learn.python.learn import learn_runner 45 from tensorflow.contrib.learn.python.learn.utils import input_fn_utils 46 from tensorflow.contrib.learn.python.learn.utils import saved_model_export_utils 47 48 _BOSTON_NUM_FEATURES = 13 49 50 51 def _get_estimator(output_dir, feature_cols): 52 """Configures DNNBoostedTreeCombinedRegressor based on flags.""" 53 learner_config = learner_pb2.LearnerConfig() 54 learner_config.learning_rate_tuner.fixed.learning_rate = ( 55 FLAGS.tree_learning_rate) 56 learner_config.regularization.l1 = 0.0 57 learner_config.regularization.l2 = FLAGS.tree_l2 58 learner_config.constraints.max_tree_depth = FLAGS.tree_depth 59 60 run_config = tf.contrib.learn.RunConfig(save_summary_steps=1) 61 62 # Create a DNNBoostedTreeCombinedRegressor estimator. 63 estimator = DNNBoostedTreeCombinedRegressor( 64 dnn_hidden_units=[int(x) for x in FLAGS.dnn_hidden_units.split(",")], 65 dnn_feature_columns=feature_cols, 66 tree_learner_config=learner_config, 67 num_trees=FLAGS.num_trees, 68 # This should be the number of examples. For large datasets it can be 69 # larger than the batch_size. 70 tree_examples_per_layer=FLAGS.batch_size, 71 model_dir=output_dir, 72 config=run_config, 73 dnn_input_layer_to_tree=True, 74 dnn_steps_to_train=FLAGS.dnn_steps_to_train) 75 return estimator 76 77 78 def _make_experiment_fn(output_dir): 79 """Creates experiment for DNNBoostedTreeCombinedRegressor.""" 80 (x_train, y_train), (x_test, 81 y_test) = tf.keras.datasets.boston_housing.load_data() 82 83 train_input_fn = tf.estimator.inputs.numpy_input_fn( 84 x={"x": x_train}, 85 y=y_train, 86 batch_size=FLAGS.batch_size, 87 num_epochs=None, 88 shuffle=True) 89 eval_input_fn = tf.estimator.inputs.numpy_input_fn( 90 x={"x": x_test}, y=y_test, num_epochs=1, shuffle=False) 91 92 feature_columns = [ 93 feature_column.real_valued_column("x", dimension=_BOSTON_NUM_FEATURES) 94 ] 95 feature_spec = tf.contrib.layers.create_feature_spec_for_parsing( 96 feature_columns) 97 serving_input_fn = input_fn_utils.build_parsing_serving_input_fn(feature_spec) 98 export_strategies = [ 99 saved_model_export_utils.make_export_strategy(serving_input_fn)] 100 return tf.contrib.learn.Experiment( 101 estimator=_get_estimator(output_dir, feature_columns), 102 train_input_fn=train_input_fn, 103 eval_input_fn=eval_input_fn, 104 train_steps=None, 105 eval_steps=FLAGS.num_eval_steps, 106 eval_metrics=None, 107 export_strategies=export_strategies) 108 109 110 def main(unused_argv): 111 learn_runner.run( 112 experiment_fn=_make_experiment_fn, 113 output_dir=FLAGS.output_dir, 114 schedule="train_and_evaluate") 115 116 117 if __name__ == "__main__": 118 tf.logging.set_verbosity(tf.logging.INFO) 119 parser = argparse.ArgumentParser() 120 # Define the list of flags that users can change. 121 parser.add_argument( 122 "--batch_size", 123 type=int, 124 default=1000, 125 help="The batch size for reading data.") 126 parser.add_argument( 127 "--output_dir", 128 type=str, 129 required=True, 130 help="Choose the dir for the output.") 131 parser.add_argument( 132 "--num_eval_steps", 133 type=int, 134 default=1, 135 help="The number of steps to run evaluation for.") 136 # Flags for configuring DNNBoostedTreeCombinedRegressor. 137 parser.add_argument( 138 "--dnn_hidden_units", 139 type=str, 140 default="8,4", 141 help="Hidden layers for DNN.") 142 parser.add_argument( 143 "--dnn_steps_to_train", 144 type=int, 145 default=1000, 146 help="Number of steps to train DNN.") 147 parser.add_argument( 148 "--tree_depth", type=int, default=4, help="Maximum depth of trees.") 149 parser.add_argument( 150 "--tree_l2", type=float, default=1.0, help="l2 regularization per batch.") 151 parser.add_argument( 152 "--tree_learning_rate", 153 type=float, 154 default=0.1, 155 help=("Learning rate (shrinkage weight) with which each " 156 "new tree is added.")) 157 parser.add_argument( 158 "--num_trees", 159 type=int, 160 default=None, 161 required=True, 162 help="Number of trees to grow before stopping.") 163 164 FLAGS, unparsed = parser.parse_known_args() 165 tf.app.run(main=main, argv=[sys.argv[0]] + unparsed) 166
