tensorflow55 《TensorFlow技术解析与实战》16 TensorFlow Debugger

01 环境

# 《TensorFlow技术解析与实战》15 TensorFlow线性代数编译框架XLA# win10 Tensorflow-gpu1.2.0-rc0 python3.5.3# CUDA v8.0 cudnn-8.0-windows10-x64-v5.1# https://github.com/tensorflow/tensorflow/blob/v1.2.0-rc0/tensorflow/python/debug/examples/debug_mnist.py# tensorflow/python/debug/examples目录下还有debug_errors.py、debug_fibonacci.py、debug_tflearn_iris.py3个测试文件可以研究。

debug_mnist.py中用到了readline函数，需要安装readline库。如果没有安装readline库，运行debug_mnist.py时，会提示ImportError: No module named 'readline'错误。
使用pip3 search readline 找到readline版本，采用pip3 install readline 安装readline，会提示readline不支持windows系统。阅读pip3 search readline结果，发现anyreadline比较合适（0.1.1版本）。安装pip3 install anyreadline后，运行 debug_mnist.py正常。

02 测试TensorFlow Debugger使用

启用TensorFlow Debugger，需要在run()之前加如下3行代码：

from tensorflow.python import debug as tf_debugsess = tf_debug.LocalCLIDebugWrapperSession(sess)# 注册过滤器has_inf_or_nan，它能判断出图中任何中间张量中是否有inf和nansess.add_tensor_filter("has_inf_or_nan", tf_debug.has_inf_or_nan)

