Update README.md
Browse files
README.md
CHANGED
|
@@ -58,6 +58,8 @@ Rules:
|
|
| 58 |
- The spatial holdout by department must not be modified. Re-splitting invalidates comparability with published results.
|
| 59 |
- For Track 3, only the Google training split may be used for training.
|
| 60 |
|
|
|
|
|
|
|
| 61 |
---
|
| 62 |
## Results
|
| 63 |
|
|
@@ -112,12 +114,28 @@ Models trained on the official BDAPPV splits using:
|
|
| 112 |
|
| 113 |
- **Optimizer:** AdamW (lr=1e-4, weight_decay=1e-4)
|
| 114 |
- **Scheduler:** Cosine annealing
|
| 115 |
-
- **Effective batch size:** 32 (
|
| 116 |
- **Early stopping:** patience=7 epochs on validation metric
|
| 117 |
- **Input size:** 400×400 px
|
| 118 |
- **Initialization:** checkpoints from [Mayer et al. (2022)](https://doi.org/10.1016/j.apenergy.2021.118469), who fine-tuned DeepLabV3-ResNet101 and InceptionV3 on 10 cm/px orthoimagery from North Rhine-Westphalia (Germany) for rooftop PV detection. These checkpoints were then further fine-tuned on BDAPPV using the splits above.
|
| 119 |
|
| 120 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 121 |
|
| 122 |
---
|
| 123 |
|
|
|
|
| 58 |
- The spatial holdout by department must not be modified. Re-splitting invalidates comparability with published results.
|
| 59 |
- For Track 3, only the Google training split may be used for training.
|
| 60 |
|
| 61 |
+
*IoU/F1 above are foreground (PV-class) only — see [`train/evaluate.py`](train/evaluate.py#L31) for the exact computation.*
|
| 62 |
+
|
| 63 |
---
|
| 64 |
## Results
|
| 65 |
|
|
|
|
| 114 |
|
| 115 |
- **Optimizer:** AdamW (lr=1e-4, weight_decay=1e-4)
|
| 116 |
- **Scheduler:** Cosine annealing
|
| 117 |
+
- **Effective batch size:** 32 (batch_size × grad_accum — see each script's defaults)
|
| 118 |
- **Early stopping:** patience=7 epochs on validation metric
|
| 119 |
- **Input size:** 400×400 px
|
| 120 |
- **Initialization:** checkpoints from [Mayer et al. (2022)](https://doi.org/10.1016/j.apenergy.2021.118469), who fine-tuned DeepLabV3-ResNet101 and InceptionV3 on 10 cm/px orthoimagery from North Rhine-Westphalia (Germany) for rooftop PV detection. These checkpoints were then further fine-tuned on BDAPPV using the splits above.
|
| 121 |
|
| 122 |
+
|
| 123 |
+
Training scripts are included in this repo, under [`train/`](train):
|
| 124 |
+
|
| 125 |
+
| File | What it does |
|
| 126 |
+
|---|---|
|
| 127 |
+
| `train_seg.py` | Trains DeepLabV3-ResNet101 (segmentation) |
|
| 128 |
+
| `train_clf.py` | Trains InceptionV3 (classification) |
|
| 129 |
+
| `evaluate.py` | Computes the IoU/F1 numbers reported below (foreground-only, on the official test split) |
|
| 130 |
+
| `dataset.py` | Dataset loaders for both tasks |
|
| 131 |
+
| `pretrained/` | Base checkpoints from Mayer et al. (2022) — starting point before BDAPPV fine-tuning |
|
| 132 |
+
|
| 133 |
+
Reproduce (segmentation, IGN):
|
| 134 |
+
\`\`\`bash
|
| 135 |
+
pip install -r train/requirements.txt
|
| 136 |
+
python train/train_seg.py --provider ign --checkpoint train/pretrained/deeplabv3_weights.tar
|
| 137 |
+
python train/evaluate.py seg --provider ign --checkpoint train/weights/deeplab_ign_best.pth
|
| 138 |
+
\`\`\`
|
| 139 |
|
| 140 |
---
|
| 141 |
|