Skip to content

Latest commit

 

History

History
81 lines (62 loc) · 3.2 KB

kaldi_memo.md

File metadata and controls

81 lines (62 loc) · 3.2 KB

Kald Install

- Kaldi installation
    - GPU Version :  Ubuntu 16.04  Cuda = 8.0 or 10.0

- Install MKL

Setting mkl-root to your installation location (eg. ~/Intel/mkl/)
Set mkl.h header file:
export CPLUS_INCLUDE_PATH="/home/luoxuan/intel/compilers_and_libraries_2019.5.281/linux/mkl/include"

- Install protobuf

sudo apt-get install autoconf automake libtool curl make g++ unzip -y
git clone https://github.com/google/protobuf.git
cd protobuf
git submodule update --init --recursive
./autogen.sh
make
make check
sudo make install
sudo ldconfig

- Install other  -> warpctc_pytorch (???)

- Compile Kaldi
Check kaldi dependency by  ./check_dependencies.sh  in kaldi/tools/extras
Install kaldi
    •	kaldi/tools/extras, run ./check_dependencies.sh and check that everything is okay
    •	kaldi/tools, run make OPENFST_VERSION=1.6.9 -j 8
    •	kaldi/src, run ./configure --shared && make depend && make -j 8
=>./configure --shared --mkl-root=/home/luoxuan/intel --mkl-libdir=/home/luoxuan/intel/compilers_and_libraries_2019.5.281/linux/mkl/lib/intel64_lin/
    •	kaldi/tools, run ./install_srilm.sh (optional)

- Install espnet

Install Kaldi CPU

  • Compile Kaldi
    • add --no-certificate-security
  • Python Environment
    • Change mac-version miniconda.sh, not linux
    • Change to espnet-conda env, instead of default conda
    • build warpctc-pytorch10-cpu, instead of pip install, cause no such
    • Install llvm(clang++) libomp to compile since fopenmp not supported in Mac gcc4.2

make CUPY_VERSION='' PYTHON_VERSION=3.7 TH_VERSION=1.2.0 unsupported option '-fopenmp'

cmake -DCMAKE_C_COMPILER="/usr/local/opt/llvm/bin/clang" -DCMAKE_CXX_COMPILER="/usr/local/opt/llvm/bin/clang++" ..

Install GPU

Kaldi architect

In egs/**/s5/
cmd.sh                     # 并行执行命令，通常分 run.pl, queue.pl 两种
config                       # 参数定制化配置文件， mfcc, decode, cmvn 等配置文件
local                        # 工程定制化内容
path.sh                    # 环境变量相关脚本
run.sh                      # 整体流程控制脚本，主入口脚本
steps                       # 存放单一步骤执行的脚本
utils                         # 存放解析文件，预处理等相关工具的脚本

要介绍了两种方式执行： 常见用法： run.pl some.log a b c 即在 bash 环境中执行 a b c 命令，并将日志输出到 some.log 文件中 并行任务： run.pl JOB=1:4 some.JOB.log a b c JOB 即在 bash 环境中执行 a b c JOB 命令，并将日志输出到 some.JOB.log 文件中, 其中 JOB 表示执行任务的名称， 任意一个 Job 失败，整体失败。

GPU
      run.pl --ngpu 3 some.log a
run.pl some.log my-prog "--opt=foo bar" foo \|  other-prog baz
and run.pl will run something like:
( my-prog '--opt=foo bar' foo |  other-prog baz ) >& some.log