学习笔记TF063:TensorFlow Debugger

TensorFlow Debugger(tfdbg)，TensorFlow专用调试器。用断点、计算机图形化展现实时数据流，可视化运行TensorFlow图形内部结构、状态。有助训练推理调试模型错误。 。


常见错误类型，非数字(nan)、无限值(inf)。tfdbg命令行界面(command line interface,CLI)。


Debugger示例。错误运行MNIST训练，通过TensorFlow Debugger找到出错地方，改正。 。


先直接执行。


    python -m tensorflow.python.debug.examples.debug_mnist


准确率第一次训练上千，后面保持较低水平。
TensorFlow Debugger，在每次调用run()前后基于终端用户界面(UI)，控制执行、检查图内部状态。


    from tensorflow.python import debug as tf_debug
    sess = tr.debug.LocalCLIDebugWrapperSession(sess)
    sess.add_tensor_filter("has_inf_or_nan", tf_debug.has_inf_or_nan)


张量值注册过滤器has_inf_on_nan，判断图中间张量是否有nan、inf值。
开启调试模式(debug)。


    python -m tensorflow.python.debug.examples.debug_mnist -debug
    python debug_mnist.py --debug=True


运行开始UI(run-start UI)，在tfdbg>后输入交互式命令，run()进入运行结束后UI(run-end UI)。连续运行10次


    tfdbg>run -t 10
找出图形第一个nan或inf值


    tfdbg> run -f has_inf_or_nan


第一行灰底字表示tfdbg在调用run()后立即停止，生成指定过滤器has_inf_or_nan中间张量。第4次调用run()，36个中间张量包含inf或nan值。首次出现在cross_entropy/Log:0。单击图中cross_entropy/Log:0,单击下划线node_info菜单项，看节点输入张量，是否有0值。


    tfdbg>pt softmax/Softmax:0


用ni命令-t标志追溯
    ni -t cross_entropy/Log
问题代码


    diff = -(y_ * tf.log(y))
修改，对tf.log输入值裁剪


    diff = y_ * tf.log(tf.clip_by_value(y, 1e-8, 1.0))


    from __future__ import absolute_import
    from __future__ import division
    from __future__ import print_function
    import argparse
    import sys
    import tensorflow as tf
    from tensorflow.examples.tutorials.mnist import input_data
    from tensorflow.python import debug as tf_debug
    IMAGE_SIZE = 28
    HIDDEN_SIZE = 500
    NUM_LABELS = 10
    RAND_SEED = 42
    def 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")
      logits = nn_layer(hidden, HIDDEN_SIZE, NUM_LABELS, "output", tf.identity)
      y = tf.nn.softmax(logits)
      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 the built-in, numerically-stable implementation to fix this
        # issue:
        #   diff = tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=logits)
        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)
      # 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)
远程调试。tfdbg offline_analyzer。设置本地、远程机器能访问共享目录。debug_utils.watch_graph函数设置运行时参数选项。运行session.run()，中间张量、运行时图像转储到共享目录。本地终端用tfdbg offline_analyzer加载、检查共享目录数据。


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


    from tensorflow.python.debug.lib import debug_utils
    # 构建图，生成session对象，省略
    run_options = tf.RunOptions()
          debug_utils.watch_graph(
              run_options,
              sess.graph,
             # 共享目录位置
              # 多个客户端执行run，应用多个不同共享目录
              debug_urls=["file://home/somebody/tfdbg_dumps_1"])
    session.run(fetches, feed_dict=feeds, options=run_options)
或用会话包装器函数DumpingDebugWrapperSession在共享目录产生训练累积文件。


    from tensorflow.python.debug.lib import debug_utils
    sess = tf_debug.DumpingDebugWrapperSession(sess, "/home/somebody/tfdbg_dumps_1", watch_fn=my_watch_fn)
    
参考资料：
《TensorFlow技术解析与实战》


欢迎推荐上海机器学习工作机会，我的微信：qingxingfengzi