Adds pixel-loss to the decoder#502
Closed
gabrieltseng wants to merge 47 commits intomainfrom
Closed
Conversation
…to gabi/exp17-ndvi-supervision # Conflicts: # scripts/vnext/single_bandset_band_dropout/single_bandset_masked_neg.py
… which suggests its too easy and anyway we saw improved performance from reducing not increasing this weight
Collaborator
Author
|
Closing in favor of #528 |
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
This PR builds on #501
Here, we add a pixel loss to the decoder.
Q: How is this different from the MAE loss?
A: MAE loss applies the same loss (e.g. L2 loss) to all modalities. This supervision loss is modality aware, and applies a classification or regression loss as appropriate.
The hypothesis with this approach was twofold: firstly, this would help maintain high resolution at higher patch sizes, and secondly that this might help with stability too.
Here, the time masking seems to really help, providing a significant boost to both pastis and so2sat.
Since the instance-contrastive loss was in part designed to add stability to the model, I wondered if we could do away with it and only have a single forward pass per batch. This is in orange below:
This seems to help a lot with MADOS (>6% improvement!) but hurts quite a bit on pastis.