Azure GPU VM Setup

Scripts to set up Azure GPU VM that install all prerequisites for latest Tensorflow on Ubuntu 18.04:

• CUDA 10 drivers
• CUDNN 7.4.1.5
• TensorRT
• pip
• virtualenv

Optional:

• Docker CE
• nvidia-docker 2.0

These scripts are not needed for EC2-AWS using the Deep Learning Base AMI, however instructions for virtual environment and Docker might still be helpful.

Instructions

In azure-vm-setup, run:

• bash cuda_setup_part1.sh to install NVIDIA drivers. This will take about 10 minutes and there will be a reboot.
• Login to the VM again and run nvidia-smi to make sure CUDA driver is installed
• bash cuda_setup_part2.sh to install CU-DNN. This will take about 20 minutes and there will be a reboot.
• Login to the VM again and run nvcc --version. Should show up as CUDA 10.1.
• Run bash test_cuda.sh to run CUDNN tests. Should show up as successful.

Run pip list --format=columns to check all base packages are there.

Run virtualenv --help to check if virtualenv is there.

STOP HERE. DO NOT PIP INSTALL ANYTHING ELSE

Attach a Disk

• Follow this instruction for AWS-EC2
• Follow this instruction for Azure

Virtual Environment Setup

Steps:

• mkdir /datadrive/envs
• virtualenv --python=python3.6 /datadrive/envs/<your_env_name>.
  • Default to empty Python 3 environment, check in /usr/bin for which Python versions are available. The virtual environment is stored on /datadrive/envs directory to save OS disk space.
  • Add --system-site-packages to inherit all the base packages that are pre-installed in the VM, however not recommended.
• To activate your environment, do source /datadrive/envs/<your_env_name>/bin/activate
  • If fancy, add this line to .bashrc : alias <your_env_name>='source /datadrive/envs/<your_env_name>/bin/activate'. Then source ~/.bashrc to take effect
  • After that, only run your-env-name to activate the environment and begin work
• To test Tensorflow-GPU, inside your virtual environment, run:
  • pip install --upgrade tensorflow-gpu. Do not pip install tensorflow as to not confuse with CPU version
  • In python, run import tensorflow as tf and tf.test.gpu_device_name(). If CUDA device shows up, congratulations!

AGAIN, DO NOT PIP INSTALL ANYTHING OUTSIDE VIRTUAL ENVIRONMENT. If you think your virtual environment is messed up, delete the folder /datadrive/envs/your_env_name and repeat the process.

Jupyter Notebook Setup

1. Open port for public access. Follow this

2. Jupyter server setup:

• Inside your virtual environment, run pip install jupyter.
• Generate a default Jupyter config in home directory by jupyter notebook --generate-config
• Run jupyter notebook password and enter your password
• Run cat /home/your-user-name/.jupyter/jupyter_notebook_config.json and copy the password starting with 'sha1:...' somewhere else
• Open the config file with nano /home/your-user-name/.jupyter/jupyter_notebook_config.py and enter the following lines
  • Change your-user-name, your-sha1-password that you copied somewhere else, and desired-port which is whichever port you opened from step 1)

c = get_config()

# Create your own password as indicated above
c.NotebookApp.password = u'your-sha1-password'

# Network and browser details. We use a fixed port (9999) so it matches
# our Azure setup, where we've allowed traffic on that port
c.NotebookApp.ip = 0.0.0.0
c.NotebookApp.port = desired-port
c.NotebookApp.open_browser = False

Docker Setup

Inside azure-vm-setup directory:

• Run bash docker.sh to install Docker CE
• Run bash nvidia_docker.sh to install Nvidia-Docker

Source:

• https://www.tensorflow.org/install/gpu#ubuntu_1804_cuda_10
• https://github.com/NVIDIA/nvidia-docker/wiki/Installation-(version-2.0)