Tensorflow官网tfdbg的使用介绍

Love The Way You Lie 2022-05-14 04:26 241阅读 0赞

直接上一个使用的例子吧，将下面的代码保存了tfdbgdemo.py

import numpy as np
    import tensorflow as tf
    from tensorflow.python import debug as tf_debug
    
    xs = np.linspace(-0.5, 0.49, 100)
    x = tf.placeholder(tf.float32, shape=[None], name="x")
    y = tf.placeholder(tf.float32, shape=[None], name="y")
    k = tf.Variable([0.0], name="k")
    y_hat = tf.multiply(k, x, name="y_hat")
    sse = tf.reduce_sum((y - y_hat) * (y - y_hat), name="sse")
    train_op = tf.train.GradientDescentOptimizer(learning_rate=0.02).minimize(sse)
    
    sess = tf.Session()
    sess.run(tf.global_variables_initializer())
    
    sess = tf_debug.LocalCLIDebugWrapperSession(sess,ui_type="readline")
    for _ in range(10):
    	sess.run(y_hat,feed_dict={x:xs,y:10*xs})
    	sess.run(train_op, feed_dict={x: xs, y: 42 * xs})

### 下面是运行结果 ###

第一步，启动Python调试，
 $> python tfdbgdemo.py –debug
 
 ...
 TTTTTT 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 GGG
 
 TensorFlow version: 1.10.0
 
 ======================================
 Session.run() call #1:
 
 Fetch(es):
 y_hat:0
 
 Feed dict:
 x:0
 y: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 values
 
 For more details, see help..
 
 
 第二步，运行程序
 tfdbg> run
 run-end: run #1: 1 fetch (y_hat:0); 2 feeds
 3 dumped tensor(s):
 
 t (ms) Size (B) Op type Tensor name
 [0.000] 170 VariableV2 k:0
 [33.354] 180 Identity k/read:0
 [55.531] 576 Mul y_hat:0
 
 
 第三步，打印想看的变量
 tfdbg> pt y_hat:0
 Tensor "y_hat:0:DebugIdentity":
 dtype: float32
 shape: (100,)
 
 array([-0., -0., -0., -0., -0., -0., -0., -0., -0., -0., -0., -0., -0.,
 -0., -0., -0., -0., -0., -0., -0., -0., -0., -0., -0., -0., -0.,
 -0., -0., -0., -0., -0., -0., -0., -0., -0., -0., -0., -0., -0.,
 -0., -0., -0., -0., -0., -0., -0., -0., -0., -0., -0., 0., 0.,
 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
 0., 0., 0., 0., 0., 0., 0., 0., 0.], dtype=float32)

Tensorflow官网tfdbg的使用介绍，原文在这里，[https://www.tensorflow.org/guide/debugger][https_www.tensorflow.org_guide_debugger]

因为reference要梯子，烦，所以在搬到这里！

### TensorFlow Debugger ###

[`tfdbg`][tfdbg] is a specialized debugger for TensorFlow. It lets you view the internal structure and states of running TensorFlow graphs during training and inference, which is difficult to debug with general-purpose debuggers such as Python's `pdb`due to TensorFlow's computation-graph paradigm.

This guide focuses on the command-line interface (CLI) of [`tfdbg`][tfdbg]. For guide on how to use the graphical user interface (GUI) of tfdbg, i.e., the **TensorBoard Debugger Plugin**, please visit [its README][].

**Note:** The TensorFlow debugger uses a [curses][]\-based text user interface. On Mac OS X, the `ncurses` library is required and can be installed with `brew install ncurses`. On Windows, curses isn't as well supported, so a [readline][]\-based interface can be used with tfdbg by installing `pyreadline` with `pip`. If you use Anaconda3, you can install it with a command such as`"C:\Program Files\Anaconda3\Scripts\pip.exe" install pyreadline`. Unofficial Windows curses packages can be downloaded [here][], then subsequently installed using `pip install <your_version>.whl`, however curses on Windows may not work as reliably as curses on Linux or Mac.

This tutorial demonstrates how to use the **tfdbg** CLI to debug the appearance of [`nan`s][nan_s] and [`inf`s][inf_s], a frequently-encountered type of bug in TensorFlow model development. The following example is for users who use the low-level[`Session`][Session] API of TensorFlow. Later sections of this document describe how to use **tfdbg** with higher-level APIs of TensorFlow, including [`tf.estimator`][tf.estimator], [`tf.keras`][tf.keras] / `keras` and [`tf.contrib.slim`][tf.contrib.slim]. To *observe* such an issue, run the following command without the debugger (the source code can be found [here][here 1]):

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

This code trains a simple neural network for MNIST digit image recognition. Notice that the accuracy increases slightly after the first training step, but then gets stuck at a low (near-chance) level:

`Accuracy at step 0: 0.1113 Accuracy at step 1: 0.3183 Accuracy at step 2: 0.098 Accuracy at step 3: 0.098 Accuracy at step 4: 0.098`

Wondering what might have gone wrong, you suspect that certain nodes in the training graph generated bad numeric values such as `inf`s and `nan`s, because this is a common cause of this type of training failure. Let's use tfdbg to debug this issue and pinpoint the exact graph node where this numeric problem first surfaced.

## Wrapping TensorFlow Sessions with tfdbg ##

To add support for tfdbg in our example, all that is needed is to add the following lines of code and wrap the Session object with a debugger wrapper. This code is already added in [debug\_mnist.py][here 1], so you can activate tfdbg CLI with the `--debug` flag at the command line.

