about using the lab server's GPU and running the TensorFlow code

Connection Server Windows-xshell xftp SSH

1. Connect to a lab server via SSH

• Using SSH connection is no stranger. GitHub and OS classes are often used
• Currently using 192.168.7.169

1. Using Tools Xshell and Xftp

• Using Xshell to connect servers and operations, the Ubuntu 16.04 LTS operating system is installed on each node of the server
• Managing Files with Xftp

1. Resources:
   Xshell+xftp SSH Tunnel Proxy
   Xshell connection to Linux server via SSH key and SSH proxy

Mac os-terminal Cyberduck

Because the computer at the lab station is a Mac that can only be re-acquainted with a wave.

1. Using terminal to establish an SSH remote connection
2. Use Cyberduck to establish SFTP connection management files (consider FileZilla)
3. Resources:
   How to use SSH to connect to a remote Linux server under Mac (including Cyberduck download)
   Use your own terminal SSH feature under Mac

Setting up the environment-virtualenv

1. Set up a virtual environment and install packages (you can also consider Anaconda)
   Setting up the Environment: virtualenv xxx_pyvirtualenv -p python3 xxx_py
   Enter the environment:source xxx_py/bin/activate
   Exit:deactivate
2. Use Tsinghua Mirror

• Temporary use
  pip install -i https://pypi.tuna.tsinghua.edu.cn/simple some-package
• Set as Default
  pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple

1. Resources:
   Tsinghua PyPI Mirror Use Help
   VIRTUALENV Introduction and basic use
   One of the essential artifacts of Python development: virtualenv
   virtualenv-Liaoche's official website

Let the TensorFlow code run on the GPU

1. GPU Occupancy Issues
   TensorFlow may occupy all GPU resources visible from sight

• To view GPU usage:gpustat
• In the Python code, add:
  os.environ[‘CUDA_VISIBLE_DEVICES‘] = ‘0‘ os.environ[‘CUDA_VISIBLE_DEVICES‘] = ‘0,1‘
• To set up using a fixed GPU:
  CUDA_VISIBLE_DEVICES=1 Only device 1 will be seen  CUDA_VISIBLE_DEVICES=0,1 Devices 0 and 1 will be visible  CUDA_VISIBLE_DEVICES=”0,1” Same as above, quotation marks are optional  CUDA_VISIBLE_DEVICES=0,2,3 Devices 0, 2, 3 will be visible; device 1 is masked
  When you run the code
  CUDA_VISIBLE_DEVICES=0 python3 main.py
• TensorFlow provides two ways to control GPU resources:
  • Dynamic application of memory during operation, how much it takes to apply

  config = tf.ConfigProto()  config.gpu_options.allow_growth = True  

  • Limiting GPU usage

  gpu_options=tf.GPUOptions(per_process_gpu_memory_fraction=0.4)  config=tf.ConfigProto(gpu_options=gpu_options)  session = tf.Session(config=config)  

1. TensorFlow Code
   There is currently no consideration of whether the GPU or CPU is manually assigned to each part of the code
   So we with tf.device(self.device): wrapped up all the network structures.
   and use config = tf.ConfigProto(gpu_options=gpu_options,allow_soft_placement=True) it to get TensorFlow to assign it himself.

2. Resources:
   TensorFlow setting up GPU and GPU memory usage
   TensorFlow using the GPU
   TensorFlow GPU Small test

about using the lab server's GPU and running the TensorFlow code