cluster-setup

How-to

When you install the cuda GPU library and tensorflow, please check wheather they are compatible. You also should check python default version and pip default version.

SSH login cluser:

ssh `username@hostaddress`

Install git on the cluster

apt-get install git

Setup from the scrtip

. `cluster-setup.sh`

SSH tricks & FAQ

1. screen

Sometimes you want to run some scripts via command line, and you may close the command line and check the results later. Then you can use screen.

• Start a screen

screen -S `your-preferred-screen-name`

• Use ctrl+A, ctrl+D to leave the screen.

• Check results later

screen -r `your-preferred-screen-name`

• Kill a screen

screen -S  `your-preferred-screen-name` -X quit

• Alternative: tmux

2. error from matplotlib

Source

You want to save image, but you may get _tkinter.TclError: no display name and no $DISPLAY environment variable. You can simply run the following command before running your code:

export MPLBACKEND="agg"

3. cuda9 not compatible for tensorflow

sudo apt autoremove cuda

use cuda8 instead

sudo apt-get install cuda-8-0

4. split data larger than 25MB to update to GitHub

• help

man split

• Command

split --byte="20M" `file_name` `prefix for the splitted file`

or

split -b 20m `file_name` `prefix for the splitted file`

e.g.

split --byte="25M" `largefile` `part`

• post process move the spliited files to a folder file_name

mv prefix* `./file_name`

e.g.

mv part* `./largefile`

• merge

cat * > `file_name`

Authors

• Jun Lu
• Wei Ma

References

• CS231n: Google Cloud Tutorial. link