`# Let your BUILD target depend on "//tensorflow/python/debug:debug_py" # (You don't need to worry about the BUILD dependency if you are using a pip #  install of open-source TensorFlow.) from tensorflow.python import debug as tf_debug sess = tf_debug.LocalCLIDebugWrapperSession(sess)`

This wrapper has the same interface as Session, so enabling debugging requires no other changes to the code. The wrapper provides additional features, including:

*  Bringing up a CLI before and after `Session.run()` calls, to let you control the execution and inspect the graph's internal state.
 *  Allowing you to register special `filters` for tensor values, to facilitate the diagnosis of issues.

In this example, we have already registered a tensor filter called [`tfdbg.has_inf_or_nan`][tfdbg.has_inf_or_nan], which simply determines if there are any `nan` or `inf` values in any intermediate tensors (tensors that are neither inputs or outputs of the`Session.run()` call, but are in the path leading from the inputs to the outputs). This filter is for `nan`s and `inf`s is a common enough use case that we ship it with the [`debug_data`][debug_data] module.

**Note:** You can also write your own custom filters. See the [API documentation][] of `DebugDumpDir.find()` for additional information.

## Debugging Model Training with tfdbg ##

Let's try training the model again, but with the `--debug` flag added this time:

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

The debug wrapper session will prompt you when it is about to execute the first `Session.run()` call, with information regarding the fetched tensor and feed dictionaries displayed on the screen.

![70][]

This is what we refer to as the *run-start CLI*. It lists the feeds and fetches to the current `Session.run` call, before executing anything.

If the screen size is too small to display the content of the message in its entirety, you can resize it.

Use the **PageUp** / **PageDown** / **Home** / **End** keys to navigate the screen output. On most keyboards lacking those keys **Fn + Up** / **Fn + Down** / **Fn + Right** / **Fn + Left** will work.

Enter the `run` command (or just `r`) at the command prompt:

`tfdbg> run`

The `run` command causes tfdbg to execute until the end of the next `Session.run()` call, which calculates the model's accuracy using a test data set. tfdbg augments the runtime Graph to dump all intermediate tensors. After the run ends, tfdbg displays all the dumped tensors values in the *run-end CLI*. For example:

![70 1][]

This list of tensors can also be obtained by running the command `lt` after you executed `run`.

### tfdbg CLI Frequently-Used Commands ###

Try the following commands at the `tfdbg>` prompt (referencing the code at`tensorflow/python/debug/examples/debug_mnist.py`):

