Skip to content
This repository was archived by the owner on Jul 18, 2024. It is now read-only.
This repository was archived by the owner on Jul 18, 2024. It is now read-only.

Add early stopping to trainer(s) #69

@tkornuta-ibm

Description

@tkornuta-ibm

Generally it is already incorporated (as best_* checkpoint contain only models from moment in time when validation loss was smallest), but despite that the trainer continues until one of the other terminal conditions will occur.

Idea:

  • monitor "total loss" (add it to trainer statistics)
  • store "minimal_loss" value and "loss_down_counter" (indicating when last time loss went down)
  • introduce config variable "early_stop" (number of episodes/epochs to stop when loss is not going down) (DEFAULT: 10?)

Metadata

Metadata

Assignees

Labels

CoreIssues related to core of the frameworkWorkerIssues related to workers

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions