ZhikangNiu/pre-train-dockerfile

An Intro to set up your Speech Docker environment and debug using VSCode

An Intro to set up your Docker environment and debug using VSCode

Created by zym22 on 2021.12.20

Based on SJTU-XLANCE Servers & Speech Pre-train Docker as examples.

Set up VSCode Remote

1. Download extensions in VSCode.

2. Set up config of multi-jumps. (For you need a jumping machine to get access the GPUs.)

# local at ~/.ssh/config
Host gauss
    HostName 202.xxx.xxx.xxx
    Port 5566
    User zym22
Host date
    HostName 192.xxx.xxx.xxx
    # ProxyJump gauss
    ProxyCommand ssh -A gauss -W %h:%p
    User zym22

3. Connect to the Server such as date.

Set up Docker environment

Before take steps in this section, make sure your terminal is like zym22@date:~$.

1. Set up Docker environment in zym22@date:~$.

docker pull zym22/pre-train:latest 
docker run -it --runtime=nvidia --name myDocker \ # choose a name
    --mount src=/mnt/xlancefs/home/zym22/data,target=/data/zym22,type=bind \ # access to your src
    # -v /mnt/xlancefs/home/zym22:/home/zym22 \
    -p 8023:22 \ # choose a port
    --ipc=host    # host shares memory with the container
    zym22/pre-train:latest /bin/bash
docker start myDocker
docker exec -it myDocker /bin/bash

2. Set up Docker environment in root@myDocker:~$.

passwd root
apt update
apt install -y openssh-server
vim /etc/ssh/sshd_config
###
# edit here
Port 22                     #Enabling Port 22
PermitRootLogin yes         #Allow root user to log in using ssh
RSAAuthentication yes       #Enabling RSA Authentication 
PubkeyAuthentication yes    #Enable public-private key pairing Authentication
# wq
###
service ssh restart
exit

3. Verify that port mapping is correct.

zym22@date:~$ docker port myDocker 22
0.0.0.0:8023

4. Edit local config at ~/.ssh/config.

# local at ~/.ssh/config
Host gauss
    HostName 202.xxx.xxx.xxx
    Port 5566
    User zym22
Host date
    HostName 192.xxx.xxx.xxx
    # ProxyJump gauss
    ProxyCommand ssh -A gauss -W %h:%p
    User zym22
Host myDocker
    HostName 192.xxx.xxx.xxx # same as date
    Port 8023
    # ProxyJump gauss
    ProxyCommand ssh -A gauss -W %h:%p
    User root

5. Connect to the Docker.

Debug using VSCode

Before take steps in this section, make sure your terminal is like root@myDocker:~$

1. Create a launch.json in .vscode folder.

# launch.json
{
    "version": "0.2.0",
    "configurations": [
        {// python -m debugpy --listen 5678 --wait-for-client main.py args
            "name": "Python: attach local",
            "type": "python",
            "request": "attach",
            "console": "integratedTerminal",
            "env": {
                "CUDA_VISIBLE_DEVICES": "1" //edit
            },
            "connect": {
                "host": "localhost",
                "port": 5678 //edit
            }
        }

    ]
}

2. Run a python file and block.

pip install debugpy
python -m debugpy --listen 5678 --wait-for-client main.py args
# An example to debug wav2vec2.0 using fairseq.
python -m debugpy --listen 5678 --wait-for-client ../miniconda3/envs/espnet/bin/fairseq-hydra-train task.data=examples/wav2vec/manifest --config-dir examples/wav2vec/config/pretraining --config-name wav2vec2_test_librispeech

3. Start debugging mode.

About the Speech Pre-train Docker (Latest tag: 20220218)

A docker for Speech, compatible with Kaldi, Espnet and Fairseq.

Basic environment：

• ubuntu=16.04
• cudnn=7 cuda=10.2
• python=3.8 torch=1.10.0 torchaudio=0.10.0
• condaenv=espnet

Basic Directory structure:

- root
	- miniconda3
	- .bashrc
	- .profile
- data
	- zym22
	- xc095
- espnet
- fairseq -> espnet/tools/fairseq
- kaldi -> espnet/tools/kaldi