Don't dealloc strings if the TensorBuffer is shared (#357)

[brychcy]

No description provided.

[saudet]

There's no more straightforward way to get the information? It doesn't sound like TF_TensorMaybeMove() is always going to return null when the content is being shared.

[brychcy]

There's no more straightforward way to get the information? It doesn't sound like TF_TensorMaybeMove() is always going to return null when the content is being shared.

If there is, I haven't found it.

BTW, what I really wonder is: Shouldn't string deallocation be done automatically on the native Tensorflow level when the last tensor that uses the TensorBuffer is released?

[saudet]

No, that doesn't happen. We need to allocate the strings manually, and deallocate them manually or we end up with memory leaks, see issue #251.

[karllessard]

Even if we are not 100% sure that this fix will cover all cases, it seemed to work for @brychcy so let's merge it to give it a try and see. Thanks @brychcy !

[karllessard]

Oops, it looks like the modified code is not formatted correctly, @brychcy can you please run mvn spotless:apply locally before pushing a newer version of this PR?

[brychcy]

Oops, it looks like the modified code is not formatted correctly, @brychcy can you please run mvn spotless:apply locally before pushing a newer version of this PR?

Done.

At first it didn't do anything for me because origin/master pointed to my version and
<ratchetFrom>origin/master</ratchetFrom>
is used.

As the whole file was reformatted, I have done it in a separate commit.

[karllessard]

Thanks a lot @brychcy for fixing this!