Models: TypeError: Input 'split_dim' of 'Split' Op has type float32 that does not match expected type of int32.

[kevinashaw]

Hi all.
I'm running a TF model that is not based on Inception and the new r0.12 is breaking it also. I should note that this model worked perfectly with r0.11.
I'm referencing issue #761 which seems to have the same error message.
Using git hash #7c16e2e and built using Bazel this afternoon (bazel build -c opt //tensorflow/tools/pip_package:build_pip_package)
Here is the error log:

TypeError                                 Traceback (most recent call last)
<ipython-input-13-c788373b2a77> in <module>()
     27 
     28 # Split data because rnn cell needs a list of inputs for the RNN inner loop
---> 29 Xnew = tf.split(0, n_steps, Xnew)
     30 # new shape: n_steps * (batch_size, n_hidden)
     31 

/Users/kevin/anaconda2/lib/python2.7/site-packages/tensorflow/python/ops/array_ops.pyc in split(value, num_or_size_splits, axis, num, name)
   1220   if isinstance(num_or_size_splits, six.integer_types):
   1221     return gen_array_ops._split(
-> 1222         split_dim=axis, num_split=num_or_size_splits, value=value, name=name)
   1223   else:
   1224     size_splits = ops.convert_to_tensor(num_or_size_splits)

/Users/kevin/anaconda2/lib/python2.7/site-packages/tensorflow/python/ops/gen_array_ops.pyc in _split(split_dim, value, num_split, name)
   3424   """
   3425   result = _op_def_lib.apply_op("Split", split_dim=split_dim, value=value,
-> 3426                                 num_split=num_split, name=name)
   3427   return result
   3428 

/Users/kevin/anaconda2/lib/python2.7/site-packages/tensorflow/python/framework/op_def_library.pyc in apply_op(self, op_type_name, name, **keywords)
    507             if input_arg.type != types_pb2.DT_INVALID:
    508               raise TypeError("%s expected type of %s." %
--> 509                               (prefix, dtypes.as_dtype(input_arg.type).name))
    510             else:
    511               # Update the maps with the default, if needed.

TypeError: Input 'split_dim' of 'Split' Op has type float32 that does not match expected type of int32.

[kevinashaw]

Not that I understand this code all that well, but I can't help but think that someone swapped the order of the parameters for the split method in line 1174 of array_ops.py.
The documentation clearly indicates the calling order for tf.split should be:

tf.split(split_dim, num_split, value, name='split')

But the actual code indicates:

def split(value, num_or_size_splits, axis=0, num=None, name="split"):

And when I correct the order of parameters to match the documentation, it seems to work better.

[asimshankar]

Apologies for the instability, there is some churn happening in the API in preparation for the 1.0 release keeping with our intention to follow semantic versioning. As a result, there are currently breaking changes between 0.11 and 0.12 (and will be more for the next release).

See RELEASE.md for a list of changes being made, for example, the change in argument ordering for tf.split, which will not match NumPy.

You might also find the conversion tool at https://github.com/tensorflow/tensorflow/tree/master/tensorflow/tools/compatibility useful to help navigate these API changes.

Hope that helps! Closing this issue out, let us know if your concerns weren't addressed.

[Nassimmd]

I am using Tensorflow 1.5.0rc0 and I still have this issue

[726663676]

I have the same issue. and address the problem by changing the parameter position:

i, j, f, o = tf.split(1, 4, hidden)

i, j, f, o = tf.split(hidden, 4, axis=1)