TensorFlow New Op: AttributeError: 'module' object has no attribute 'custom_op'

Question

I am creating new ops (https://www.tensorflow.org/extend/adding_an_op) for TensorFlow (r1.0) running both on x86 and ARMv7.

Minor code modifications are necessary to run TensorFlow on ARMv7, but this guide helps a lot: https://github.com/samjabrahams/tensorflow-on-raspberry-pi/blob/master/GUIDE.md.

But I noticed that the custom operations do not work on my ARMv7 installation of TensorFlow.

For example, when I test my custom operation in a Python script on ARMv7:

import tensorflow as tf
_custom_op_module = tf.load_op_library('custom_op.so')
custom_op = _custom_op_module.add_stub

I get the following error (that does not show up on x86):

$ python test_custom_op.py 
  Traceback (most recent call last):
  File "custom_op.py", line 3, in <module>
    add_stub = _custom_op_module.add_stub
  AttributeError: 'module' object has no attribute 'custom_op'

I further investigated the issue, and apparently there is not my custom operation in the .so library file.

$ python
>>> import tensorflow as tf
>>> _custom_op_module = tf.load_op_library('custom_op.so')
>>> dir(_custom_op_module)
>>> ['LIB_HANDLE', 'OP_LIST', '_InitOpDefLibrary', '__builtins__', '__doc__', '__name__', '__package__', '_collections', '_common_shapes', '_op_def_lib', '_op_def_library', '_op_def_pb2', '_op_def_registry', '_ops', '_text_format']
>>> _custom_op_module.OP_LIST

>>>

The same commands on x86 have the following output:

>>> import tensorflow as tf
>>> _custom_op_module = tf.load_op_library('custom_op.so')
>>> dir(_custom_op_module)
>>> ['LIB_HANDLE', 'OP_LIST', '_InitOpDefLibrary', '__builtins__', '__doc__', '__name__', '__package__', '_add_stub_outputs', '_collections', '_common_shapes', '_op_def_lib', '_op_def_library', '_op_def_pb2', '_op_def_registry', '_ops', '_text_format', 'custom_op']
>>> _custom_op_module.OP_LIST
op {
  name: "CustomOp"
  ...
}
>>>

Does anybody have similar issue? Can we consider this a bug?

Nicole Finnie · Accepted Answer · 2018-11-29 13:55:34Z

1

I hit a similar issue with a similar error message when I tried to load my new op, however, my problem was I tried to register a customized op that had the same op name as tensorflow, and that led to a name collision. Changing the name fixed it without recompiling TF.

The error message I encountered: AttributeError: module '6e237d88703da016805889179d3f5baa' has no attribute 'custom_op'

answered Nov 29, 2018 at 13:55

Nicole Finnie

1,62016 silver badges13 bronze badges

Sign up to request clarification or add additional context in comments.

Comments

GDG · Accepted Answer · 2017-03-01 16:09:46Z

0

Apparently, recompiling and re-installing the TF made it works.

answered Mar 1, 2017 at 16:09

GDG

4096 silver badges15 bronze badges

Collectives™ on Stack Overflow

TensorFlow New Op: AttributeError: 'module' object has no attribute 'custom_op'

2 Answers 2

Comments

Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

2 Answers 2

Comments

Comments

Your Answer

Sign up or log in

Post as a guest

Related