unet_segmentation: raise the epoch budget and add early stopping - #60
Open
nilshof01 wants to merge 1 commit into
Open
unet_segmentation: raise the epoch budget and add early stopping#60nilshof01 wants to merge 1 commit into
nilshof01 wants to merge 1 commit into
Conversation
The pipeline trained a fixed 5 epochs with no early stopping and no best-epoch restore (the DINOv3 pipeline has both). Measured at 32 training chips (6 regions x 5 folds), the best epoch fell in the 19-30 range in every region; the best val F1 reachable within 5 epochs vs 30 epochs + early stopping: Banepa 48.2->69.1, Jakarta 54.8->61.6, Accra 37.1->48.3, Bogota 15.0->23.1, Manila 61.9->67.2, Nairobi 40.3->51.7. epochs default 5 -> 30; new early_stop_patience (default 5, 0 disables) monitoring validation loss with best-epoch weight restore. Compute cost stays bounded by early stopping, and warm starts that converge early (e.g. the building-pretrained base) stop well before the cap. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The pipeline trained a fixed 5 epochs with no early stopping and no best-epoch restore (the DINOv3 pipeline has both). Measured at 32 training chips (6 regions x 5 folds), the best epoch fell in the 19-30 range in every region; the best val F1 reachable within 5 epochs vs 30 epochs + early stopping: Banepa 48.2->69.1, Jakarta 54.8->61.6, Accra 37.1->48.3, Bogota 15.0->23.1, Manila 61.9->67.2, Nairobi 40.3->51.7.
epochs default 5 -> 30; new early_stop_patience (default 5, 0 disables) monitoring validation loss with best-epoch weight restore. Compute cost stays bounded by early stopping, and warm starts that converge early (e.g. the building-pretrained base) stop well before the cap.