Build upon previous training

Description

Develop a feature in which you can retrain using previous knowledge, that is, for instance:

Imagine we train an AutoML process with `max_runtime_secs=3000` and we are still unhappy about the results. We would like to continue training, but we would like to train using this previous training, not starting all over again.

We would like this feature because we are researching about a platform that includes something like "MLaaS", and verticals would pay for a limited CPU/RAM/GPU consumption. So this means this parameter will always be set, due to financial restrictions. So if after finalizing, a vertical's desire is to continue training to improve the AUC or minimize the losses, we should give them the option to retrain, building upon the last training, thus avoiding "extra" charges for the same repeated training. We know we can change the "seed" option, but that would not be sufficient.

Assignee

Unassigned

Fix versions

None

Reporter

Sergio Fernández

Support ticket URL

None

Labels

None

Affected Spark version

None

Customer Request Type

None

Task progress

None

ReleaseNotesHidden

None

CustomerVisible

No

Components

Priority

Minor
Configure