none
Best compute to use for computationally intensive ML RRS feed

  • Question

  • Hi There, 

    I've been trying to implement a fairly complex ML time series projection using the new ML Designer.   (For simplicity, I thought I would try it in the ML Designer rather than struggling through a Jupyter notebook (which i am not very familiar with)).

    It turned out that the ML model I was wanting to use was not offered as one of the OOB algorithms in the designer, so I ended up creating a Python script task within the designer that implemented my ML model using a seasonal ARIMA model.

    The script ends up calling ARMIA several thousand times as part of the hyper-parameter selection, and that is taking nearly 35 minutes to execute each time I want to run the experiment....  In the classic designer, it only took 1 minute to execute the same code, so I am pretty disappointed with the performance in the new designer thus far.

    I tried creating a couple of different compute targets (A DS15-V2 with 20 CPUs and lots of memory for example), but it didn't seem to make a scrap of difference to the execution time.

    Does anyone have suggestions as to the proper compute target to run a computationally intensive Python script in through ML Designer?

    Maybe I should just bite the bullet and implement this in a notebook instead?... though I fear that I will still run into the issue of long execution times no matter what.

    Hoping the community can offer some advice here.

    Cheers,

    -Sheldon

      

     

    Sunday, January 12, 2020 2:07 PM

Answers

  • Unfortunately I can't share this one, as the dataset is from our companies production environment.

    I subsequently was in contact with one of the MS Cloud Advocates and was told that performance is generally an issue right now with this preview version, and the product team is working hard to address it.

    So till then we just have to be patient :) (and find things to do while our experiments run ;) )

    • Marked as answer by slyttle Monday, January 13, 2020 8:21 PM
    Monday, January 13, 2020 8:20 PM

All replies

  • Hi,

    Can you please add your project to the Azure AI gallery(https://gallery.azure.ai) and share the link to check?

    Thanks

    • Marked as answer by slyttle Monday, January 13, 2020 8:17 PM
    • Unmarked as answer by slyttle Monday, January 13, 2020 8:21 PM
    Monday, January 13, 2020 8:05 AM
    Moderator
  • Unfortunately I can't share this one, as the dataset is from our companies production environment.

    I subsequently was in contact with one of the MS Cloud Advocates and was told that performance is generally an issue right now with this preview version, and the product team is working hard to address it.

    So till then we just have to be patient :) (and find things to do while our experiments run ;) )

    • Marked as answer by slyttle Monday, January 13, 2020 8:21 PM
    Monday, January 13, 2020 8:20 PM