add comp requirements to model card
Browse files
README.md
CHANGED
|
@@ -1,13 +1,13 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: cc-by-4.0
|
| 3 |
-
tags:
|
| 4 |
-
- weather
|
| 5 |
-
- numerical-weather-prediction
|
| 6 |
-
- gnn
|
| 7 |
-
- autoregressive
|
| 8 |
-
- medium-range-forecast
|
| 9 |
-
pipeline_tag: graph-ml
|
| 10 |
-
---
|
| 11 |
|
| 12 |
# Model Card for FastNet-v1.1
|
| 13 |
|
|
@@ -89,10 +89,25 @@ Weighting was applied per-variable and was proportional to pressure level.
|
|
| 89 |
#### Preprocessing
|
| 90 |
Training data from each pressure level (including surface-level variables) are separately standardised to have zero mean and unit standard deviation, over latitude, longitude and time. Orography is rescaled to the unit interval, and land-sea mask, solar hour angle, time of year, latitude, and longitude require no change as they are already suitably normalised.
|
| 91 |
|
| 92 |
-
### Speed, Sizes, Times
|
| 93 |
-
|
| 94 |
## Evaluation
|
| 95 |
|
| 96 |
FastNet is evaluated against ERA5 data using the WeatherBench 2 software package analysis for 2022. We compute the full rollout for all forecast valid times in 2022, re-gridding the output to a 1.5 degree latitude-longitude grid using a conservative re-gridding scheme.
|
| 97 |
|
| 98 |
-
We also evaluate against the Met Office Global Model operational
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-4.0
|
| 3 |
+
tags:
|
| 4 |
+
- weather
|
| 5 |
+
- numerical-weather-prediction
|
| 6 |
+
- gnn
|
| 7 |
+
- autoregressive
|
| 8 |
+
- medium-range-forecast
|
| 9 |
+
pipeline_tag: graph-ml
|
| 10 |
+
---
|
| 11 |
|
| 12 |
# Model Card for FastNet-v1.1
|
| 13 |
|
|
|
|
| 89 |
#### Preprocessing
|
| 90 |
Training data from each pressure level (including surface-level variables) are separately standardised to have zero mean and unit standard deviation, over latitude, longitude and time. Orography is rescaled to the unit interval, and land-sea mask, solar hour angle, time of year, latitude, and longitude require no change as they are already suitably normalised.
|
| 91 |
|
|
|
|
|
|
|
| 92 |
## Evaluation
|
| 93 |
|
| 94 |
FastNet is evaluated against ERA5 data using the WeatherBench 2 software package analysis for 2022. We compute the full rollout for all forecast valid times in 2022, re-gridding the output to a 1.5 degree latitude-longitude grid using a conservative re-gridding scheme.
|
| 95 |
|
| 96 |
+
We also evaluate against the Met Office Global Model operational
|
| 97 |
+
|
| 98 |
+
## Computational Requirements
|
| 99 |
+
|
| 100 |
+
### Hardware
|
| 101 |
+
For inference, FastNet-global runs on a single NVIDIA A100 GPU, though CPU inference is also supported. When using the GPU model file, we recommend a GPU with compute capability ≥ 8.0 (i.e. A100 or equivalent/higher).
|
| 102 |
+
|
| 103 |
+
### Software
|
| 104 |
+
|
| 105 |
+
Inference dependencies are intentionally lightweight. Key packages include:
|
| 106 |
+
- torch — runs the model (load + rollout)
|
| 107 |
+
- xarray / zarr / dask — geospatial and meteorological data handling
|
| 108 |
+
- pyyaml — parsing inference configuration yaml files
|
| 109 |
+
|
| 110 |
+
Training dependencies (not required for inference, listed here for attribution):
|
| 111 |
+
- torch-harmonics — spherical harmonic transforms
|
| 112 |
+
- lightning — ML training framework
|
| 113 |
+
- icosphere — icosahedral grid handling
|