<table> 
 <thead> 
 <tr> 
 <th>Command</th> 
 <th>Syntax or Option</th> 
 <th>Explanation</th> 
 <th>Example</th> 
 </tr> 
 </thead> 
 <tbody> 
 <tr> 
 <td><code>lt</code></td> 
 <td>&nbsp;</td> 
 <td>List dumped tensors.</td> 
 <td><code>lt</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-n &lt;name_pattern&gt;</code></td> 
 <td>List dumped tensors with names matching given regular-expression pattern.</td> 
 <td><code>lt -n Softmax.*</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-t &lt;op_pattern&gt;</code></td> 
 <td>List dumped tensors with op types matching given regular-expression pattern.</td> 
 <td><code>lt -t MatMul</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-f &lt;filter_name&gt;</code></td> 
 <td>List only the tensors that pass a registered tensor filter.</td> 
 <td><code>lt -f has_inf_or_nan</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-f &lt;filter_name&gt; -fenn &lt;regex&gt;</code></td> 
 <td>List only the tensors that pass a registered tensor filter, excluding nodes with names matching the regular expression.</td> 
 <td><code>lt -f has_inf_or_nan</code><code>-fenn&nbsp;.*Sqrt.*</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-s &lt;sort_key&gt;</code></td> 
 <td>Sort the output by given&nbsp;<code>sort_key</code>, whose possible values are&nbsp;<code>timestamp</code>&nbsp;(default),&nbsp;<code>dump_size</code>,&nbsp;<code>op_type</code>&nbsp;and&nbsp;<code>tensor_name</code>.</td> 
 <td><code>lt -s dump_size</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-r</code></td> 
 <td>Sort in reverse order.</td> 
 <td><code>lt -r -s dump_size</code></td> 
 </tr> 
 <tr> 
 <td><code>pt</code></td> 
 <td>&nbsp;</td> 
 <td>Print value of a dumped tensor.</td> 
 <td>&nbsp;</td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>pt &lt;tensor&gt;</code></td> 
 <td>Print tensor value.</td> 
 <td><code>pt hidden/Relu:0</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>pt &lt;tensor&gt;[slicing]</code></td> 
 <td>Print a subarray of tensor, using&nbsp;<a href="http://www.numpy.org/" rel="nofollow">numpy</a>-style array slicing.</td> 
 <td><code>pt hidden/Relu:0[0:50,:]</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-a</code></td> 
 <td>Print the entirety of a large tensor, without using ellipses. (May take a long time for large tensors.)</td> 
 <td><code>pt -a hidden/Relu:0[0:50,:]</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-r &lt;range&gt;</code></td> 
 <td>Highlight elements falling into specified numerical range. Multiple ranges can be used in conjunction.</td> 
 <td><code>pt hidden/Relu:0 -a -r [[-inf,-1],[1,inf]]</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-n &lt;number&gt;</code></td> 
 <td>Print dump corresponding to specified 0-based dump number. Required for tensors with multiple dumps.</td> 
 <td><code>pt -n 0 hidden/Relu:0</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-s</code></td> 
 <td>Include a summary of the numeric values of the tensor (applicable only to non-empty tensors with Boolean and numeric types such as&nbsp;<code>int*</code>&nbsp;and&nbsp;<code>float*</code>.)</td> 
 <td><code>pt -s hidden/Relu:0[0:50,:]</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-w</code></td> 
 <td>Write the value of the tensor (possibly sliced) to a Numpy file using&nbsp;<a href="https://docs.scipy.org/doc/numpy-1.13.0/reference/generated/numpy.save.html" rel="nofollow"><code>numpy.save()</code></a></td> 
 <td><code>pt -s hidden/Relu:0 -w /tmp/relu.npy</code></td> 
 </tr> 
 <tr> 
 <td><code>@[coordinates]</code></td> 
 <td>&nbsp;</td> 
 <td>Navigate to specified element in&nbsp;<code>pt</code>&nbsp;output.</td> 
 <td><code>@[10,0]</code>&nbsp;or&nbsp;<code>@10,0</code></td> 
 </tr> 
 <tr> 
 <td><code>/regex</code></td> 
 <td>&nbsp;</td> 
 <td><a href="https://linux.die.net/man/1/less" rel="nofollow">less</a>-style search for given regular expression.</td> 
 <td><code>/inf</code></td> 
 </tr> 
 <tr> 
 <td><code>/</code></td> 
 <td>&nbsp;</td> 
 <td>Scroll to the next line with matches to the searched regex (if any).</td> 
 <td><code>/</code></td> 
 </tr> 
 <tr> 
 <td><code>pf</code></td> 
 <td>&nbsp;</td> 
 <td>Print a value in the feed_dict to&nbsp;<code>Session.run</code>.</td> 
 <td>&nbsp;</td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>pf &lt;feed_tensor_name&gt;</code></td> 
 <td>Print the value of the feed. Also note that the&nbsp;<code>pf</code>command has the&nbsp;<code>-a</code>,&nbsp;<code>-r</code>&nbsp;and&nbsp;<code>-s</code>&nbsp;flags (not listed below), which have the same syntax and semantics as the identically-named flags of&nbsp;<code>pt</code>.</td> 
 <td><code>pf input_xs:0</code></td> 
 </tr> 
 <tr> 
 <td>eval</td> 
 <td>&nbsp;</td> 
 <td>Evaluate arbitrary Python and numpy expression.</td> 
 <td>&nbsp;</td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>eval &lt;expression&gt;</code></td> 
 <td>Evaluate a Python / numpy expression, with numpy available as&nbsp;<code>np</code>&nbsp;and debug tensor names enclosed in backticks.</td> 
 <td><code>eval "np.matmul((`output/Identity:0` / `Softmax:0`).T,`Softmax:0`)"</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-a</code></td> 
 <td>Print a large-sized evaluation result in its entirety, i.e., without using ellipses.</td> 
 <td><code>eval -a 'np.sum(`Softmax:0`,&nbsp;axis=1)'</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-w</code></td> 
 <td>Write the result of the evaluation to a Numpy file using&nbsp;<a href="https://docs.scipy.org/doc/numpy-1.13.0/reference/generated/numpy.save.html" rel="nofollow"><code>numpy.save()</code></a></td> 
 <td><code>eval -a 'np.sum(`Softmax:0`,&nbsp;axis=1)' -w /tmp/softmax_sum.npy</code></td> 
 </tr> 
 <tr> 
 <td><code>ni</code></td> 
 <td>&nbsp;</td> 
 <td>Display node information.</td> 
 <td>&nbsp;</td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-a</code></td> 
 <td>Include node attributes in the output.</td> 
 <td><code>ni -a hidden/Relu</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-d</code></td> 
 <td>List the debug dumps available from the node.</td> 
 <td><code>ni -d hidden/Relu</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-t</code></td> 
 <td>Display the Python stack trace of the node's creation.</td> 
 <td><code>ni -t hidden/Relu</code></td> 
 </tr> 
 <tr> 
 <td><code>li</code></td> 
 <td>&nbsp;</td> 
 <td>List inputs to node</td> 
 <td>&nbsp;</td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-r</code></td> 
