README.md

---
license: other
license_name: ANRF Open License
license_link: https://anrfonline.in/ANRF/AbstractFilePath?FileType=E&FileName=OL_AISE.pdf&PathKey=DOCUMENT_TEMPLATE
tags:
  - segmentation
  - remote-sensing
  - flood
  - pytorch
---

# ANRF AISEHack Phase 2 — ensemble artifacts

## Checkpoints (bucket root)

These are the **weights** behind the five-model majority-vote ensemble (~0.227 public LB).

| File | Model |
|------|--------|
| `siamese_efficientnet_b4_epoch45.ckpt` | Dual-branch Siamese U-Net, EfficientNet-B4 |
| `kd_transductive_epoch35.ckpt` | Same architecture, knowledge-distilled |
| `siamese_efficientnet_b7_epoch35.ckpt` | Siamese U-Net, EfficientNet-B7 |
| `baseline_21ch_swa.pth` | 21-channel U-Net (SWA weights) |
| `convnext_large_3class_best.pth` | ConvNeXt-Large 3-class U-Net |

## Prediction CSVs (`ensemble_five_csv/`)

Folder **`ensemble_five_csv/`** in this bucket holds the five **submission CSVs** (RLE masks) for those models, with the preferred filenames (`01_…` through `05_…`). They match the stack used for the public LB submission. Sync with:

`hf sync hf://buckets/itikelabhaskar/AISE_PHASE2/ensemble_five_csv ./local_folder`

**License:** ANRF Open License — see `LICENSE_ANRF.txt` in this repo and the competition PDF.

**Code:** [GitHub — `anrf_flood_phase2_submission`](https://github.com/itikelabhaskar/AISE_PHASE2) (notebooks + `config/models_manifest.json`).

**Inference:** Training code for re-running the models lives outside this bundle (e.g. `phase2_final/experiments/` in the author’s codebase). The release notebooks **merge precomputed CSVs**; full re-inference needs matching SAR/aux preprocessing.