Mingkun Huang

Since I'm not quite familiar with Lingvo, you'd best ask Google to open source their version of RNN-T.

You mean use tensorflow binding in this pytorch training code?

Of course you can. Just convert pytorch tensor into tensorflow tensor, and call tensorflow rnnt loss.

No need.

Try to remove python 'RegisterShape', then add a [SetShapeFn](https://www.tensorflow.org/guide/create_op#shape_functions_in_c) after [warprnnt_op.cc:13](https://github.com/HawkAaron/warp-transducer/blob/master/tensorflow_binding/src/warprnnt_op.cc#L13).

@housebaby which kind of GPU did you use?

Well, the gradidents here are to log probability, \frac{\partial L}{\partial \log p} = \frac{\partial L}{\partial p} \frac{\partial \exp{\log p}}{\partial \log p} = p \frac{\partial L}{\partial p}.

Please run test.py outside the pytorch_binding dir.

Here is my setup: https://github.com/HawkAaron/warp-ctc/blob/pytorch_bindings/tensorflow_binding/setup.py

To make things easy, you can run the following code in bash cmdline: ```bash export CUDA_HOME=/path/to/cuda ```