 <td>List the inputs to node, recursively (the input tree.)</td> 
 <td><code>li -r hidden/Relu:0</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-d &lt;max_depth&gt;</code></td> 
 <td>Limit recursion depth under the&nbsp;<code>-r</code>&nbsp;mode.</td> 
 <td><code>li -r -d 3 hidden/Relu:0</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-c</code></td> 
 <td>Include control inputs.</td> 
 <td><code>li -c -r hidden/Relu:0</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-t</code></td> 
 <td>Show op types of input nodes.</td> 
 <td><code>li -t -r hidden/Relu:0</code></td> 
 </tr> 
 <tr> 
 <td><code>lo</code></td> 
 <td>&nbsp;</td> 
 <td>List output recipients of node</td> 
 <td>&nbsp;</td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-r</code></td> 
 <td>List the output recipients of node, recursively (the output tree.)</td> 
 <td><code>lo -r hidden/Relu:0</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-d &lt;max_depth&gt;</code></td> 
 <td>Limit recursion depth under the&nbsp;<code>-r</code>&nbsp;mode.</td> 
 <td><code>lo -r -d 3 hidden/Relu:0</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-c</code></td> 
 <td>Include recipients via control edges.</td> 
 <td><code>lo -c -r hidden/Relu:0</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-t</code></td> 
 <td>Show op types of recipient nodes.</td> 
 <td><code>lo -t -r hidden/Relu:0</code></td> 
 </tr> 
 <tr> 
 <td><code>ls</code></td> 
 <td>&nbsp;</td> 
 <td>List Python source files involved in node creation.</td> 
 <td>&nbsp;</td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-p &lt;path_pattern&gt;</code></td> 
 <td>Limit output to source files matching given regular-expression path pattern.</td> 
 <td><code>ls -p&nbsp;.*debug_mnist.*</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-n</code></td> 
 <td>Limit output to node names matching given regular-expression pattern.</td> 
 <td><code>ls -n Softmax.*</code></td> 
 </tr> 
 <tr> 
 <td><code>ps</code></td> 
 <td>&nbsp;</td> 
 <td>Print Python source file.</td> 
 <td>&nbsp;</td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>ps &lt;file_path&gt;</code></td> 
 <td>Print given Python source file source.py, with the lines annotated with the nodes created at each of them (if any).</td> 
 <td><code>ps /path/to/source.py</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-t</code></td> 
 <td>Perform annotation with respect to Tensors, instead of the default, nodes.</td> 
 <td><code>ps -t /path/to/source.py</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-b &lt;line_number&gt;</code></td> 
 <td>Annotate source.py beginning at given line.</td> 
 <td><code>ps -b 30 /path/to/source.py</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-m &lt;max_elements&gt;</code></td> 
 <td>Limit the number of elements in the annotation for each line.</td> 
 <td><code>ps -m 100 /path/to/source.py</code></td> 
 </tr> 
 <tr> 
 <td><code>run</code></td> 
 <td>&nbsp;</td> 
 <td>Proceed to the next Session.run()</td> 
 <td><code>run</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-n</code></td> 
 <td>Execute through the next&nbsp;<code>Session.run</code>&nbsp;without debugging, and drop to CLI right before the run after that.</td> 
 <td><code>run -n</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-t &lt;T&gt;</code></td> 
 <td>Execute&nbsp;<code>Session.run</code>&nbsp;<code>T - 1</code>&nbsp;times without debugging, followed by a run with debugging. Then drop to CLI right after the debugged run.</td> 
 <td><code>run -t 10</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-f &lt;filter_name&gt;</code></td> 
 <td>Continue executing&nbsp;<code>Session.run</code>&nbsp;until any intermediate tensor triggers the specified Tensor filter (causes the filter to return&nbsp;<code>True</code>).</td> 
 <td><code>run -f has_inf_or_nan</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-f &lt;filter_name&gt; -fenn &lt;regex&gt;</code></td> 
 <td>Continue executing&nbsp;<code>Session.run</code>&nbsp;until any intermediate tensor whose node names doesn't match the regular expression triggers the specified Tensor filter (causes the filter to return&nbsp;<code>True</code>).</td> 
 <td><code>run -f has_inf_or_nan -fenn&nbsp;.*Sqrt.*</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>--node_name_filter &lt;pattern&gt;</code></td> 
 <td>Execute the next&nbsp;<code>Session.run</code>, watching only nodes with names matching the given regular-expression pattern.</td> 
 <td><code>run --node_name_filter Softmax.*</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>--op_type_filter &lt;pattern&gt;</code></td> 
 <td>Execute the next&nbsp;<code>Session.run</code>, watching only nodes with op types matching the given regular-expression pattern.</td> 
 <td><code>run --op_type_filter Variable.*</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>--tensor_dtype_filter &lt;pattern&gt;</code></td> 
 <td>Execute the next&nbsp;<code>Session.run</code>, dumping only Tensors with data types (<code>dtype</code>s) matching the given regular-expression pattern.</td> 
 <td><code>run --tensor_dtype_filter int.*</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>-p</code></td> 
 <td>Execute the next&nbsp;<code>Session.run</code>&nbsp;call in profiling mode.</td> 
 <td><code>run -p</code></td> 
 </tr> 
 <tr> 
 <td><code>ri</code></td> 
 <td>&nbsp;</td> 
 <td>Display information about the run the current run, including fetches and feeds.</td> 
 <td><code>ri</code></td> 
 </tr> 
 <tr> 
 <td><code>config</code></td> 
 <td>&nbsp;</td> 
 <td>Set or show persistent TFDBG UI configuration.</td> 
 <td>&nbsp;</td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>set</code></td> 
 <td>Set the value of a config item: { <code>graph_recursion_depth</code>,&nbsp;<code>mouse_mode</code>}.</td> 
 <td><code>config set graph_recursion_depth 3</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>show</code></td> 
 <td>Show current persistent UI configuration.</td> 
 <td><code>config show</code></td> 
 </tr> 
 <tr> 
 <td><code>version</code></td> 
 <td>&nbsp;</td> 
 <td>Print the version of TensorFlow and its key dependencies.</td> 
 <td><code>version</code></td> 
 </tr> 
 <tr> 
 <td><code>help</code></td> 
 <td>&nbsp;</td> 
 <td>Print general help information</td> 
 <td><code>help</code></td> 
 </tr> 
 <tr> 
 <td>&nbsp;</td> 
 <td><code>help &lt;command&gt;</code></td> 
 <td>Print help for given command.</td> 
 <td><code>help lt</code></td> 
 </tr> 
 </tbody> 
</table>

