Skip to content

Latest commit

 

History

History
166 lines (111 loc) · 4.61 KB

installing-cat-tutorial.md

File metadata and controls

166 lines (111 loc) · 4.61 KB

Installation of CAT

This is a step-by-step tutorial of installation of CAT, especially for beginners, including the installation of dependencies PyTorch and Kaldi.

Dependencies

CAT requires two (main) dependencies: PyTorch and Kaldi. Some other dependent packages such as h5py for reading data is not included here, which can be easily installed with pip/pip3.

We recommend to install CAT and its dependencies in conda environment, otherwise you may require root permission for installing some libraries.

And note that CAT is relied on CUDA availability. Directly running with CPU is not supported yet.

PyTorch

  1. Check your CUDA version.

    $ whereis cuda

    The output includes the path to your CUDA, normally /usr/local/cuda, then check the version file.

    $ cat <path to CUDA>/version.txt

  2. Install the corresponding PyTorch version.

    The official guide of PyTorch gives very details of installing the latest stable version with different methods.

    For previous version of PyTorch or if your CUDA version not listed, refer to this and find your suitable version.

    Python2 with PyTorch 0.4+ or Python3 with PyTorch 1.0+ are supported by CAT.

Kaldi

Please refer to this step-by-step installation guide: http://jrmeyer.github.io/asr/2016/01/26/Installing-Kaldi.html

Note that IRSTLM tool is required. The directory of kaldi is needed in following steps, denoted as <path to kaldi>.

OpenFST

  1. Create a temporary directory and change the working directory to it.

    $ mkdir tmpfst && cd tmpfst

  2. Get OpenFST-1.6.7 file and extract.

    # fetch the file from source
$ wget http://www.openfst.org/twiki/pub/FST/FstDownload/openfst-1.6.7.tar.gz
# extract
$ tar -xf openfst-1.6.7.tar.gz
# change directory to extracted one
$ cd openfst-1.6.7/

  3. Make configuration of compiling. Without --prefix option, the OpenFST will be installed to /usr/local. Individual directory for current user is recommended.

    $ ./configure --prefix=<path to openfst>

  4. Make and install. This step may require root permission up to the <path to openfst> you specified in previous step.

    make && make install

  5. [OPTIONAL] Remove the temporary directory.

    $ cd ../..
$ rm -r tmpfst/

Install CAT

  1. Get source codes from GitHub. Denote the path as <path to CAT>.

    $ git clone https://github.com/thu-spmi/CAT

  2. Install the kaldi-patch of CAT into kaldi.

    # copy the patch file into kaldi
$ cp <path to CAT>/src/kaldi-patch/latgen-faster.cc <path to kaldi>/src/bin/
# change directory
$ cd <path to kaldi>/src/bin/

    Edit the Makefile:

    Add a string "latgen-faster" to the BINFILES list. Then compile.

    $ make

    If everything goes well, there should be two new files in <path to kaldi>/src/bin/: latgen-faster and latgen-faster.o.

  3. Install CTC-CRF module.

    # change directory
$ cd <path to CAT>/src/ctc_crf/

    CTC-CRF module will be installed as a python module. So, before going to next step, please ensure the python (or python3) in your current path is the one you are to use with CAT:

    # for python2
$ which python
# for python3
$ which python3

    For Python2 installation, run following command and go to step 4.

    # run as root if the command raise a permission denied error.
$ make OPENFST=<path to openfst>

    For Python3, firstly open the Makefile and comment the last line of command. Like this

    ...
CTCCRF: GPUCTC GPUDEN PATHWEIGHT
#	python setup.py

    Then compile dependencies of CTC-CRF module.

    $ make OPENFST=<path to openfst>

    Finally, install the CTC-CRF module.

    # run as root if the command raise a permission denied error.
$ python3 setup_1_0.py install

  4. Try import CTC-CRF module.

    $ python
# or python3
$ python3
>>> import torch
>>> import ctc_crf_base

    There should be no error message thrown out. And note that because CTC-CRF module is relied on torch, you should import torch before import ctc_crf_base.

  5. Further tiny works.

    Change the KALDI_ROOT=... in <path to CAT>/scripts/ctc-crf/kaldi_io.py and <path to CAT>/egs/<task>/path.sh to <path to kaldi>. Here the <task> is your ASR task directory, like wsj and swbd.

  6. Enjoy it! 🚀