I’m available for freelance work. Let’s talk »

Debugging TensorFlow coverage

Sunday 27 October 2019

It started with a coverage.py issue: Coverage not working for TensorFlow Model call function. A line in the code is executing, but coverage.py marks it as unexecuted. How could that be?

TensorFlow was completely new to me. I knew it had some unusual execution semantics, but I didn’t know what it involved. What could it be doing that would interfere with coverage measurement? Once I had instructions to reproduce the issue, I could see that it was true: a line that clearly printed output during the test run was marked as unexecuted.

The code in question was in a file called routenet_model.py. It had a line like this:

print('******** in call ******************')

It was the only such line in the code, and sure enough, the test output showed that “**** in call ****” text, so the line was definitely running.

The first step was to see who was calling the product code. It seemed like something about a caller was getting in the way, since other code in that file was marked as executed. I added this to get a stack trace at that point:

import inspect
print("\n".join("%30s : %s:%d" % (t[3],t[1],t[2]) for t in inspect.stack()[::-1]))
print('******** in call ******************')

When I re-ran the test, I saw a long stack trace that ended like this (I’ve abbreviated some of the file paths):

              ...     ...
           compute : site-packages/tensorflow/python/ops/map_fn.py:257
          <lambda> : /private/tmp/bug856/demo-routenet/tests/utils/test_utils.py:31
          __call__ : site-packages/tensorflow/python/keras/engine/base_layer.py:634
           wrapper : site-packages/tensorflow/python/autograph/impl/api.py:146
    converted_call : site-packages/tensorflow/python/autograph/impl/api.py:453
          tf__call : /var/folders/j2/gr3cj3jn63s5q8g3bjvw57hm0000gp/T/tmps9vwjn47.py:10
    converted_call : site-packages/tensorflow/python/autograph/impl/api.py:349
 _call_unconverted : site-packages/tensorflow/python/autograph/impl/api.py:258
******** in call ******************

This stack shows the function name, the file path, and the line number in a compact way. It’s a useful enough debugging helper that I have it as a vim abbreviation.

Hmm, interesting: there’s a temporary Python file (tmps9vwjn47.py) in the call stack. That’s definitely unusual. The file is gone by the time the tests are done, so to get the contents, I grab the filename from the stack trace, and copy the contents elsewhere:

import inspect
print("\n".join("%30s : %s:%d" % (t[3],t[1],t[2]) for t in inspect.stack()[::-1]))
with open("/tmp/bug856/weird.py", "w") as fout:
    with open(inspect.stack()[2].filename) as fin:
        fout.write(fin.read())
print('******** in call ******************')

I named the copied file “weird.py” because a temporary Python file is weird any time, but this is where it gets really weird: weird.py is a 528-line Python file, but it doesn’t have the function indicated in the stack trace: there’s nothing named tf__call in it. The stack trace also indicates that line 10 is running, but line 10 is a comment:

 1# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
2#
3# Licensed under the Apache License, Version 2.0 (the "License");
4# you may not use this file except in compliance with the License.
5# You may obtain a copy of the License at
6#
7#     http://www.apache.org/licenses/LICENSE-2.0
8#
9# Unless required by applicable law or agreed to in writing, software
10# distributed under the License is distributed on an "AS IS" BASIS,
11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12# See the License for the specific language governing permissions and
13# limitations under the License.
14# ==============================================================================
15"""Control flow statements: loops, conditionals, etc."""
16
17from __future__ import absolute_import
18from __future__ import division
19from __future__ import print_function
20
21... etc ...

Something is truly weird here. To add to the confusion, I can print the entire FrameInfo from the stack trace, with:

print(repr(inspect.stack()[2]))

and it shows:

FrameInfo(
    frame=<frame at 0x7ff1f0e05080, file '/var/folders/j2/gr3cj3jn63s5q8g3bjvw57hm0000gp/T/tmp1ii_1_na.py', line 14, code tf__call>,
    filename='/var/folders/j2/gr3cj3jn63s5q8g3bjvw57hm0000gp/T/tmp1ii_1_na.py',
    lineno=14,
    function='tf__call',
    code_context=["      print(ag__.converted_call(repr, None, ag__.ConversionOptions(recursive=True, force_conversion=False, optional_features=(), internal_convert_user_code=True), (ag__.converted_call('stack', inspect, ag__.ConversionOptions(recursive=True, force_conversion=False, optional_features=(), internal_convert_user_code=True), (), None)[2],), None))\n"],
    index=0
)

The code_context attribute there shows a plausible line of code, but it doesn’t correspond to the code in the file at all. This is a twist on a long-standing gotcha with Python stack traces. When Python code is running, it has filenames and line numbers in the frames on the call stack, but it doesn’t keep the source of the code it runs. To populate a stack trace with lines of code, it reads the file on disk. The classic problem with this is that the file on disk may have changed since the code started running. So the lines of source in the stack trace might be wrong because they are newer than the actual code that is running.

