[go: nahoru, domu]
Skip to content

Commit

Permalink
[tfdbg2] Update tensorflow/python/debug/README.md
Browse files Browse the repository at this point in the history 
In this CL, we
- remove obsolete information about tfdbg v1
- add up-to-date description of tfdbg v2

PiperOrigin-RevId: 366923367
Change-Id: I80aeb0f8ee2255fabeb83ffb806ebbfd300c5bf4
  • Loading branch information
@caisq @tensorflower-gardener
caisq authored and tensorflower-gardener committed Apr 6, 2021
1 parent e04290d commit 04480a4
Showing 1 changed file with 44 additions and 47 deletions.
91 changes: 44 additions & 47 deletions tensorflow/python/debug/README.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -3,53 +3,50 @@
[TOC]

TensorFlow Debugger (TFDBG) is a specialized debugger for TensorFlow's computation
graphs. It provides access to internal graph structures and tensor values at
TensorFlow runtime.

<!-- TODO(cais): Add release notes starting from 1.3. -->

## Why TFDBG?

In TensorFlow's current
[computation-graph framework](https://www.tensorflow.org/get_started/get_started#the_computational_graph),
almost all actual computation after graph construction happens in a single
Python function, namely
[tf.Session.run](https://www.tensorflow.org/api_docs/python/tf/Session#run).
Basic Python debugging tools such as [pdb](https://docs.python.org/2/library/pdb.html)
cannot be used to debug `Session.run`, due to the fact that TensorFlow's graph
execution happens in the underlying C++ layer. C++ debugging tools such as
[gdb](https://www.gnu.org/software/gdb/) are not ideal either, because of their
inability to recognize and organize the stack frames and variables in a way
relevant to TensorFlow's operations, tensors and other graph constructs.

TFDBG addresses these limitations. Among the features provided by TFDBG, the
following ones are designed to facilitate runtime debugging of TensorFlow
models:

* Easy access through session wrappers
* Easy integration with common high-level APIs, such as
[TensorFlow Estimators](https://www.tensorflow.org/guide/estimators) and
[Keras](https://keras.io/)
* Inspection of runtime tensor values and node connections
* Conditional breaking after runs that generate tensors satisfying given
predicates, which makes common debugging tasks such as tracing the origin
of infinities and [NaNs](https://en.wikipedia.org/wiki/NaN) easier
* Association of nodes and tensors in graphs with Python source lines
* Profiling of models at the level of graph nodes and Python source lines.
(Omitted internal-only feature)
* A [gRPC](https://grpc.io/)-based remote debugging protocol, which allows us to
build a browser-based graphical user interface (GUI) for TFDBG: the
[TensorBoard Debugger Plugin](https://github.com/tensorflow/tensorboard/blob/master/tensorboard/plugins/debugger/README.md).
runtime. TFDBG in TensorFlow 2.x provides access to:

## How to use TFDBG?

* For a walkthrough of TFDBG command-line interface, see https://www.tensorflow.org/guide/debugger.
* For information on the web GUI of TFDBG (TensorBoard Debugger Plugin), see
[this README](https://github.com/tensorflow/tensorboard/blob/master/tensorboard/plugins/debugger/README.md).
* For programmatic use of the API of TFDBG, see https://www.tensorflow.org/api_docs/python/tfdbg.
- Tensor values during [eager](https://www.tensorflow.org/guide/eager) and
[graph](https://www.tensorflow.org/api_docs/python/tf/Graph) execution.
- Structure of computation graphs
- Source code and stack traces associated with these execution and
graph-execution events.

## How to use TFDBG?

## Related Publications

* Cai, S., Breck E., Nielsen E., Salib M., Sculley D. (2016) TensorFlow Debugger:
Debugging Dataflow Graphs for Machine Learning. https://research.google.com/pubs/pub45789.html
TFDBG in TensorFlow 2.x consists of a Python API that enables dumping debug data
to the file system (namely `tf.debugging.experimental.enable_dump_debug_info()`)
and a TensorBoard-based GUI that provides an interactive visualization of the
debug data (i.e., *TensorBoard Debugger V2 Plugin*).

`enable_dump_debug_info()` offers a number of levels of tensor-value
instrumentation varying in the amount of information dumped and the incurred
performance overhead.

See the API documentation of
[`tf.debugging.experimental.enable_dump_debug_info()`](https://www.tensorflow.org/api_docs/python/tf/debugging/experimental/enable_dump_debug_info)

For a detailed walkthrough of the GUI TensorBoard Debugger V2 Plugin, see
[Debugging Numerical Issues in TensorFlow Programs Using TensorBoard Debugger
V2](https://www.tensorflow.org/tensorboard/debugger_v2).

## Known issues and limitations

1. Using `tf.debugging.experimental.enable_dumpp_debug_info()` leads to
performance penalty on your TensorFlow program. The amount of slowdown
varied depending on whether you are using TensorFlow on CPU, GPUs, or TPUs.
The performance penalty is the highest on TPUs, followed by GPUs, and lowest
on CPU.
2. `tf.debugging.experimental.enable_dump_debug_info()` is currently
incompatible with
[model saving/loading and checkpointing](https://www.tensorflow.org/tutorials/keras/save_and_load)

## Legacy API for TensorFlow 1.x

TensorFlow 1.x's execution paradigm is different from that of TensorFlow v2; it
is based on the deprecated
[`tf.Session`](https://www.tensorflow.org/api_docs/python/tf/compat/v1/Session)
If you are using TensorFlow 1.x, you can use the deprecated
`tf_debug.LocalCLIDebugWrapperSession` wrapper for `tf.Session`
to inspect tensor values and other types of debug information in a
terminal-based command-line interface. For details, see
[this blog post](https://developers.googleblog.com/2017/02/debug-tensorflow-models-with-tfdbg.html).

0 comments on commit 04480a4

Please sign in to comment.