tf.split_v() is removed in master?

[wookayin]

A bit of history:

• Until <= 0.12, tf.split() has the following signature:
  tf.split(split_dim, num_split, value, name='split')
• In commit f803bd7 (since 0.12), a new op split_v() is introduced:
  tf.split_v(value, size_splits, split_dim=0, num=None, name="split_v")
• In commit a46b6d2 (master), tf.split_v() is finally renamed to tf.split():
  tf.split(value, num_or_size_splits, axis=0, num=None, name="split")

Due to these changes, the signature of tf.split() has been changed. AFAIK TensorFlow will having some of breaking changes after 0.12 (after the end of 2016); is this backward-incompatible API change intended?

If so, my suggestion is that tf.split_v(), which has been introduced in 0.12, should not be removed in the newer versions as well. In the current master, tf.split_v is non-existent.

AttributeError: 'module' object has no attribute 'split_v'

I am reporthing this (minor) issue because I am frequently switching the tensorflow versions, from r0.12 (stable branch) to master (the breaking? future), and thus I need a way to write a code that is both compatible in those two versions. However, due to the change of tf.split(), it seems that I cannot achieve it at the moment.

Thanks,

[aselle]

We are making several backward incompatible changes. The final state is what you are seeing in the latest master branch where split_v() is removed and split() exists with the keyword arguments

tf.split(value, num_or_size_splits, axis=0, num=None, name="split")

To switch between versions, you will need to creat a local compatiblity wrapper that keys off of tf.__version__ or uses exceptions to try the newest form and if the keyword arguments don't match use the previous form (this is probably easier). Hope this helps.

[wookayin]

Thanks for the feedback. I was assuming backward-incompatible changes, but also was confused since I could find there is a similar change on tf.concat(): in r0.12 we have both concat_v2() and concat() (I expect the v2 method will replace the original soon).

However, if the changes are intended, perhaps we might not need to add a method such as tf.split_v() because it eventually would be a legacy, though the changes seem quite radical. For now, it would be better for me to try alternative approaches as you suggested.

[tfboyd]

Automatically closing due to lack of recent activity. We hope that you were able to resolve it on your own. However, since this is a support issue rather than a bug or feature request, you will probably get more information by posting it on StackOverflow.

[wookayin]

Yeah, I think this is not necessary any more since it has been a while after TF 1.0 came out.