Note that each time you enter a command, a new screen output will appear. This is somewhat analogous to web pages in a browser. You can navigate between these screens by clicking the `<--` and `-->` text arrows near the top-left corner of the CLI.

### Other Features of the tfdbg CLI ###

In addition to the commands listed above, the tfdbg CLI provides the following additional features:

* To navigate through previous tfdbg commands, type in a few characters followed by the Up or Down arrow keys. tfdbg will show you the history of commands that started with those characters.
 * To navigate through the history of screen outputs, do either of the following:
 
 * Use the `prev` and `next` commands.
 * Click underlined `<--` and `-->` links near the top left corner of the screen.
 * Tab completion of commands and some command arguments.
 * To redirect the screen output to a file instead of the screen, end the command with bash-style redirection. For example, the following command redirects the output of the pt command to the `/tmp/xent_value_slices.txt`file:

> `tfdbg> pt cross_entropy/Log:0[:, 0:10] > /tmp/xent_value_slices.txt`

### Finding `nan`s and `inf`s ###

In this first `Session.run()` call, there happen to be no problematic numerical values. You can move on to the next run by using the command `run` or its shorthand `r`.

> TIP: If you enter `run` or `r` repeatedly, you will be able to move through the `Session.run()` calls in a sequential manner.
> 
> You can also use the `-t` flag to move ahead a number of `Session.run()` calls at a time, for example: 
> 
> `tfdbg> run -t 10`

Instead of entering `run` repeatedly and manually searching for `nan`s and `inf`s in the run-end UI after every `Session.run()` call (for example, by using the `pt` command shown in the table above) , you can use the following command to let the debugger repeatedly execute `Session.run()` calls without stopping at the run-start or run-end prompt, until the first `nan` or `inf` value shows up in the graph. This is analogous to *conditional breakpoints* in some procedural-language debuggers:

`tfdbg> run -f has_inf_or_nan`

> NOTE: The preceding command works properly because a tensor filter called `has_inf_or_nan` has been registered for you when the wrapped session is created. This filter detects `nan`s and `inf`s (as explained previously). If you have registered any other filters, you can use "run -f" to have tfdbg run until any tensor triggers that filter (cause the filter to return True). 
> 
> `def my_filter_callable(datum, tensor):   # A filter that detects zero-valued scalars.   return len(tensor.shape) == 0 and tensor == 0.0 sess.add_tensor_filter('my_filter', my_filter_callable)`
> 
> Then at the tfdbg run-start prompt run until your filter is triggered: 
> 
> `tfdbg> run -f my_filter`

See [this API document][API documentation] for more information on the expected signature and return value of the predicate `Callable` used with `add_tensor_filter()`.

![70 2][]As the screen display indicates on the first line, the `has_inf_or_nan` filter is first triggered during the fourth `Session.run()` call: an [Adam optimizer][] forward-backward training pass on the graph. In this run, 36 (out of the total 95) intermediate tensors contain `nan` or `inf` values. These tensors are listed in chronological order, with their timestamps displayed on the left. At the top of the list, you can see the first tensor in which the bad numerical values first surfaced: `cross_entropy/Log:0`.

To view the value of the tensor, click the underlined tensor name `cross_entropy/Log:0` or enter the equivalent command:

`tfdbg> pt cross_entropy/Log:0`

Scroll down a little and you will notice some scattered `inf` values. If the instances of `inf` and `nan` are difficult to spot by eye, you can use the following command to perform a regex search and highlight the output:

`tfdbg> /inf`

Or, alternatively:

`tfdbg> /(inf|nan)`

You can also use the `-s` or `--numeric_summary` command to get a quick summary of the types of numeric values in the tensor:

`tfdbg> pt -s cross_entropy/Log:0`

From the summary, you can see that several of the 1000 elements of the `cross_entropy/Log:0` tensor are `-inf`s (negative infinities).

Why did these infinities appear? To further debug, display more information about the node `cross_entropy/Log` by clicking the underlined `node_info` menu item on the top or entering the equivalent node\_info (`ni`) command:

`tfdbg> ni cross_entropy/Log`

![70 3][]

You can see that this node has the op type `Log` and that its input is the node `Softmax`. Run the following command to take a closer look at the input tensor:

`tfdbg> pt Softmax:0`

Examine the values in the input tensor, searching for zeros:

`tfdbg> /0\.000`

Indeed, there are zeros. Now it is clear that the origin of the bad numerical values is the node `cross_entropy/Log`taking logs of zeros. To find out the culprit line in the Python source code, use the `-t` flag of the `ni` command to show the traceback of the node's construction:

`tfdbg> ni -t cross_entropy/Log`

If you click "node\_info" at the top of the screen, tfdbg automatically shows the traceback of the node's construction.

From the traceback, you can see that the op is constructed at the following line: [`debug_mnist.py`][debug_mnist.py]:

`diff = y_ * tf.log(y)`

**tfdbg** has a feature that makes it easy to trace Tensors and ops back to lines in Python source files. It can annotate lines of a Python file with the ops or Tensors created by them. To use this feature, simply click the underlined line numbers in the stack trace output of the `ni -t <op_name>` commands, or use the `ps` (or `print_source`) command such as: `ps /path/to/source.py`. For example, the following screenshot shows the output of a `ps` command.

![70 4][]

### Fixing the problem ###

To fix the problem, edit `debug_mnist.py`, changing the original line:

`diff = -(y_ * tf.log(y))`

to the built-in, numerically-stable implementation of softmax cross-entropy:

`diff = tf.losses.softmax_cross_entropy(labels=y_, logits=logits)`

Rerun with the `--debug` flag as follows:

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

At the `tfdbg>` prompt, enter the following command:

run -f has_inf_or_nan`

Confirm that no tensors are flagged as containing `nan` or `inf` values, and accuracy now continues to rise rather than getting stuck. Success!

## Debugging TensorFlow Estimators ##

This section explains how to debug TensorFlow programs that use the `Estimator` APIs. Part of the convenience provided by these APIs is that they manage `Session`s internally. This makes the `LocalCLIDebugWrapperSession`described in the preceding sections inapplicable. Fortunately, you can still debug them by using special `hook`s provided by [`tfdbg`][tfdbg].

