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 """Demo of the tfdbg curses CLI: Locating the source of bad numerical values. 16 17 The neural network in this demo is larged based on the tutorial at: 18 tensorflow/examples/tutorials/mnist/mnist_with_summaries.py 19 20 But modifications are made so that problematic numerical values (infs and nans) 21 appear in nodes of the graph during training. 22 """ 23 from __future__ import absolute_import 24 from __future__ import division 25 from __future__ import print_function 26 27 import argparse 28 import sys 29 30 import tensorflow as tf 31 32 from tensorflow.examples.tutorials.mnist import input_data 33 from tensorflow.python import debug as tf_debug 34 35 36 IMAGE_SIZE = 28 37 HIDDEN_SIZE = 500 38 NUM_LABELS = 10 39 RAND_SEED = 42 40 41 42 def main(_): 43 # Import data 44 mnist = input_data.read_data_sets(FLAGS.data_dir, 45 one_hot=True, 46 fake_data=FLAGS.fake_data) 47 48 def feed_dict(train): 49 if train or FLAGS.fake_data: 50 xs, ys = mnist.train.next_batch(FLAGS.train_batch_size, 51 fake_data=FLAGS.fake_data) 52 else: 53 xs, ys = mnist.test.images, mnist.test.labels 54 55 return {x: xs, y_: ys} 56 57 sess = tf.InteractiveSession() 58 59 # Create the MNIST neural network graph. 60 61 # Input placeholders. 62 with tf.name_scope("input"): 63 x = tf.placeholder( 64 tf.float32, [None, IMAGE_SIZE * IMAGE_SIZE], name="x-input") 65 y_ = tf.placeholder(tf.float32, [None, NUM_LABELS], name="y-input") 66 67 def weight_variable(shape): 68 """Create a weight variable with appropriate initialization.""" 69 initial = tf.truncated_normal(shape, stddev=0.1, seed=RAND_SEED) 70 return tf.Variable(initial) 71 72 def bias_variable(shape): 73 """Create a bias variable with appropriate initialization.""" 74 initial = tf.constant(0.1, shape=shape) 75 return tf.Variable(initial) 76 77 def nn_layer(input_tensor, input_dim, output_dim, layer_name, act=tf.nn.relu): 78 """Reusable code for making a simple neural net layer.""" 79 # Adding a name scope ensures logical grouping of the layers in the graph. 80 with tf.name_scope(layer_name): 81 # This Variable will hold the state of the weights for the layer 82 with tf.name_scope("weights"): 83 weights = weight_variable([input_dim, output_dim]) 84 with tf.name_scope("biases"): 85 biases = bias_variable([output_dim]) 86 with tf.name_scope("Wx_plus_b"): 87 preactivate = tf.matmul(input_tensor, weights) + biases 88 89 activations = act(preactivate) 90 return activations 91 92 hidden = nn_layer(x, IMAGE_SIZE**2, HIDDEN_SIZE, "hidden") 93 logits = nn_layer(hidden, HIDDEN_SIZE, NUM_LABELS, "output", tf.identity) 94 y = tf.nn.softmax(logits) 95 96 with tf.name_scope("cross_entropy"): 97 # The following line is the culprit of the bad numerical values that appear 98 # during training of this graph. Log of zero gives inf, which is first seen 99 # in the intermediate tensor "cross_entropy/Log:0" during the 4th run() 100 # call. A multiplication of the inf values with zeros leads to nans, 101 # which is first in "cross_entropy/mul:0". 102 # 103 # You can use the built-in, numerically-stable implementation to fix this 104 # issue: 105 # diff = tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=logits) 106 107 diff = -(y_ * tf.log(y)) 108 with tf.name_scope("total"): 109 cross_entropy = tf.reduce_mean(diff) 110 111 with tf.name_scope("train"): 112 train_step = tf.train.AdamOptimizer(FLAGS.learning_rate).minimize( 113 cross_entropy) 114 115 with tf.name_scope("accuracy"): 116 with tf.name_scope("correct_prediction"): 117 correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1)) 118 with tf.name_scope("accuracy"): 119 accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32)) 120 121 sess.run(tf.global_variables_initializer()) 122 123 if FLAGS.debug and FLAGS.tensorboard_debug_address: 124 raise ValueError( 125 "The --debug and --tensorboard_debug_address flags are mutually " 126 "exclusive.") 127 if FLAGS.debug: 128 sess = tf_debug.LocalCLIDebugWrapperSession(sess, ui_type=FLAGS.ui_type) 129 elif FLAGS.tensorboard_debug_address: 130 sess = tf_debug.TensorBoardDebugWrapperSession( 131 sess, FLAGS.tensorboard_debug_address) 132 133 # Add this point, sess is a debug wrapper around the actual Session if 134 # FLAGS.debug is true. In that case, calling run() will launch the CLI. 135 for i in range(FLAGS.max_steps): 136 acc = sess.run(accuracy, feed_dict=feed_dict(False)) 137 print("Accuracy at step %d: %s" % (i, acc)) 138 139 sess.run(train_step, feed_dict=feed_dict(True)) 140 141 142 if __name__ == "__main__": 143 parser = argparse.ArgumentParser() 144 parser.register("type", "bool", lambda v: v.lower() == "true") 145 parser.add_argument( 146 "--max_steps", 147 type=int, 148 default=10, 149 help="Number of steps to run trainer.") 150 parser.add_argument( 151 "--train_batch_size", 152 type=int, 153 default=100, 154 help="Batch size used during training.") 155 parser.add_argument( 156 "--learning_rate", 157 type=float, 158 default=0.025, 159 help="Initial learning rate.") 160 parser.add_argument( 161 "--data_dir", 162 type=str, 163 default="/tmp/mnist_data", 164 help="Directory for storing data") 165 parser.add_argument( 166 "--ui_type", 167 type=str, 168 default="curses", 169 help="Command-line user interface type (curses | readline)") 170 parser.add_argument( 171 "--fake_data", 172 type="bool", 173 nargs="?", 174 const=True, 175 default=False, 176 help="Use fake MNIST data for unit testing") 177 parser.add_argument( 178 "--debug", 179 type="bool", 180 nargs="?", 181 const=True, 182 default=False, 183 help="Use debugger to track down bad values during training. " 184 "Mutually exclusive with the --tensorboard_debug_address flag.") 185 parser.add_argument( 186 "--tensorboard_debug_address", 187 type=str, 188 default=None, 189 help="Connect to the TensorBoard Debugger Plugin backend specified by " 190 "the gRPC address (e.g., localhost:1234). Mutually exclusive with the " 191 "--debug flag.") 192 FLAGS, unparsed = parser.parse_known_args() 193 tf.app.run(main=main, argv=[sys.argv[0]] + unparsed) 194
