This article presents instructions on how to launch a Virtual Machine (VM) from the Deep Learning Supercloud Workspace Image (SWI). For information about SWIs see here. To understand what is included in the Deep Learning SWI see this link.
The instructions in this article assume you have access to the SuperCloud and SWI catalogue.
Step 1: Launch and connect to a Deep Learning VM
A detailed instruction on how to create and launch VMs on the Supercloud is provided here. To use the deep learning capabilities provided by Firmus, you need to launch an accelerated Flavour, e.g., accelerated.small.i using the deep learning SWI as a Source, i.e., supercloud- deep-learning. See below figures for more details. Feel free to increase the Volume Size or leave as is.
Once you launch the VM and establish a SSH connection, you will see the welcome page as below:
Step 2: Activate the Deep Learning Environment
The Deep Learning capabilities are provided through Conda virtual environments. To get a list of available environments, enter:
conda env list
Based on your chosen WSI, you’ll get a couple of options. The Deep Learning SWI currently supports TensorFlow workloads.
To activate a Conda environment, try:
source activate tensorflow27_py39
The command line would now appear as:
To get a detailed list of packages inside the environment, try:
You can add/remove/update packages as you would normally do in a conda environment, e.g.,
conda install NEWPACKAGE
Step 3: Sanity check
In order to check the availability of TensorFlow and the visibility of the GPU devices, you can run the following script from your commandline. It is a simple script running benchmarks on opensource TensorFlow based models. Store this script as test_setup.py and execute the code by typing python test_setup.py all. The script measures the performance of the setup based on the execution time on training (training as an argument), inferencing (inference as an argument), or both (all as an argument). Completion of this program confirms the heath of the deep learning Conda environment.