Server 84

• Only user amax is allowed to use sudo on server 84. If need to use, you can
  • post a new issue here.
  • contact chenming or luzai, to access sudo from amax.
• If need to download config files, contact luzai, he will
  • distribute an id_rsa file so that you can access this private repo.
  • or add you to this repo as one of contributors.
• Feel free to issue and fix errors, including typos!

Connect to Internet

Here provide a method to share Internet of your laptop to server.

Prerequisite:

• openssh-client
  • windows system can try:
    • cygwin and install openssh-client
    • linux subsystem and install openssh-client
    • [Potentially work] some gui manipulations in putty, I fail to find the way. Tell me if you are familiar with putty.
  • ubuntu: sudo apt-get install openssh-client

Forward Dynamic Port

In the most case, this works:

ssh -gfNTD 1080 username@serverip
proxychains curl ip.gs

If 1080 is being used, you can

• use existed 1080, and of course, other user’s Internet. In the most time, luzai will share his Internet to server.
• Start a new port, as shown below

cp /etc/proxychains.conf ./
vim ./proxychains.conf # and modify 1080 to 1081
ssh -gfNTD 1081 username@serverip
proxychains curl ip.gs

Post Work

For convenience, you can add this to .bashrc or other rc files you are using.

alias git='proxychains git'
alias pip='proxychains pip2'
alias pip2='proxychains pip2'
alias pip3='proxychains pip3'
alias wget='proxychains wget -c '
alias conda='proxychains conda'

Be careful to use tensorflow!

Limit to one Device

In terminal, before running program, set environment variable.

export CUDA_VISIBLE_DEVICES=0,1
python train.py
python resume.py

CUDA_VISIBLE_DEVICES=0,1 python train.py

In python, set environment variable.

def init_dev(n=(0,)):
    from os.path import expanduser

    logging.info('use gpu {}'.format(n))
    home = expanduser("~")
    if isinstance(n, int):
        n = (n,)
    devs = ''
    for n_ in n:
        devs += str(n_) + ','
    os.environ["CUDA_VISIBLE_DEVICES"] = devs
    # set_env('PATH', home + '/cuda-8.0/bin' )
    # set_env('LD_LIBRARY_PATH', '/usr/local/cuda-8.0/lib64')


def set_env(key, value):
    os.environ[key] = value + ':' + os.environ[key]

In python: use tf.device when construct graph.

Allow Growth

In order to know how much memory I consume, really.

def allow_growth_tf():
    import tensorflow as tf
    _sess_config = tf.ConfigProto(allow_soft_placement=True)
    _sess_config.gpu_options.allow_growth = True
    return _sess_config


def allow_growth_keras():
    import tensorflow as tf
    tf_graph = tf.get_default_graph()
    _sess_config = allow_growth_tf()
    sess = tf.Session(config=_sess_config, graph=tf_graph)
    import keras.backend as K
    K.set_session(sess)
    return sess

Other config files

System config: Apt source, use one of:

• [x] ustc’s ipv6 mirror. (Now we are using)
• zju’s 14.04 mirror.
• aliyun 14.04 mirror.

Copy them to your home path, i.e., ~.

• .condarc: use ustc’s mirror
• .pip: use tsinghua’s mirror

Copy them to current directory to override system config.

• proxychains.conf: proxy for http(s) and sock5
• tsocks.conf: proxy for sock5