Follow

Using GPUs with Deep Learning Libraries on Domino

GPUs can speed up training of deep learning models. Many deep learning libraries like Neon, Theano, and Tensorflow are compatible with GPUs. 

In order to use these deep learning libraries with GPUs in Domino, you will need to 
  1. Select a hardware tier that has a GPU
  2. Select an environment that has the proper software installed to take advantage of the GPU.
 
Selecting the Hardware Tier
 
For users in Domino's hosted environment: There are 3 hardware tiers with GPU capabilities: GPU [g2.2xlarge], GPU Large [g2.8xlarge], and P2 8XLarge GPU [p2.8xlarge]. For more information about each hardware tier, refer to the AWS page on EC2 instance types. To learn more about how to select your hardware tier, refer to our help article.
 
For VPC and on-premise customers: Please check with your administrators on the availability of hardware with GPU capabilities.
 
Selecting a GPU capable environment 
 
For users in Domino's hosted environment: Domino provides a globally available environment called [Domino Managed] GPU Tools. This environment comes with GPU enabled Tensorflow, Theano, and Neon. To select the environment, navigate to Settings page of your project, click on the dropdown menu in the Compute Environments section, and select GPU Tools.
 
 
For VPC or on-premises customers: Contact your Domino administrator to request a global environment with GPU enabled package.
If you would like to create your own custom environments with GPU enabled packages, follow the example link for the specific library below. You will find commands to use in your custom environment in the respective Overview sections.
 
Examples
 
Below are links to example Domino projects for that use GPU enabled environments and GPU machines to train LSTM models for sentiment analysis on the IMDB dataset. You will be able to find references and the required commands to build an environment in the Overview sections for each project. Each project also includes a script to execute the training (lstm.py) and a Jupyter notebook, which shows how to call the trained model and do a prediction.
Was this article helpful?
0 out of 0 found this helpful

Comments