Combination of fast_rnnt and fast_emit

[Butterfly-c]

Is there any version that take advantage of fast_emit?

[pkufool]

Sorry for the late reply, we have a discussion in k2 repository (k2-fsa/k2#955) and we are doing experiments, finally we will add something to make symbols emit earier.

[pkufool]

@Butterfly-c The delay_penalty in k2-fsa/k2#955 has been merged, you can try it. It behaves as good as fast_emit. If you really want the fast_emit, you can try k2-fsa/k2#1069 (by installing k2 or modifying fast_rnnt according to this PR).

[BOTQQ-Lee]

help help help

[csukuangfj]

Please post, all, all, of the logs. The screenshot contains little information and it is hard for us to figure out what bad things happened.

[BOTQQ-Lee]

[csukuangfj]

please first use

pip install cmake

and then re-try.

[csukuangfj]

Please post, all, all, of the logs. The screenshot contains little information and it is hard for us to figure out what bad things happened.

[csukuangfj]

-- Could NOT find CUDNN (missing: CUDNN_LIBRARY_PATH CUDNN_INCLUDE_PATH)

From the error log, you have not installed cuDNN yet.

Also, your installed CUDA is 9.2, which is quite old. I am not sure whether it works.
I suggest that you upgrade your CUDA and PyTorch.

[csukuangfj]

CMake Error at pybind11-subbuild/pybind11-populate-prefix/src/pybind11-populate-stamp/download-pybind11-populate.cmake:170 (message):
Each download failed!

You have network connection problems with github.com. Please re-try.

[csukuangfj]

-- No NVCC detected. Disable CUDA support

From the above logs, I suggest installing a CUDA version of PyTorch. Otherwise, it would be slow in training with CPU.

[csukuangfj]

From the output of cmake

-- C++ Standard version: 14
-- No NVCC detected. Disable CUDA support

[csukuangfj]

How did you compile fast_rnnt? Do you have the compilation logs?

The error logs show that you have compiled a CPU version of fast_rnnt.