TensorFlow binding for CTC-CRF. An extension of thu-spmi CAT for Tensorflow.

This is a modified version of thu-spmi/CAT. I just modify the code to the new CPP Extensions API style of Tensorflow and refact the gpu_den code.

• Cuda Toolkit/CuDNN
• gcc/g++ 5.5.0
• Python3
• TensorFlow
• OpenFst
• Kaldi

Because CTC-CRF operator is based on CUDA Toolkit, so you should setting CUDA environment. For details, you can follow this link or TensorFlow official link.

1. Install CUDA Toolkit

• Follow this link to download and install CUDA Toolkit 10.1 for your Linux distribution.

• Installation instructions can be found here

• Install CUDNN

  • Go to https://developer.nvidia.com/rdp/cudnn-download
  • Create a user profile if needed and log in
  • Select cuDNN v7.6.5 (Nov 5, 2019), for CUDA 10.1
  • Download cuDNN v7.6.5 Library for Linux
  • Follow the instructions under Section 2.3.1 of the CuDNN Installation Guide to install CuDNN.

2. Environment Setup

   Append the following lines to ~/.bashrc or ~/.zshrc.

   export PATH=/usr/local/cuda-10.1/bin${PATH:+:${PATH}}
   export LD_LIBRARY_PATH=/usr/local/cuda-10.1/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}

3. Install TensorFlow with Anaconda virtual environment

   Create a virtual environment is recommended. You can choose Conda or venv. Here I use Conda as an example.

   # Install TensorFlow/cuda/nvcc first, reference is here:
   conda create --name tf pip python==3.7
   conda activate tf
   conda install tensorflow-gpu==1.15.0

4. Install CTC-CRF TensorFlow wrapper

   setting your TENSORFLOW_SRC_PATH and OPENFST.

   NOTE: This is an example, please don't simply copy to your terminal:

   # Create a symlink libtensorflow_framework.so.1 which references the original file  libtensorflow_framework.so
   ln -s /home/huanglk/anaconda3_202002/envs/tf_subcomp/lib/python3.7/site-packages/tensorflow_core/libtensorflow_framework.so.1 /home/huanglk/anaconda3_202002/envs/tf_subcomp/lib/python3.7/site-packages/tensorflow_core/libtensorflow_framework.so
   
   # export TENSORFLOW_SRC_PATH
   export TENSORFLOW_SRC_PATH=/home/huanglk/anaconda3_202002/envs/tf_subcomp/lib/python3.7/site-packages/tensorflow_core/
   
   # export OPENFST
   export OPENFST=/usr/local/

   • It will compile three modules with gcc/g++, include GPUCTC, PATHWEIGHT and GPUDEN.
   • It is worth mentioning that if the version of gcc/g++ >= 5.0.0 and less than 6.0.0 will be helpful for following pipeline.
   • Finally, Makefile will exetucate python3 ./setup.py install for CTC-CRF TensorFlow wrapper.

   Now, you can install CTC-Crf TensorFlow wrapper warp-ctc-crf.

   # Install warp_ctc_crf
   cd warp_ctc_crf
   make -j 32

• Test:

  • tests/test_ctc_crf_op.py is a simple test for Tensorflow API for CAT.

• Tensorflow API：

  • setup.py: ctc_crf_loss setup script. It only test in g++ 5.0+ and Tensorflow 1.14.0
  • ctc_crf_op_kernel.cc: Tensorflow C++ binding.
  • ctc_crf_tensorflow/__init__.py: Tensorflow API for ctc_crf_loss.

• Others:

  • gpu_ctc/: Just follow thu-spmi/CAT
  • gpu_den/: Code refactoring from thu-spmi/CAT.

CAT. https://github.com/thu-spmi/CAT.