14 days ago
-
Add SegFormer-B0 (IGN) checkpoint, config and metrics
Hi @gabrielkasmi — here is the SegFormer-B0 (IGN) checkpoint you asked for, trained as part of [deeppvmapper#20](https://github.com/gabrielkasmi/deeppvmapper/pull/20) (issue #11).
This PR adds a self-contained `segformer-b0/` folder:
| File | Description |
| --- | --- |
| `segformer_ign_best.pth` | Raw PyTorch `state_dict`, epoch 47 (3.71 M params) |
| `config.yaml` | Exact training config |
| `metrics.json` | Per-epoch history + best-epoch summary |
| `evaluate.py` | Reference-protocol evaluation (see below) |
| `README.md` | Training recipe and full results |
**Results — canonical test split (1239 IGN positives):**
| Protocol | IoU (PV) | Pixel F1 |
| --- | --- | --- |
| Reference (mean of per-batch metrics, batch 8, 416²) | **0.7356** | **0.8437** |
| Pixel-aggregated over the split | 0.7445 | 0.8535 |
Instance F1 0.8549, pixel precision 0.8705, recall 0.8372 (aggregated).
**Two practical notes:**
1. `train/evaluate.py` builds `deeplabv3_resnet101` and loads a raw DeepLab sta…
-
Update train/train_seg.py
-
Update README.md
-
Upload 9 files
add source training code. Weights landing in `weights` are in the root of the folder.
-
Update README.md
-
Update README.md
-
Update README.md
-
Update README.md
-
Upload model.py with huggingface_hub
-
Update README.md
-
Update README.md
-
Upload model.py with huggingface_hub
-
Delete .DS_Store
-
Upload folder using huggingface_hub
-
initial commit
