1 # Copyright 2016 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 """Regression using the DNNRegressor Estimator.""" 16 17 from __future__ import absolute_import 18 from __future__ import division 19 from __future__ import print_function 20 21 import tensorflow as tf 22 23 import imports85 # pylint: disable=g-bad-import-order 24 25 STEPS = 5000 26 PRICE_NORM_FACTOR = 1000 27 28 29 def main(argv): 30 """Builds, trains, and evaluates the model.""" 31 assert len(argv) == 1 32 (train, test) = imports85.dataset() 33 34 # Switch the labels to units of thousands for better convergence. 35 def normalize_price(features, labels): 36 return features, labels / PRICE_NORM_FACTOR 37 38 train = train.map(normalize_price) 39 test = test.map(normalize_price) 40 41 # Build the training input_fn. 42 def input_train(): 43 return ( 44 # Shuffling with a buffer larger than the data set ensures 45 # that the examples are well mixed. 46 train.shuffle(1000).batch(128) 47 # Repeat forever 48 .repeat()) 49 50 # Build the validation input_fn. 51 def input_test(): 52 return test.shuffle(1000).batch(128) 53 54 # The first way assigns a unique weight to each category. To do this you must 55 # specify the category's vocabulary (values outside this specification will 56 # receive a weight of zero). Here we specify the vocabulary using a list of 57 # options. The vocabulary can also be specified with a vocabulary file (using 58 # `categorical_column_with_vocabulary_file`). For features covering a 59 # range of positive integers use `categorical_column_with_identity`. 60 body_style_vocab = ["hardtop", "wagon", "sedan", "hatchback", "convertible"] 61 body_style = tf.feature_column.categorical_column_with_vocabulary_list( 62 key="body-style", vocabulary_list=body_style_vocab) 63 make = tf.feature_column.categorical_column_with_hash_bucket( 64 key="make", hash_bucket_size=50) 65 66 feature_columns = [ 67 tf.feature_column.numeric_column(key="curb-weight"), 68 tf.feature_column.numeric_column(key="highway-mpg"), 69 # Since this is a DNN model, convert categorical columns from sparse 70 # to dense. 71 # Wrap them in an `indicator_column` to create a 72 # one-hot vector from the input. 73 tf.feature_column.indicator_column(body_style), 74 # Or use an `embedding_column` to create a trainable vector for each 75 # index. 76 tf.feature_column.embedding_column(make, dimension=3), 77 ] 78 79 # Build a DNNRegressor, with 2x20-unit hidden layers, with the feature columns 80 # defined above as input. 81 model = tf.estimator.DNNRegressor( 82 hidden_units=[20, 20], feature_columns=feature_columns) 83 84 # Train the model. 85 model.train(input_fn=input_train, steps=STEPS) 86 87 # Evaluate how the model performs on data it has not yet seen. 88 eval_result = model.evaluate(input_fn=input_test) 89 90 # The evaluation returns a Python dictionary. The "average_loss" key holds the 91 # Mean Squared Error (MSE). 92 average_loss = eval_result["average_loss"] 93 94 # Convert MSE to Root Mean Square Error (RMSE). 95 print("\n" + 80 * "*") 96 print("\nRMS error for the test set: ${:.0f}" 97 .format(PRICE_NORM_FACTOR * average_loss**0.5)) 98 99 print() 100 101 102 if __name__ == "__main__": 103 # The Estimator periodically generates "INFO" logs; make these logs visible. 104 tf.logging.set_verbosity(tf.logging.INFO) 105 tf.app.run(main=main) 106
