Update README.md
Browse files
README.md
CHANGED
|
@@ -51,45 +51,59 @@ python train_kaggle.py \
|
|
| 51 |
--clean-episode-ratio 0.80 \
|
| 52 |
--event-spatial-correlation 0.50 \
|
| 53 |
--device auto
|
|
|
|
| 54 |
|
| 55 |
-
Requires ZoneEquivariantMaskablePolicy from
|
| 56 |
|
| 57 |
-
Load
|
|
|
|
|
|
|
| 58 |
from sb3_contrib import MaskablePPO
|
| 59 |
|
| 60 |
model = MaskablePPO.load("run_nz2_triage/final_model.zip")
|
| 61 |
action, _ = model.predict(obs, action_masks=masks, deterministic=True)
|
|
|
|
|
|
|
|
|
|
| 62 |
|
| 63 |
-
|
| 64 |
python mnn_export.py \
|
| 65 |
--checkpoint run_nz2_triage/final_model.zip \
|
| 66 |
--output weather_rl_model.mnn \
|
| 67 |
--quantize int8 \
|
| 68 |
--n-zones 2
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 69 |
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
Research stack only: no on-call, feed SLA, or regulatory certification.
|
| 81 |
-
Training is largely synthetic for the moment.
|
| 82 |
-
Full-tour belief or episode-length metrics are not zone-selection skill.
|
| 83 |
-
MNN/C++ path is export-oriented, not a validated production edge deployment.
|
| 84 |
-
|
| 85 |
-
Dependencies
|
| 86 |
torch>=2.0
|
| 87 |
gymnasium>=0.29
|
| 88 |
stable-baselines3>=2.0
|
| 89 |
sb3-contrib>=2.0
|
| 90 |
numpy>=1.24
|
|
|
|
| 91 |
|
| 92 |
While GPU is recommended for training, CPUs can be used successfully.
|
| 93 |
|
| 94 |
-
License
|
|
|
|
| 95 |
AGPL-3.0. See repository for full terms.
|
|
|
|
| 51 |
--clean-episode-ratio 0.80 \
|
| 52 |
--event-spatial-correlation 0.50 \
|
| 53 |
--device auto
|
| 54 |
+
```
|
| 55 |
|
| 56 |
+
Requires `ZoneEquivariantMaskablePolicy` from `gru_weather_policy.py` (not the string `"MultiInputPolicy"`).
|
| 57 |
|
| 58 |
+
### Load
|
| 59 |
+
|
| 60 |
+
```python
|
| 61 |
from sb3_contrib import MaskablePPO
|
| 62 |
|
| 63 |
model = MaskablePPO.load("run_nz2_triage/final_model.zip")
|
| 64 |
action, _ = model.predict(obs, action_masks=masks, deterministic=True)
|
| 65 |
+
```
|
| 66 |
+
|
| 67 |
+
### Edge export
|
| 68 |
|
| 69 |
+
```bash
|
| 70 |
python mnn_export.py \
|
| 71 |
--checkpoint run_nz2_triage/final_model.zip \
|
| 72 |
--output weather_rl_model.mnn \
|
| 73 |
--quantize int8 \
|
| 74 |
--n-zones 2
|
| 75 |
+
```
|
| 76 |
+
|
| 77 |
+
## Validation status
|
| 78 |
+
|
| 79 |
+
| Claim | Status |
|
| 80 |
+
|-------|--------|
|
| 81 |
+
| Weight transfer across `n_zones` changes | Supported by design |
|
| 82 |
+
| Zone-order shuffle avoids fixed-slot lock-in | Observed |
|
| 83 |
+
| Risk-conditioned zone choice under triage (synthetic, single-dirty) | Observed (~mid-80s% accuracy; holds when evaluated at higher spatial correlation than used in training) |
|
| 84 |
+
| Product skill vs curated L1 labels | Scorer/product path available; L1 density limits apply |
|
| 85 |
+
| Multi-seed replication of the triage recipe | Open |
|
| 86 |
+
| Held-out real product / emit metrics | Open |
|
| 87 |
|
| 88 |
+
## Limitations
|
| 89 |
+
|
| 90 |
+
- Research stack only: no on-call, feed SLA, or regulatory certification.
|
| 91 |
+
- Training is largely synthetic for the moment.
|
| 92 |
+
- Full-tour belief or episode-length metrics are not zone-selection skill.
|
| 93 |
+
- MNN/C++ path is export-oriented, not a validated production edge deployment.
|
| 94 |
+
|
| 95 |
+
## Dependencies
|
| 96 |
+
|
| 97 |
+
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 98 |
torch>=2.0
|
| 99 |
gymnasium>=0.29
|
| 100 |
stable-baselines3>=2.0
|
| 101 |
sb3-contrib>=2.0
|
| 102 |
numpy>=1.24
|
| 103 |
+
```
|
| 104 |
|
| 105 |
While GPU is recommended for training, CPUs can be used successfully.
|
| 106 |
|
| 107 |
+
## License
|
| 108 |
+
|
| 109 |
AGPL-3.0. See repository for full terms.
|