[`tfdbg`][tfdbg] can debug the [`train()`][train], [`evaluate()`][evaluate] and [`predict()`][predict] methods of tf-learn `Estimator`s. To debug `Estimator.train()`, create a `LocalCLIDebugHook` and supply it in the `hooks` argument. For example:

``# First, let your BUILD target depend on "//tensorflow/python/debug:debug_py" # (You don't need to worry about the BUILD dependency if you are using a pip #  install of open-source TensorFlow.) from tensorflow.python import debug as tf_debug # Create a LocalCLIDebugHook and use it as a monitor when calling fit(). hooks = [tf_debug.LocalCLIDebugHook()] # To debug `train`: classifier.train(input_fn,                  steps=1000,                  hooks=hooks)``

Similarly, to debug `Estimator.evaluate()` and `Estimator.predict()`, assign hooks to the `hooks` parameter, as in the following example:

``# To debug `evaluate`: accuracy_score = classifier.evaluate(eval_input_fn,                                      hooks=hooks)["accuracy"] # To debug `predict`: predict_results = classifier.predict(predict_input_fn, hooks=hooks)``

[debug\_tflearn\_iris.py][debug_tflearn_iris.py], contains a full example of how to use the tfdbg with `Estimator`s. To run this example, do:

`python -m tensorflow.python.debug.examples.debug_tflearn_iris --debug`

The `LocalCLIDebugHook` also allows you to configure a `watch_fn` that can be used to flexibly specify what `Tensor`s to watch on different `Session.run()` calls, as a function of the `fetches` and `feed_dict` and other states. See [this API doc][] for more details.

## Debugging Keras Models with TFDBG ##

To use TFDBG with [tf.keras][], let the Keras backend use a TFDBG-wrapped Session object. For example, to use the CLI wrapper:

``import tensorflow as tf from tensorflow.python import debug as tf_debug tf.keras.backend.set_session(tf_debug.LocalCLIDebugWrapperSession(tf.Session())) # Define your keras model, called "model". # Calls to `fit()`, 'evaluate()` and `predict()` methods will break into the # TFDBG CLI. model.fit(...) model.evaluate(...) model.predict(...)``

With minor modification, the preceding code example also works for the [non-TensorFlow version of Keras][] running against a TensorFlow backend. You just need to replace [`tf.keras.backend`][tf.keras.backend] with `keras.backend`.

## Debugging tf-slim with TFDBG ##

TFDBG supports debugging of training and evaluation with [tf-slim][]. As detailed below, training and evaluation require slightly different debugging workflows.

### Debugging training in tf-slim ###

To debug the training process, provide `LocalCLIDebugWrapperSession` to the `session_wrapper` argument of `slim.learning.train()`. For example:

`import tensorflow as tf from tensorflow.python import debug as tf_debug # ... Code that creates the graph and the train_op ... tf.contrib.slim.learning.train(     train_op,     logdir,     number_of_steps=10,     session_wrapper=tf_debug.LocalCLIDebugWrapperSession)`

### Debugging evaluation in tf-slim ###

To debug the evaluation process, provide `LocalCLIDebugHook` to the `hooks` argument of `slim.evaluation.evaluate_once()`. For example:

`import tensorflow as tf from tensorflow.python import debug as tf_debug # ... Code that creates the graph and the eval and final ops ... tf.contrib.slim.evaluation.evaluate_once(     '',     checkpoint_path,     logdir,     eval_op=my_eval_op,     final_op=my_value_op,     hooks=[tf_debug.LocalCLIDebugHook()])`

## Offline Debugging of Remotely-Running Sessions ##

Often, your model is running on a remote machine or a process that you don't have terminal access to. To perform model debugging in such cases, you can use the `offline_analyzer` binary of [`tfdbg`][tfdbg] (described below). It operates on dumped data directories. This can be done to both the lower-level `Session` API and the higher-level `Estimator` API.

### Debugging Remote tf.Sessions ###

If you interact directly with the [`tf.Session`][Session] API in `python`, you can configure the `RunOptions` proto that you call your `Session.run()` method with, by using the method [`tfdbg.watch_graph`][tfdbg.watch_graph]. This will cause the intermediate tensors and runtime graphs to be dumped to a shared storage location of your choice when the `Session.run()` call occurs (at the cost of slower performance). For example:

`from tensorflow.python import debug as tf_debug # ... Code where your session and graph are set up... run_options = tf.RunOptions() tf_debug.watch_graph(       run_options,       session.graph,       debug_urls=["file:///shared/storage/location/tfdbg_dumps_1"]) # Be sure to specify different directories for different run() calls. session.run(fetches, feed_dict=feeds, options=run_options)`

Later, in an environment that you have terminal access to (for example, a local computer that can access the shared storage location specified in the code above), you can load and inspect the data in the dump directory on the shared storage by using the `offline_analyzer` binary of [`tfdbg`][tfdbg]. For example:

`python -m tensorflow.python.debug.cli.offline_analyzer \     --dump_dir=/shared/storage/location/tfdbg_dumps_1`

The `Session` wrapper `DumpingDebugWrapperSession` offers an easier and more flexible way to generate file-system dumps that can be analyzed offline. To use it, simply wrap your session in a `tf_debug.DumpingDebugWrapperSession`. For example:

`# Let your BUILD target depend on "//tensorflow/python/debug:debug_py # (You don't need to worry about the BUILD dependency if you are using a pip #  install of open-source TensorFlow.) from tensorflow.python import debug as tf_debug sess = tf_debug.DumpingDebugWrapperSession(     sess, "/shared/storage/location/tfdbg_dumps_1/", watch_fn=my_watch_fn)`

The `watch_fn` argument accepts a `Callable` that allows you to configure what `tensor`s to watch on different `Session.run()` calls, as a function of the `fetches` and `feed_dict` to the `run()` call and other states.

### C++ and other languages ###

If your model code is written in C++ or other languages, you can also modify the `debug_options` field of `RunOptions`to generate debug dumps that can be inspected offline. See [the proto definition][] for more details.

