安装cuda 非root_无root权限离线安装ESPNet

分手后的思念是犯贱 2022-11-19 12:52 861阅读 0赞

ESPNet依赖于Kaldi，首先需要编译好Kaldi。

shayxu：无ROOT权限离线安装kaldizhuanlan.zhihu.com

![Image 1][]

到GitHub上下载ESPNet。

espnetgithub.com

git clone https://github.com/espnet/espnet
    
    # 使用镜像
    git clone https://github.com.cnpmjs.org/espnet/espnet

进入espnet/tools文件夹下。

创建一个指向Kaldi根目录的软链接，指定CUDA位置和anaconda位置。

cd <espnet-root>/tools
    
    ln -s <kaldi-root> .
    
    . ./setup_cuda_env.sh <cuda-root>  # . ./setup_cuda_env.sh /opt/lib/cuda-8.0
    
    ./setup_anaconda.sh [output-dir-name] [conda-env-name] [python-version] 
    # ./setup_anaconda.sh /home/mkws/ruixu11/anaconda

打开espnet/tools文件夹下的Makefile文件，按照代码一步一步手动安装相应依赖。

sctk.done:
        ./installers/install_sctk.sh
        touch sctk.done
    
    # install_sctk.sh
      wget -T 10 https://github.com/espnet/kaldi-bin/releases/download/v0.0.2/sctk-2.4.10-20151007-1312Z.tar.bz2 || 
      wget -T 10 -t 3 ftp://jaguar.ncsl.nist.gov/pub/sctk-2.4.10-20151007-1312Z.tar.bz2|| 
      wget --no-check-certificate -T 10 http://www.openslr.org/resources/4/sctk-2.4.10-20151007-1312Z.tar.bz2
      # 下载sctk
      tar xojf sctk-2.4.10-20151007-1312Z.tar.bz2
      ln -s sctk-2.4.10 sctk
      # 解压并创建软连接
      cd sctk
      make config
      touch .configured
      make all && make install && make doc
      # 编译安装sctk。

编译并安装sctk，创建sctk.done空文件标志stck安装成功。

sph2pipe.done:
        ./installers/install_sph2pipe.sh
        touch sph2pipe.done
    
    # install_sph2pipe.sh
       wget -T 10 https://github.com/espnet/kaldi-bin/releases/download/v0.0.2/sph2pipe_v2.5.tar.gz || 
       wget -T 10 -t 3 http://www.openslr.org/resources/3/sph2pipe_v2.5.tar.gz || 
       wget --no-check-certificate -T 10  https://sourceforge.net/projects/kaldi/files/sph2pipe_v2.5.tar.gz
       # 下载sph2pipe
       tar --no-same-owner -xzf sph2pipe_v2.5.tar.gz
       # 解压sph2pipe压缩包
       cd sph2pipe_v2.5
       gcc -o sph2pipe  *.c -lm
       # 编译安装sph2pipe。

编译安装sph2pipe 。

warp-ctc.done: pytorch.done conda_packages.done
        . ./activate_python.sh && ./installers/install_warp-ctc.sh
        touch warp-ctc.done
    
    # install_warp-ctc.sh
      pip install warpctc-pytorch==${
         warpctc_version}+torch"${
         torch_version}".cuda"${
         cuda_version}"
      # warpctc_version=0.2.1,
      # pytorch>=1.1.0可通过pip直接安装warp-ctc
    
     # 若pytorch低于1.1.0，则手动安装
      git clone https://github.com/espnet/warp-ctc.git
      cd warp-ctc
            git checkout -b pytorch-0.4 remotes/origin/pytorch-0.4
            mkdir build
    
            (
                set -euo pipefail
                cd build && cmake .. && ${
         MAKE}  #MAKE=make
            )
    
            python3 -m pip install cffi
            (
                set -euo pipefail
                cd pytorch_binding && python3 setup.py installl
            )

编译安装warp-ctc。

warp-transducer.done: pytorch.done conda_packages.done
       . ./activate_python.sh && ./installers/install_warp-transducer.sh
       . ./activate_python.sh &&  ./installers/install_warp-rnnt.sh $(CUDA_VERSION)
       touch warp-transducer.done
    
    # install_warp-transducer.sh
      git clone https://github.com/HawkAaron/warp-transducer.git
      (
        set -euo pipefail
        cd warp-transducer
    
        mkdir build
        (
            set -euo pipefail
            cd build && cmake .. && make
        )
    
        (
            set -euo pipefail
            cd pytorch_binding && python3 setup.py install
        )
      )
    
    # install_warp-rnnt.sh
      git clone https://github.com/1ytic/warp-rnnt
    
      (
        set -euo pipefail
        cd warp-rnnt/pytorch_binding && python3 setup.py install
      )

编译安装warp-transducer。

chainer_ctc.done: chainer.done
        . ./activate_python.sh && ./installers/install_chainer_ctc.sh
        touch chainer_ctc.done
    
    # install_chainer_ctc.sh
      git clone https://github.com/jheymann85/chainer_ctc.git
      cd chainer_ctc && chmod +x install_warp-ctc.sh && ./install_warp-ctc.sh && python3 -m pip install .
      # 调用warp-ctc安装脚本，再执行一遍warp-ctc安装。
      # 将warp-ctc文件夹放置在tools文件夹下，将git clone命令注释了，再执行install_chainer_ctc.sh脚本。

编译安装chainer\_ctc。

nkf.done:
        ./installers/install_nkf.sh
        touch nkf.done
    
    pyopenjtalk.done: espnet.done conda_packages.done
        . ./activate_python.sh && ./installers/install_pyopenjtalk.sh
        touch pyopenjtalk.done
    
    moses.done:
    	git clone --depth 1 https://github.com/moses-smt/mosesdecoder.git moses
    	touch moses.done
    
    mwerSegmenter.done:
    	./installers/install_mwerSegmenter.sh
    	touch mwerSegmenter.done
    
    kenlm.done: espnet.done conda_packages.done
    	. ./activate_python.sh && ./installers/install_kenlm.sh
    	touch kenlm.done
    
    ./installers/install_pesq.sh
    
    py3mmseg.done: espnet.done
    	. ./activate_python.sh && ./installers/install_py3mmseg.sh
    	touch py3mmseg.done

安装各个依赖。

. ./activate_python.sh; python3 check_install.py

检查安装是否成功。

从openslr下载wave\_yesno数据集，并解压到/espnet/egs/yesno/asr1文件夹下。

openslr.orgwww.openslr.org

![d260a9469666f9d64e877503839a5cd6.png][]

运行run.sh脚本，观察训练结果，若无错误发生，则安装成功。

bash run.sh

[Image 1]: 
[d260a9469666f9d64e877503839a5cd6.png]: /images/20221119/ea2d6cddd7dd44a5963296078fa96e94.png