tf.function raises OSError "could not get source code" when run in a Python shell

[ageron]

System information

• Have I written custom code (as opposed to using a stock example script provided in TensorFlow):
  Yes
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04):
  Mac OS X 10.13.6
• Mobile device (e.g. iPhone 8, Pixel 2, Samsung Galaxy) if the issue happens on mobile device:
  N/A
• TensorFlow installed from (source or binary):
  binary
• TensorFlow version (use command below):
  VERSION="1.13.0-dev20181226" (this is the TF 2.0-preview)
  GIT_VERSION="b'v1.12.0-5133-gc343196842'"
• Python version:
  3.6.6
• Bazel version (if compiling from source):
  N/A
• GCC/Compiler version (if compiling from source):
  N/A
• CUDA/cuDNN version:
  N/A
• GPU model and memory:
  N/A

Describe the current behavior
tf.function raises an OSError exception with the message "could not get source code" when I run it in a Python shell. The error does not occur in Jupyter or ipython, and it does not occur when I revert back to TF 1.12.0.

Describe the expected behavior
I expect no error.

Code to reproduce the issue

Run this in a Python shell:

import tensorflow as tf

@tf.function
def foo():
    return 42

foo()

Other info / logs

Here is the output of the above commands when I run them in a Python shell:

(tf2) $ python
Python 3.6.6 (default, Jun 28 2018, 05:43:53)
[GCC 4.2.1 Compatible Apple LLVM 9.1.0 (clang-902.0.39.2)] on darwin
Type "help", "copyright", "credits" or "license" for more information.
>>> import tensorflow as tf
>>> @tf.function
... def foo():
...     return 42
...
>>> foo()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/Users/ageron/.virtualenvs/tf2/lib/python3.6/site-packages/tensorflow/python/eager/def_function.py", line 336, in __call__
    self._initialize(args, kwds)
  File "/Users/ageron/.virtualenvs/tf2/lib/python3.6/site-packages/tensorflow/python/eager/def_function.py", line 309, in _initialize
    *args, **kwds))
  File "/Users/ageron/.virtualenvs/tf2/lib/python3.6/site-packages/tensorflow/python/eager/function.py", line 1024, in _get_concrete_function_internal_garbage_collected
    graph_function, _, _ = self._maybe_define_function(args, kwargs)
  File "/Users/ageron/.virtualenvs/tf2/lib/python3.6/site-packages/tensorflow/python/eager/function.py", line 1247, in _maybe_define_function
    arg_names=arg_names),
  File "/Users/ageron/.virtualenvs/tf2/lib/python3.6/site-packages/tensorflow/python/framework/func_graph.py", line 456, in func_graph_from_py_func
    func_outputs = python_func(*func_args, **func_kwargs)
  File "/Users/ageron/.virtualenvs/tf2/lib/python3.6/site-packages/tensorflow/python/eager/def_function.py", line 261, in wrapped_fn
    return weak_wrapped_fn().__wrapped__(*args, **kwds)
  File "/Users/ageron/.virtualenvs/tf2/lib/python3.6/site-packages/tensorflow/python/framework/func_graph.py", line 449, in wrapper
    ), *args, **kwargs)
  File "/Users/ageron/.virtualenvs/tf2/lib/python3.6/site-packages/tensorflow/python/autograph/impl/api.py", line 293, in converted_call
    experimental_partial_types=partial_types)
  File "/Users/ageron/.virtualenvs/tf2/lib/python3.6/site-packages/tensorflow/python/autograph/impl/api.py", line 415, in to_graph
    arg_values, arg_types)
  File "/Users/ageron/.virtualenvs/tf2/lib/python3.6/site-packages/tensorflow/python/autograph/impl/conversion.py", line 175, in entity_to_graph
    node, name, ns = function_to_graph(o, program_ctx, arg_values, arg_types)
  File "/Users/ageron/.virtualenvs/tf2/lib/python3.6/site-packages/tensorflow/python/autograph/impl/conversion.py", line 338, in function_to_graph
    node, source = parser.parse_entity(f)
  File "/Users/ageron/.virtualenvs/tf2/lib/python3.6/site-packages/tensorflow/python/autograph/pyct/parser.py", line 34, in parse_entity
    source = tf_inspect.getsource(entity)
  File "/Users/ageron/.virtualenvs/tf2/lib/python3.6/site-packages/tensorflow/python/util/tf_inspect.py", line 327, in getsource
    return _inspect.getsource(tf_decorator.unwrap(object)[1])
  File "/opt/local/Library/Frameworks/Python.framework/Versions/3.6/lib/python3.6/inspect.py", line 968, in getsource
    lines, lnum = getsourcelines(object)
  File "/opt/local/Library/Frameworks/Python.framework/Versions/3.6/lib/python3.6/inspect.py", line 955, in getsourcelines
    lines, lnum = findsource(object)
  File "/opt/local/Library/Frameworks/Python.framework/Versions/3.6/lib/python3.6/inspect.py", line 786, in findsource
    raise OSError('could not get source code')
OSError: could not get source code

Below is the output when I run the exact same commands in an ipython shell (within the same virtualenv):

(tf2) $ ipython
Python 3.6.6 (default, Jun 28 2018, 05:43:53)
Type 'copyright', 'credits' or 'license' for more information
IPython 7.2.0 -- An enhanced Interactive Python. Type '?' for help.

In [1]: import tensorflow as tf

In [2]: @tf.function
   ...: def foo():
   ...:     return 42
   ...:

In [3]: foo()
2019-01-08 13:02:33.785802: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
Out[3]: <tf.Tensor: id=5, shape=(), dtype=int32, numpy=42>

The result is also good in Jupyter (within the same virtualenv).

This is the output when I switch to another virtualenv based on TF 1.12.0:

(tf1) $ python
Python 3.6.6 (default, Jun 28 2018, 05:43:53)
[GCC 4.2.1 Compatible Apple LLVM 9.1.0 (clang-902.0.39.2)] on darwin
Type "help", "copyright", "credits" or "license" for more information.
>>> import tensorflow as tf
>>> tf.enable_eager_execution()
>>> @tf.contrib.eager.defun
... def foo():
...     return 42
...
>>> foo()
2019-01-08 12:55:53.087099: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
<tf.Tensor: id=5, shape=(), dtype=int32, numpy=42>

[martinwicke]

I believe this is an autograph issue?

[mdanatg]

Yes, it's a currently known limitation. Libraries like dill seem to be are able to get past that, we'll look into those.

In the mean time, you can use @tf.function(autograph=False) inside a Python shell, but that means you'll have to use graph code.

We could revert to autograph=False automatically and output a warning in cases like this, I wonder if that would be a better user experience?

[ageron]

Thanks for your feedback @mdanatg. Indeed, until a fix is found, I like the idea of reverting to autograph=False and outputting a warning. Or perhaps simply having a more explicit error message?

[mdanatg]

Agreed, we'll definitely add a better error message.

[alextp]

We now revert to autograph=False, so closing this issue for now.