### Debugging Remotely-Running Estimators ###

If your remote TensorFlow server runs `Estimator`s, you can use the non-interactive `DumpingDebugHook`. For example:

`# Let your BUILD target depend on "//tensorflow/python/debug:debug_py # (You don't need to worry about the BUILD dependency if you are using a pip #  install of open-source TensorFlow.) from tensorflow.python import debug as tf_debug hooks = [tf_debug.DumpingDebugHook("/shared/storage/location/tfdbg_dumps_1")]`

Then this `hook` can be used in the same way as the `LocalCLIDebugHook` examples described earlier in this document. As the training, evalution or prediction happens with `Estimator`, tfdbg creates directories having the following name pattern: `/shared/storage/location/tfdbg_dumps_1/run_<epoch_timestamp_microsec>_<uuid>`. Each directory corresponds to a `Session.run()` call that underlies the `fit()` or `evaluate()` call. You can load these directories and inspect them in a command-line interface in an offline manner using the `offline_analyzer` offered by tfdbg. For example:

`python -m tensorflow.python.debug.cli.offline_analyzer \     --dump_dir="/shared/storage/location/tfdbg_dumps_1/run_<epoch_timestamp_microsec>_<uuid>"`

## Frequently Asked Questions ##

**Q**: *Do the timestamps on the left side of the `lt` output reflect actual performance in a non-debugging session?*

**A**: No. The debugger inserts additional special-purpose debug nodes to the graph to record the values of intermediate tensors. These nodes slow down the graph execution. If you are interested in profiling your model, check out

1.  The profiling mode of tfdbg: `tfdbg> run -p`.
2.  [tfprof][] and other profiling tools for TensorFlow.

**Q**: *How do I link tfdbg against my `Session` in Bazel? Why do I see an error such as "ImportError: cannot import name debug"?*

**A**: In your BUILD rule, declare dependencies: `"//tensorflow:tensorflow_py"` and `"//tensorflow/python/debug:debug_py"`. The first is the dependency that you include to use TensorFlow even without debugger support; the second enables the debugger. Then, In your Python file, add:

`from tensorflow.python import debug as tf_debug # Then wrap your TensorFlow Session with the local-CLI wrapper. sess = tf_debug.LocalCLIDebugWrapperSession(sess)`

**Q**: *Does tfdbg help debug runtime errors such as shape mismatches?*

**A**: Yes. tfdbg intercepts errors generated by ops during runtime and presents the errors with some debug instructions to the user in the CLI. See examples:

`# Debugging shape mismatch during matrix multiplication. python -m tensorflow.python.debug.examples.debug_errors \     --error shape_mismatch --debug # Debugging uninitialized variable. python -m tensorflow.python.debug.examples.debug_errors \     --error uninitialized_variable --debug`

**Q**: *How can I let my tfdbg-wrapped Sessions or Hooks run the debug mode only from the main thread?*

**A**: This is a common use case, in which the `Session` object is used from multiple threads concurrently. Typically, the child threads take care of background tasks such as running enqueue operations. Often, you want to debug only the main thread (or less frequently, only one of the child threads). You can use the `thread_name_filter` keyword argument of `LocalCLIDebugWrapperSession` to achieve this type of thread-selective debugging. For example, to debug from the main thread only, construct a wrapped `Session` as follows:

`sess = tf_debug.LocalCLIDebugWrapperSession(sess, thread_name_filter="MainThread$")`

The above example relies on the fact that main threads in Python have the default name `MainThread`.

**Q**: *The model I am debugging is very large. The data dumped by tfdbg fills up the free space of my disk. What can I do?*

**A**: You might encounter this problem in any of the following situations:

*  models with many intermediate tensors
 *  very large intermediate tensors
 *  many [`tf.while_loop`][tf.while_loop] iterations

There are three possible workarounds or solutions:

*  The constructors of `LocalCLIDebugWrapperSession` and `LocalCLIDebugHook` provide a keyword argument, `dump_root`, to specify the path to which tfdbg dumps the debug data. You can use it to let tfdbg dump the debug data on a disk with larger free space. For example:

`# For LocalCLIDebugWrapperSession sess = tf_debug.LocalCLIDebugWrapperSession(dump_root="/with/lots/of/space") # For LocalCLIDebugHook hooks = [tf_debug.LocalCLIDebugHook(dump_root="/with/lots/of/space")]`

Make sure that the directory pointed to by dump\_root is empty or nonexistent. [`tfdbg`][tfdbg] cleans up the dump directories before exiting.

*  Reduce the batch size used during the runs.
 *  Use the filtering options of tfdbg's `run` command to watch only specific nodes in the graph. For example:
    
    `tfdbg> run --node_name_filter .*hidden.* tfdbg> run --op_type_filter Variable.* tfdbg> run --tensor_dtype_filter int.*`
    
    The first command above watches only nodes whose name match the regular-expression pattern `.*hidden.*`. The second command watches only operations whose name match the pattern `Variable.*`. The third one watches only the tensors whose dtype match the pattern `int.*` (e.g., `int32`).

**Q**: *Why can't I select text in the tfdbg CLI?*

**A**: This is because the tfdbg CLI enables mouse events in the terminal by default. This [mouse-mask][] mode overrides default terminal interactions, including text selection. You can re-enable text selection by using the command `mouse off` or `m off`.