But that’s not what we’re seeing here. Now the line of code in the stack trace doesn’t match the file on disk at all. It seems to correspond to what is running, but not what is on disk. The reason is that Python uses a module called linecache to read the line of source. As the name implies, linecache caches file contents, so that reading many different lines from a file won’t try to open the same file many times.

What must have happened here is that the file had one program in it, and then was read (and cached) by linecache for some reason. That first program is what is running. Then the file was re-written with a second program. Linecache checks the modification time to invalidate the cache, but if the file was rewritten quickly enough to not have a different modification time, then the stale cache would be used. This is why the stack trace has the correct line of code, even though the file on disk doesn’t.

A quick look in the __pycache__ directory in the tmp directory shows a .pyc file, and if I dump it with show_pyc.py, I can see that it has the code I’m interested in. But rather than try to read disassembled bytecode, I can get the source from the stale copy in linecache!

import inspect
print("\n".join("%30s : %s:%d" % (t[3],t[1],t[2]) for t in inspect.stack()[::-1]))
with open("/tmp/bug856/better.py", "w") as fout:
    import linecache
    fout.write("".join(linecache.getlines(inspect.stack()[2].filename)))
print('******** in call ******************')

When I run this, I get a file better.py that makes clear why coverage.py claimed the original line wasn’t executed. Here’s the start of better.py:

 1def create_converted_entity_factory():
2
3  def create_converted_entity(ag__, ag_source_map__, ag_module__):
4
5    def tf__call(self, inputs, training=None):
6      do_return = False
7      retval_ = ag__.UndefinedReturnValue()
8      import inspect, sys
9      print(ag__.converted_call('join', '\n', ag__.ConversionOptions(recursive=True, force_conversion=False, optional_features=(), internal_convert_user_code=True), (('%30s : %s:%d' % (t[3], t[1], t[2]) for t in ag__.converted_call('getouterframes', inspect, ag__.ConversionOptions(recursive=True, force_conversion=False, optional_features=(), internal_convert_user_code=True), (ag__.converted_call('_getframe', sys, ag__.ConversionOptions(recursive=True, force_conversion=False, optional_features=(), internal_convert_user_code=True), (), None),), None)[::-1]),), None))
10      with open('/tmp/bug856/better.py', 'w') as fout:
11        import linecache
12        ag__.converted_call('write', fout, ag__.ConversionOptions(recursive=True, force_conversion=False, optional_features=(), internal_convert_user_code=True), (ag__.converted_call('join', '', ag__.ConversionOptions(recursive=True, force_conversion=False, optional_features=(), internal_convert_user_code=True), (ag__.converted_call('getlines', linecache, ag__.ConversionOptions(recursive=True, force_conversion=False, optional_features=(), internal_convert_user_code=True), (ag__.converted_call('stack', inspect, ag__.ConversionOptions(recursive=True, force_conversion=False, optional_features=(), internal_convert_user_code=True), (), None)[2].filename,), None),), None),), None)
13      print('******** in call ******************')
14
15    ... lines omitted ...
16
17    return tf__call
18  return create_converted_entity

This is the code from our original routenet_model.py (including all the debugging code that I put in there), translated into some kind of annotated form. The reason coverage.py said the product code wasn’t run is because it wasn’t run! A copy of the code was run.

Now I realize something about inspect.stack(): the first frame it shows is your caller. If I had used a stack trace that showed the current frame first, it would have shown that my debugging code was not in the file I thought it was.

It turns out that inspect.stack() is a one-line helper using other things:

def stack(context=1):
    """Return a list of records for the stack above the caller's frame."""
    return getouterframes(sys._getframe(1), context)

Changing my stack trace one-liner to use getoutframes(sys._getframe()) is better, but is still confusing in this case because TensorFlow rewrites function calls, including sys._getframe, so the resulting stack trace ends with:

__call__ : site-packages/tensorflow/python/keras/engine/base_layer.py:634
       wrapper : site-packages/tensorflow/python/autograph/impl/api.py:146
converted_call : site-packages/tensorflow/python/autograph/impl/api.py:453
      tf__call : /var/folders/j2/gr3cj3jn63s5q8g3bjvw57hm0000gp/T/tmpcwhc1y2a.py:10
converted_call : site-packages/tensorflow/python/autograph/impl/api.py:321

Even now, I can’t quite wrap my head around why it comes out that way.

The next step is to decide what to do about this. The converted code has a parameter called ag_source_map__, which is a map from converted code back to source code. This could be used to get the coverage right, perhaps in a plugin, but I need to hear from TensorFlow people to see what would be the best approach. I’ve written a TensorFlow issue to start the conversation.

Comments

Add a comment:

Ignore this:
Leave this empty:
Name is required. Either email or web are required. Email won't be displayed and I won't spam you. Your web site won't be indexed by search engines.
Don't put anything here:
Leave this empty:
Comment text is Markdown.