Semester 5 undergraduate research project. Core idea: supervise a SegFormer-B0's
internal attention against the ground-truth forest mask during training (an
Attention Consistency Loss), using an adaptation of Gradient-weighted
Attention Rollout for SegFormer's spatial-reduction attention, and a new
quantitative interpretability metric (AAMO). Benchmarked against U-Net and
vanilla SegFormer-B0 baselines. Full motivation, methodology, and math are in
Research_Proposal_extended.pdf; see also
Research_Proposal .pdf for the original version.
See CONTRIBUTING.md for the repo's folder/naming/documentation conventions before adding new work.
| # | Name | Role |
|---|---|---|
| 1 | Chanupa | Data & Baselines |
| 2 | Kalana | Transformer Lead |
| 3 | Dinura | Loss & Training |
| 4 | Lasana | Evaluation Lead |
| 5 | Dhinanjaya | Writing & Technical Integration Lead |
One folder per person, Phase1/<Name>-Person<N>[_Role]/, each with its own
README.md describing what's inside. Start there.
| Folder | Owner | Contents |
|---|---|---|
Phase1/Chanupa-Person1/ |
Person 1 | U-Net baseline notebook + full-dataset results |
Phase1/Kalana-Person2/ |
Person 2 | SegFormer-B0 training, dataset images/masks, checkpoints |
Phase1/Dinura-Person3/ |
Person 3 | Attention hooks, adapted Grad-Rollout, Attention Consistency Loss, math formulation, tests |
Phase1/Lasana-Person4_Evaluation/ |
Person 4 | Shared eval package: metrics, efficiency, AAMO, ablation runner, comparison tables |
Phase1/Dhinanjaya-Person5/ |
Person 5 (you) | Writing (related work, paper drafts), integration notes |
shared/— shared Weights & Biases experiment-tracking config, opt-in for all four technical members. Seeshared/README.md.
Literature_Review_Papers/— source PDFs cited in the proposal's related work.Research_Proposal_extended.pdf,Research_Proposal .pdf— the proposal documents.
These predate the Phase1/<Name>-PersonN/ convention and are not where
current work happens — don't build on them, but don't delete them either,
they contain early results/notes that informed the Phase 1 folders above.
Dinura/— early dev iterations (dev0,dev1) andproject1-Do-not-use/(explicitly marked unused by its own README).Lasana/— pre-Phase1 U-Net training + ViT XAI postprocessing research; superseded byPhase1/Lasana-Person4_Evaluation/.ViT_XAI_Segmentation/— early standalone ViT explainability experiment (pseudo-concepts, not the SegFormer/AAMO pipeline used now).
Phase 2 (Weeks 5–12: full ablation study + full paper) will follow the same
Phase2/<Name>-Person<N>/ convention once it starts. See
CONTRIBUTING.md.