Machine Learning Designer is too slow!

Question

I am teaching ML next week. I have usually been using ML Studio (Classic), but thought it would be time to transfer to ML the ML Designer. It is great in many ways, but it is just too slow for demo and teaching purposes.

1000 rows from an Azure DB in the same region. The simplest of experiments (Select Columns, Split Rows, Train Boosted Decission Tree, Score and Evaluate) on a STANDARD_DS5_V2 (16 Cores, 56 GB RAM, 112 GB Disk) compute target, which was the most expensive I could choose on my VS subscription. Creating a model takes about eight minutes!

Or to phrase it as a question: Am I doing anything wrong? Is there any way to make the Designer usable?

Thanks,
Christian

Answer 1

@CVinter The classic studio used a fixed environment or compute which was always available so it was faster to execute simple tasks but was limited in many ways to the applications that could be run on the compute as there was no flexibility to select the type of compute to perform advanced tasks with a customer's existing compute.

Designer on the other hand provides the ability to select the required compute type but currently there is a limitation of setting up this compute for the job and hence the time to complete the overall job seems to be higher but for scenarios where heavy compute processing is required the overall time to run the job is significantly reduced. Our team is addressing this scenario to reduce the overhead time to setup this compute which should eventually reduce the time to run simple tasks on designer. One of our team member from Azure ML Designer provided more details on a similar question in this thread. I hope this helps. @Lu Zhang (AI) @Lu Zhang (MSFT) FYI.