02.01 运行python debug_mnist.py，结果如下：

# debug_mnist.py# Copyright 2016 The TensorFlow Authors. All Rights Reserved.## Licensed under the Apache License, Version 2.0 (the "License");# you may not use this file except in compliance with the License.# You may obtain a copy of the License at##     http://www.apache.org/licenses/LICENSE-2.0## Unless required by applicable law or agreed to in writing, software# distributed under the License is distributed on an "AS IS" BASIS,# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.# See the License for the specific language governing permissions and# limitations under the License.# =============================================================================="""Demo of the tfdbg curses CLI: Locating the source of bad numerical values.The neural network in this demo is larged based on the tutorial at:  tensorflow/examples/tutorials/mnist/mnist_with_summaries.pyBut modifications are made so that problematic numerical values (infs and nans)appear in nodes of the graph during training."""from __future__ import absolute_importfrom __future__ import divisionfrom __future__ import print_functionimport argparseimport sysimport tensorflow as tffrom tensorflow.examples.tutorials.mnist import input_datafrom tensorflow.python import debug as tf_debugIMAGE_SIZE = 28HIDDEN_SIZE = 500NUM_LABELS = 10RAND_SEED = 42def main(_):  # Import data  mnist = input_data.read_data_sets(FLAGS.data_dir,                                    one_hot=True,                                    fake_data=FLAGS.fake_data)  def feed_dict(train):    if train or FLAGS.fake_data:      xs, ys = mnist.train.next_batch(FLAGS.train_batch_size,                                      fake_data=FLAGS.fake_data)    else:      xs, ys = mnist.test.images, mnist.test.labels    return {x: xs, y_: ys}  sess = tf.InteractiveSession()  # Create the MNIST neural network graph.  # Input placeholders.  with tf.name_scope("input"):    x = tf.placeholder(        tf.float32, [None, IMAGE_SIZE * IMAGE_SIZE], name="x-input")    y_ = tf.placeholder(tf.float32, [None, NUM_LABELS], name="y-input")  def weight_variable(shape):    """Create a weight variable with appropriate initialization."""    initial = tf.truncated_normal(shape, stddev=0.1, seed=RAND_SEED)    return tf.Variable(initial)  def bias_variable(shape):    """Create a bias variable with appropriate initialization."""    initial = tf.constant(0.1, shape=shape)    return tf.Variable(initial)  def nn_layer(input_tensor, input_dim, output_dim, layer_name, act=tf.nn.relu):    """Reusable code for making a simple neural net layer."""    # Adding a name scope ensures logical grouping of the layers in the graph.    with tf.name_scope(layer_name):      # This Variable will hold the state of the weights for the layer      with tf.name_scope("weights"):        weights = weight_variable([input_dim, output_dim])      with tf.name_scope("biases"):        biases = bias_variable([output_dim])      with tf.name_scope("Wx_plus_b"):        preactivate = tf.matmul(input_tensor, weights) + biases      activations = act(preactivate)      return activations  hidden = nn_layer(x, IMAGE_SIZE**2, HIDDEN_SIZE, "hidden")  y = nn_layer(hidden, HIDDEN_SIZE, NUM_LABELS, "softmax", act=tf.nn.softmax)  with tf.name_scope("cross_entropy"):    # The following line is the culprit of the bad numerical values that appear    # during training of this graph. Log of zero gives inf, which is first seen    # in the intermediate tensor "cross_entropy/Log:0" during the 4th run()    # call. A multiplication of the inf values with zeros leads to nans,    # which is first in "cross_entropy/mul:0".    #    # You can use clipping to fix this issue, e.g.,    #   diff = y_ * tf.log(tf.clip_by_value(y, 1e-8, 1.0))    diff = y_ * tf.log(y)    with tf.name_scope("total"):      cross_entropy = -tf.reduce_mean(diff)  with tf.name_scope("train"):    train_step = tf.train.AdamOptimizer(FLAGS.learning_rate).minimize(        cross_entropy)  with tf.name_scope("accuracy"):    with tf.name_scope("correct_prediction"):      correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))    with tf.name_scope("accuracy"):      accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))  sess.run(tf.global_variables_initializer())  if FLAGS.debug:    sess = tf_debug.LocalCLIDebugWrapperSession(sess, ui_type=FLAGS.ui_type)    sess.add_tensor_filter("has_inf_or_nan", tf_debug.has_inf_or_nan)  # Add this point, sess is a debug wrapper around the actual Session if  # FLAGS.debug is true. In that case, calling run() will launch the CLI.  for i in range(FLAGS.max_steps):    acc = sess.run(accuracy, feed_dict=feed_dict(False))    print("Accuracy at step %d: %s" % (i, acc))    sess.run(train_step, feed_dict=feed_dict(True))if __name__ == "__main__":  parser = argparse.ArgumentParser()  parser.register("type", "bool", lambda v: v.lower() == "true")  parser.add_argument(      "--max_steps",      type=int,      default=10,      help="Number of steps to run trainer.")  parser.add_argument(      "--train_batch_size",      type=int,      default=100,      help="Batch size used during training.")  parser.add_argument(      "--learning_rate",      type=float,      default=0.025,      help="Initial learning rate.")  parser.add_argument(      "--data_dir",      type=str,      default="/tmp/mnist_data",      help="Directory for storing data")  parser.add_argument(      "--ui_type",      type=str,      default="curses",      help="Command-line user interface type (curses | readline)")  parser.add_argument(      "--fake_data",      type="bool",      nargs="?",      const=True,      default=False,      help="Use fake MNIST data for unit testing")  parser.add_argument(      "--debug",      type="bool",      nargs="?",      const=True,      default=False,      help="Use debugger to track down bad values during training")  FLAGS, unparsed = parser.parse_known_args()  tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)

02.02 进入debug调试界面

运行python debug_mnist.py --debug=True

02.03 run -f has_inf_or_nan

02.04 ni -t cross_entropy/Log

tfdbg> ni -t cross_entropy/LogNode cross_entropy/Log  Op: Log  Device: /job:localhost/replica:0/task:0/gpu:0  1 input(s) + 0 control input(s):    1 input(s):      [Softmax] softmax/Softmax  3 recipient(s) + 0 control recipient(s):    3 recipient(s):      [Shape] train/gradients/cross_entropy/mul_grad/Shape_1      [Mul] cross_entropy/mul      [Mul] train/gradients/cross_entropy/mul_grad/mulTraceback of node construction:0: debug_mnist.py  Line:     177  Function: <module>  Text:     "tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)"1: C:\Python35\lib\site-packages\tensorflow\python\platform\app.py  Line:     48  Function: run  Text:     "_sys.exit(main(_sys.argv[:1] + flags_passthrough))"2: debug_mnist.py  Line:     105  Function: main  Text:     "diff = y_ * tf.log(y)"3: C:\Python35\lib\site-packages\tensorflow\python\ops\gen_math_ops.py  Line:     1118  Function: log  Text:     "result = _op_def_lib.apply_op("Log", x=x, name=name)"4: C:\Python35\lib\site-packages\tensorflow\python\framework\op_def_library.py  Line:     767  Function: apply_op  Text:     "op_def=op_def)"5: C:\Python35\lib\site-packages\tensorflow\python\framework\ops.py  Line:     2506  Function: create_op  Text:     "original_op=self._default_original_op, op_def=op_def)"6: C:\Python35\lib\site-packages\tensorflow\python\framework\ops.py  Line:     1269  Function: __init__  Text:     "self._traceback = _extract_stack()"