**Q**: *Why does the tfdbg CLI show no dumped tensors when I debug code like the following?*

`a = tf.ones([10], name="a") b = tf.add(a, a, name="b") sess = tf.Session() sess = tf_debug.LocalCLIDebugWrapperSession(sess) sess.run(b)`

**A**: The reason why you see no data dumped is because every node in the executed TensorFlow graph is constant-folded by the TensorFlow runtime. In this exapmle, `a` is a constant tensor; therefore, the fetched tensor `b` is effectively also a constant tensor. TensorFlow's graph optimization folds the graph that contains `a` and `b` into a single node to speed up future runs of the graph, which is why [`tfdbg`][tfdbg] does not generate any intermediate tensor dumps. However, if `a` were a[`tf.Variable`][tf.Variable], as in the following example:

`import numpy as np a = tf.Variable(np.ones[10], name="a") b = tf.add(a, a, name="b") sess = tf.Session() sess.run(tf.global_variables_initializer()) sess = tf_debug.LocalCLIDebugWrapperSession(sess) sess.run(b)`

the constant-folding would not occur and [`tfdbg`][tfdbg] should show the intermediate tensor dumps.

**Q**: I am debugging a model that generates unwanted infinities or NaNs. But there are some nodes in my model that are known to generate infinities or NaNs in their output tensors even under completely normal conditions. How can I skip those nodes during my `run -f has_inf_or_nan` actions?

**A**: Use the `--filter_exclude_node_names` (`-fenn` for short) flag. For example, if you known you have a node with name matching the regular expression `.*Sqrt.*` that generates infinities or NaNs regardless of whether the model is behaving correctly, you can exclude the nodes from the infinity/NaN-finding runs with the command `run -f has_inf_or_nan -fenn .*Sqrt.*`.

**Q**: Is there a GUI for tfdbg?

**A**: Yes, the **TensorBoard Debugger Plugin** is the GUI of tfdbg. It offers features such as inspection of the computation graph, real-time visualization of tensor values, continuation to tensor and conditional breakpoints, and tying tensors to their graph-construction source code, all in the browser environment. To get started, please visit [its README][].

Except as otherwise noted, the content of this page is licensed under the [Creative Commons Attribution 3.0 License][], and code samples are licensed under the [Apache 2.0 License][]. For details, see our [Site Policies][]. Java is a registered trademark of Oracle and/or its affiliates.

Last updated August 8, 2018.

[https_www.tensorflow.org_guide_debugger]: https://www.tensorflow.org/guide/debugger
[tfdbg]: https://www.tensorflow.org/api_docs/python/tfdbg
[its README]: https://github.com/tensorflow/tensorboard/blob/master/tensorboard/plugins/debugger/README.md
[curses]: https://en.wikipedia.org/wiki/Curses_%28programming_library%29
[readline]: https://en.wikipedia.org/wiki/GNU_Readline
[here]: https://www.lfd.uci.edu/~gohlke/pythonlibs/#curses
[nan_s]: https://en.wikipedia.org/wiki/NaN
[inf_s]: https://en.wikipedia.org/wiki/Infinity
[Session]: https://www.tensorflow.org/api_docs/python/tf/Session
[tf.estimator]: https://www.tensorflow.org/api_docs/python/tf/estimator
[tf.keras]: https://www.tensorflow.org/api_docs/python/tf/keras
[tf.contrib.slim]: https://www.tensorflow.org/api_docs/python/tf/contrib/slim
[here 1]: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/debug/examples/debug_mnist.py
[tfdbg.has_inf_or_nan]: https://www.tensorflow.org/api_docs/python/tfdbg/has_inf_or_nan
[debug_data]: https://www.tensorflow.org/api_guides/python/tfdbg#Classes_for_debug_dump_data_and_directories
[API documentation]: https://www.tensorflow.org/api_docs/python/tfdbg/DebugDumpDir#find
[70]: /images/20220514/a7384107c29a4083abe3d02f5f92e013.png
[70 1]: /images/20220514/e4da48f9f8d346c281c76b7260901725.png
[70 2]: /images/20220514/87e80093b80d437abe48860681731ecd.png
[Adam optimizer]: https://www.tensorflow.org/api_docs/python/tf/train/AdamOptimizer
[70 3]: /images/20220514/d1b80f777a0747289a58d2600eaf5a04.png
[debug_mnist.py]: https://www.github.com/tensorflow/tensorflow/blob/r1.10/tensorflow/python/debug/examples/debug_mnist.py
[70 4]: /images/20220514/855ed013cf2e47c58f6f269ccb1bb3b2.png
[train]: https://www.tensorflow.org/api_docs/python/tf/estimator/Estimator#train
[evaluate]: https://www.tensorflow.org/api_docs/python/tf/estimator/Estimator#evaluate
[predict]: https://www.tensorflow.org/api_docs/python/tf/estimator/Estimator#predict
[debug_tflearn_iris.py]: https://www.github.com/tensorflow/tensorflow/blob/r1.10/tensorflow/python/debug/examples/debug_tflearn_iris.py
[this API doc]: https://www.tensorflow.org/api_docs/python/tfdbg/DumpingDebugWrapperSession#__init__
[non-TensorFlow version of Keras]: https://keras.io/
[tf.keras.backend]: https://www.tensorflow.org/api_docs/python/tf/keras/backend
[tf-slim]: https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/slim
[tfdbg.watch_graph]: https://www.tensorflow.org/api_docs/python/tfdbg/watch_graph
[the proto definition]: https://www.github.com/tensorflow/tensorflow/blob/r1.10/tensorflow/core/protobuf/debug.proto
[tfprof]: https://github.com/tensorflow/tensorflow/tree/master/tensorflow/core/profiler
[tf.while_loop]: https://www.tensorflow.org/api_docs/python/tf/while_loop
[mouse-mask]: https://linux.die.net/man/3/mousemask
[tf.Variable]: https://www.tensorflow.org/api_docs/python/tf/Variable
[Creative Commons Attribution 3.0 License]: https://creativecommons.org/licenses/by/3.0/
[Apache 2.0 License]: https://www.apache.org/licenses/LICENSE-2.0
[Site Policies]: https://developers.google.com/terms/site-policies