Clarify California scope and add CONUS retraining recipe
Browse files
models/wildfire_fm/README.md
CHANGED
|
@@ -13,6 +13,13 @@ models/wildfire_fm/checkpoints/seed_123/best_firms_prauc.pt
|
|
| 13 |
|
| 14 |
Use `modeling_unet.py` to instantiate the compact U-Net architecture before loading a checkpoint state dict.
|
| 15 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 16 |
## Input channels
|
| 17 |
|
| 18 |
The checkpoints expect a fixed 16-channel tensor in `[channel, y, x]` order.
|
|
|
|
| 13 |
|
| 14 |
Use `modeling_unet.py` to instantiate the compact U-Net architecture before loading a checkpoint state dict.
|
| 15 |
|
| 16 |
+
## Training scope
|
| 17 |
+
|
| 18 |
+
The released checkpoints are California regional weights trained for the 5 km
|
| 19 |
+
EPSG:5070 California grid. They should not be described as nationwide-trained
|
| 20 |
+
weights. A machine-readable scope file is stored in `training_scope.json`, and a
|
| 21 |
+
CONUS retraining recipe is provided in `../../training/NATIONWIDE_RETRAINING.md`.
|
| 22 |
+
|
| 23 |
## Input channels
|
| 24 |
|
| 25 |
The checkpoints expect a fixed 16-channel tensor in `[channel, y, x]` order.
|