Closed lissyx closed 7 years ago
This seems like a bug in TensorFlow not an API change.
The documentation[1] states the first argument inputs
is a normal Tensor. However it also states labels
the second argument is a SparseTensor.
We pass inputs
as a normal Tensor and labels
as a SparseTensor. So there should be no problem.
Ah ah, next time, use named arguments I guess :D https://github.com/tensorflow/tensorflow/commit/e72ad6333ee0204b7d2df93df1f7a85ec4b9ab10
Yeah, that's an annoying API change for no good reason.
I guess we should address this issue by simply using named arguments here.
Got most of the issues figured out, we should just ensure about that tf.variable_scope() thing.
FYI the PR for TensorFlow is at https://github.com/mozilla/tensorflow/pull/7 and it contains not only fixes, but also updates warpctc python code to remove deprecated use of tf.concat(), changes PIP package naming (hopefully would avoid new upstream to install over warpctc build). With that build I could run LDC93S1 sample with both CTC and WarpCTC, successfully.
Using #324 as well as mozilla/tensorflow#7, I am able to run on LDC93S1 (tested only CTC, have not tested WarpCTC yet).
On top of the changes for 0.12, we need to change back tf.concat_v2() to tf.concat()
Overfitting LDC93S1 without any problem.
This PR https://github.com/mozilla/tensorflow/pull/11 should be good enough now.
PR mozilla/tensorflow#11 merged in https://github.com/mozilla/tensorflow/commit/616a8ea8626e318e98daa794197ec85a0bb8794a
Automation has been switched to TF 1.0.0, relying on upstream official package (because of segfault when using ours. Debugging this is halted until further notice)
This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.