2

I'm running some projects with H2o AutoML using Sagemaker notebook instances, and I would like to know if H2o AutoML can benefit from a GPU Sagemaker instance, if so, how should I configure the notebook?

Marcel Mendes Reis
  • 107
  • 10
  • Can you share more about how you are using H2O AutoML and what version of H2O it is? – Erin LeDell Sep 05 '19 at 21:46
  • @ErinLeDell, for sure! I believe it's the 3.2 version of H2O AutoML. I work as a management consultant, as part of my job I use ML to predict the likelihood of events as well as predicting continuous variables using regression. Addiotanly, I've been using ML for forecasting. Specifically, I'm in a project which I need to predict the likelihood of an investment portfolio to fall in a window of time. In this project, I'll need to train a huge dataset, and then, I thought I could take advantage of GPU processing to reduce the training time. Usually, I use notebooks on Sagemaker. – Marcel Mendes Reis Sep 05 '19 at 23:26
  • You might want to double-check the version number, we are at 3.26 right now, so it's probably something close to that. – Erin LeDell Sep 05 '19 at 23:45
  • 1
    Yes, @ErinLeDell I realized it's the 3.26. – Marcel Mendes Reis Sep 05 '19 at 23:55

1 Answers1

1

H2O AutoML contains a handful of algorithms and one of them is XGBoost, which has been part of H2O AutoML since H2O version 3.22.0.1. XGBoost is the only GPU-capable algorithm inside of H2O AutoML, however, a lot of the models that are trained in AutoML are XGBoost models, so it still can be useful to utilize a GPU. Keep in mind that you must use H2O 3.22 or above to use this feature.

My suggestion is to test it on a GPU-enabled instance and compare the results to a non-GPU instance and see if it's worth the extra cost.

Erin LeDell
  • 8,704
  • 1
  • 19
  • 35
  • Thanks, @ErinLeDell. Do I need any additional set-up to enable de GPU instance to work with H2O AutoML? Like, install some libraries or anything else? – Marcel Mendes Reis Sep 06 '19 at 00:00
  • You need to check that you have Nvidia GPUs and CUDA 8 installed. You can monitor the GPU usage during the AutoML run via the command line using the `nvidia-smi` command. – Erin LeDell Sep 06 '19 at 07:13
  • Thanks, @ErinLeDell, another doubt! To get both installed, can I only install CUDA Toolkit? I saw on Cuda's website that the most recent version of CUDA is 10.1.243, could I install that version(10) or I need specifically to install the version 8? – Marcel Mendes Reis Sep 06 '19 at 13:49
  • @Erin LeDell, can the H2O AutoML XGBoost utilize TPU in Google Colab? – Md Hishamur Rahman Jul 03 '20 at 14:26
  • 1
    @Md.HishamurRahman No, there's no TPU support, just GPU. – Erin LeDell Jul 17 '20 at 17:39