如果所示：与debug_mnist.py相关的记录只有一条，并且指明是105行的”diff = y_ * tf.log(y)”。分析后，修改该语句为：diff = y_ * tf.log(tf.clip_by_value(y, 1e-8, 1.0))。重新运行。

02.05 python debug_mnist.py

02.06 tfdbg help

tfdbg> helpTTTTTT FFFF DDD  BBBB   GGG  TT   F    D  D B   B G  TT   FFF  D  D BBBB  G  GG  TT   F    D  D B   B G   G  TT   F    DDD  BBBB   GGGUpcoming run:======================================Session.run() call #1:Fetch(es):  accuracy/accuracy/Mean:0Feed dict(s):  input/x-input:0  input/y-input:0======================================Select one of the following commands to proceed ---->  run:    Execute the run() call with debug tensor-watching  run -n:    Execute the run() call without debug tensor-watching  run -t <T>:    Execute run() calls (T - 1) times without debugging, then execute run() once more with debugging and drop back to the CLI  run -f <filter_name>:    Keep executing run() calls until a dumped tensor passes a given, registered filter (conditional breakpoint mode)    Registered filter(s):        * has_inf_or_nan  invoke_stepper:    Use the node-stepper interface, which allows you to interactively step through nodes involved in the graph run() call and inspect/modify their valuesFor more details, see help..help  Aliases: h  Print this help message.invoke_stepper  Aliases: s  Invoke stepper (cont, step, breakpoint, etc.)  optional arguments:    -h, --help  show this help message and exitrun  Aliases: r  Run through, with or without debug tensor watching.  optional arguments:    -h, --help            show this help message and exit    -t TIMES, --times TIMES                          How many Session.run() calls to proceed with.    -n, --no_debug        Run through without debug tensor watching.    -f TILL_FILTER_PASS, --till_filter_pass TILL_FILTER_PASS                          Run until a tensor in the graph passes the specified                          filter.    --node_name_filter NODE_NAME_FILTER                          Regular-expression filter for node names to be watched                          in the run, e.g., loss, reshape.*    --op_type_filter OP_TYPE_FILTER                          Regular-expression filter for op type to be watched in                          the run, e.g., (MatMul|Add), Variable.*    --tensor_dtype_filter TENSOR_DTYPE_FILTER                          Regular-expression filter for tensor dtype to be                          watched in the run, e.g., (float32|float64), int.*    -p, --profile         Run and profile TensorFlow graph execution.run_info  Aliases: ri  Display information about this Session.run() call.  optional arguments:    -h, --help  show this help message and exit

03 远程调试方法(未测试)

采用tfdbg的offiline_analyzer。设置一个本地和远程主机都能访问的共享目录，如：/home/somebody/tfdbg_dumps_1。然后，通过debug_utils.watch_graph函数设置运行时的参数选项。在运行session.run()时就会将中间张量和运行时的图像转储到共享目录中。方法如下：

from tensorflow.python.debug import debug_utils# 此处代码如：构建图，生成session对象，略run_options = tf.RunOptions()debug_utils.watch_graph(  run_options,  session.graph,  debug_utils=[file:///home/somebody/tfdbg_dumps_1]) # 共享目录位置# 这里如果用多个客户端执行run，应该使用多个不同的共享目录session.run(fetches, feed_dict=feeds, options=run_options)

这样，在本地终端上就可以使用tfdbg的offline_analyzer来加载和检查共享目录中的数据。方法如下：

python -m tensorflow.python.debug.cli.offline_analyzer --dump_dir=/home/somebody/tfdbg_dumps_1

另外一种更简单和灵活的方法是使用回话的包装器函数DumpingDebugWrapperSession来在共享目录中产生训练中的累积文件。如下：

from tensorflow.python.debug import debug_utilssess = tf_debug.DumpingDebugWrapperSession(sess, "/home/somebody/tfdbg_dumps_1/", watch_fn=my_watch_fn)