[TOC]

TensorFlow debugger (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.

NOTE: The system requirements of tfdbg on supported external platforms include the following. On Mac OS X, the ncurses library is required. It can be installed with brew install homebrew/dupes/ncurses. On Windows, pyreadline is required. If you use Anaconda3, you can install it with a command such as "C:\Program Files\Anaconda3\Scripts\pip.exe" install pyreadline.

This tutorial demonstrates how to use the tfdbg command-line interface (CLI) to debug the appearance of nans and infs, a frequently-encountered type of bug in TensorFlow model development. The following example is for users who use the low-level Session API of TensorFlow. A later section of this document describes how to use tfdbg with a higher-level API, namely tf-learn Estimators and Experiments. To observe such an issue, run the following command without the debugger (the source code can be found here):

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 infs and nans, 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.

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, 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)
sess.add_tensor_filter("has_inf_or_nan", tf_debug.has_inf_or_nan)

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}, 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 nans and infs is a common enough use case that we ship it with the @{$python/tfdbg#Classes_for_debug_dump_data_and_directories$debug_data} module.

Note: You can also write your own custom filters. See the @{tfdbg.DebugDumpDir.find$API documentation} of DebugDumpDir.find() for additional information.

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.

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:

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

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

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.

In addition to the commands listed above, the tfdbg CLI provides the following addditional 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

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 nans and infs 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 we have registered a filter for nans and infs called has_inf_or_nan (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).

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 for more information on the expected signature and return value of the predicate Callable used with add_tensor_filter().

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 -infs (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

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

tfdbg> pt softmax/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:

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.

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.nn.softmax_cross_entropy_with_logits(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!

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

Currently, tfdbg can debug the @{tf.contrib.learn.BaseEstimator.fit$fit()} @{tf.contrib.learn.BaseEstimator.evaluate$evaluate()} methods of tf-learn Estimators. To debug Estimator.fit(), create a LocalCLIDebugHook and supply it in the monitors 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()]

classifier.fit(x=training_set.data,
               y=training_set.target,
               steps=1000,
               monitors=hooks)

To debug Estimator.evaluate(), assign hooks to the hooks parameter, as in the following example:

accuracy_score = classifier.evaluate(x=test_set.data,
                                     y=test_set.target,
                                     hooks=hooks)["accuracy"]

debug_tflearn_iris.py, based on {$tflearn$tf-learn's iris tutorial}, contains a full example of how to use the tfdbg with Estimators. To run this example, do:

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

Experiment is a construct in tf.contrib.learn at a higher level than Estimator. It provides a single interface for training and evaluating a model. To debug the train() and evaluate() calls to an Experiment object, you can use the keyword arguments train_monitors and eval_hooks, respectively, when calling its constructor. 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

hooks = [tf_debug.LocalCLIDebugHook()]

ex = experiment.Experiment(classifier,
                           train_input_fn=iris_input_fn,
                           eval_input_fn=iris_input_fn,
                           train_steps=FLAGS.train_steps,
                           eval_delay_secs=0,
                           eval_steps=1,
                           train_monitors=hooks,
                           eval_hooks=hooks)

ex.train()
accuracy_score = ex.evaluate()["accuracy"]

To build and run the debug_tflearn_iris example in the Experiment mode, do:

python -m tensorflow.python.debug.examples.debug_tflearn_iris \
    --use_experiment --debug

The LocalCLIDebugHook also allows you to configure a watch_fn that can be used to flexibly specify what Tensors to watch on different Session.run() calls, as a function of the fetches and feed_dict and other states. See @{tfdbg.DumpingDebugWrapperSession.init$this API doc} for more details.

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

import tensorflow as tf
from keras import backend as keras_backend
from tensorflow.python import debug as tf_debug

keras_backend.set_session(tf_debug.LocalCLIDebugWrapperSession(tf.Session()))

# Define your keras model, called "model".
model.fit(...)  # This will break into the TFDBG CLI.

TFDBG currently supports only training with 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)

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 (described below). It operates on dumped data directories. This can be done to both the lower-level Session API and the higher-level Estimator and Experiment APIs.

If you interact directly with the tf.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}. 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. 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 tensors to watch on different Session.run() calls, as a function of the fetches and feed_dict to the run() call and other states.

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.

If your remote TensorFlow server runs Estimators, 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 and/or evalution of Estimator or Experiment happens, 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>"

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} 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 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 does not generate any intermediate tensor dumps. However, if a were a @{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 should show the intermediate tensor dumps.