CODEWITHSUNDEEP
google-cloud-ml
anon_swe
anon_swe

Reputation: 9345

Google Cloud ML Engine: Hyperparameter Tuning Cost as Function of MaxParallelTrials

Quick question about hyperparameter tuning on ML Engine. Assuming I'm setting maxParallelTrials to a value less than the number of cores on the ML Engine instance I'm using, is there any difference in cost between:

maxTrials: 20 maxParallelTrials: 1

and

maxTrials: 80 maxParallelTrials: 4

Since they should take basically the same amount of time? Or is the number of cores being used a factor in calculating ML Units Consumed?

Thanks!

Upvotes: 0

Views: 100

Answers (1)

ch_mike
ch_mike

Reputation: 1576

There shouldn’t be any differences if they take the same amount of time (perhaps there could be a little overhead).

The number of cores used is not a factor. ML Units are calculated based on the region and the scale tier you choose.

Training pricing is calculated according to this formula:

(training units * base price / 60) * job duration in minutes

Where training units depends on the scale tier and region you select and base price depends on the region. For example, training in US with a PREMIUM_1 scale tier, training units would be 49.323 and base price 0.49 USD. You can find the complete pricing table here.

Upvotes: 1

Related Questions