Installation

Ubuntu Swap

$ free
$ sudo gedit /etc/sysctl.conf
#################################################
# Set Swap usable availability
#        Swap <-------> physical memory
# default: 60%                40%
vm.swappiness=20

Hide mounted Drives logo infiles sidebar

Use the mount option x-gvfs-hide in /etc/fstab to hide it in nautilus, for example.

For example, a line in /etc/fstab would become:

/dev/sda1 /mnt/sda1 ext4   defaults,x-gvfs-hide       0     2

Install Nvidia Driver, cuda, cuDnn, Anaconda, Tensorflow

Nvidia Driver can be installed by CUDA (Some times doesn't work)

$ sudo apt-get purge nvidia*
# Note this might remove your cuda installation as well
$ sudo apt-get autoremove 

For a new PC you still need $ sudo apt-get install build-essential gcc-multilib dkms to install the depends.

• In this condiction, When "X server is running"

$ sudo service lightdm stop
$ rm /tmp X*-lock

• Add to path to ~/.bashrc

export PATH=/usr/local/cuda/bin:${PATH}
export LD_LIBRARY_PATH=/usr/local/cuda/bin:${LD_LIBRARY_PATH}

• When you update your Ubuntu system, it may cause a login loop.

1. check the authority of the .Xauthority $ ls -lah if root then $ sudo chown username:username .Xauthority.
2. If after this the login loop keeps,reinstall the Nvidia driver

• Anaconda couldn't install check the owner of anaconda

Independent install Nvidia Driver, follow this

ATTENTION

• If you can not solve the nouveau problem, try -no-x-check -no-nouveau-check, the check is not necessary, in this

multiple CUDA

$ cd /usr/local   #cuda is a soft link
$ stat cuda

File: 'cuda' -> '/usr/local/cuda-9.0'
  Size: 19        	Blocks: 0          IO Block: 4096   symbolic link
Device: 801h/2049d	Inode: 15337567    Links: 1
Access: (0777/lrwxrwxrwx)  Uid: (    0/    root)   Gid: (    0/    root)
Access: 2018-12-27 22:38:59.461332996 +0900
Modify: 2018-12-27 22:38:51.621298141 +0900
Change: 2018-12-27 22:38:51.621298141 +0900
 Birth: -

cuDNN

$ sudo cp cuda/include/cudnn.h /usr/local/cuda/include
$ sudo cp cuda/lib64/libcudnn* /usr/local/cuda/lib64
$ sudo chmod a+r /usr/local/cuda/include/cudnn.h /usr/local/cuda/lib64/libcudnn*

openCV with videoCapture

$ pip uninstall opencv-python
$ conda remove opencv. 

$ conda install -c anaconda opencv

With the method mentioned above, we meet the problem

The function is not implemented. Rebuild the library with Windows, GTK+ 2.x or
Carbon support. If you are on Ubuntu or Debian, install libgtk2.0-dev and
pkg-config, then re-run cmake or configure script

which solved by below

$ conda install -c menpo opencv=2.4.11

VS Code

• VS Code cannot open

$ code --verbose
[main 20:19:26] Startup error: 
Error: EACCES: permission denied, mkdir '/home/<user>/.config/Code/CachedData'

# sure enough the folder ~/.config/Code had root access permissions for some reason. Deleted the folder using sudo.

$ rm -rf /home/<user>/.config/Code 

• VS Code could not install extensions
  The owner is root.
  $ sudo chown username:username VSCode/

exFAT in Ubuntu

exFAT works fine in Win and MacOS, for Ubuntu,
$ sudo apt install exfat-fuse exfat-utils reboot to make it work. (In my case, without logout, the disk can be mount, but fail to write.)

Reproduce Keras in Tensorflow for 'Deep Image Homographt Estimation'

When we turn the Keras model into normal tf Graph, with totally same parameters and setting, the Loss does NOT constringe. The prediction of our tf network is arrange in a very narrow range around 0 after thousands of training rounds, while the Keras model grows into a reasonable range(single digit).
When we decrease the learning rate, Adam(lr=1e-4), the tf model also constringe. BUT the rate of the tf model is slow than the Keras model and the Keras also have a better monotonicity. Still waiting for the final constringency loss difference.
By compareing the two models, the same model, same optimizer & learning_rate. The difference may caused by the initializer or the model.compile

A small problem about how to update sublime $ sudo apt-get install sublime-text

VPN & SSR Setting

A Google cloud platform(GCP) based method is published here. Two types (VPN/SSR) have been tested.

SSR

here is a reference for SSR. But there is a problem that the connection will be blocked after several hours and changing port number can bring the connection back for several hours.

VPN

The Windows 10 firewall bring some troubles on both these two types. Here is the setting of Registry and [here] is the setting of the firewall rule.

• PPTP

Here is a reference for PPTP.

• L2TP

Here is a reference for L2PT.