work in progress

README.md

@@ -4,5 +4,3 @@
 See https://microsoft.github.io/aurora/intro.html
 
 See https://github.com/microsoft/aurora
-
-

src/ai_models_aurora/model.py

@@ -7,25 +7,119 @@
 
 import logging
 
-from anemoi.inference.plugin import AIModelPlugin
-from aurora import Aurora, AuroraSmall, rollout
+import numpy as np
+import torch
+from ai_models.model import Model
+from aurora import Aurora
+from aurora import Batch
+from aurora import Metadata
+from aurora import rollout
 
 LOG = logging.getLogger(__name__)
 
 
-class AuroraModel(AIModelPlugin):
-    expver = "auro"
+class AuroraModel(Model):
+
+    # Input
+    area = [90, 0, -90, 360 - 0.25]
+    grid = [0.25, 0.25]
+
+    surf_vars = ("2t", "10u", "10v", "msl")
+    static_vars = ("lsm", "z", "slt")
+    atmos_vars = ("z", "u", "v", "t", "q")
+    levels = (1000, 925, 850, 700, 600, 500, 400, 300, 250, 200, 150, 100, 50)
 
+    lagged = (-6, 0)
 
-    # Download
-    download_files = ["checkpoint.ckpt"]
+    #  For the MARS requets
+    param_sfc = surf_vars + static_vars
+    param_level_pl = (atmos_vars, levels)
+
+    # Output
+
+    expver = "auro"
 
     def __init__(self, **kwargs):
         super().__init__(**kwargs)
-        self.model =Aurora()
+        self.model = Aurora()
 
     def run(self):
-        model = AuroraSmall()
-        model.load_checkpoint("microsoft/aurora", "aurora-0.25-small-pretrained.ckpt")
+        LOG.info("Running Aurora model")
+        model = Aurora(use_lora=False)  # Model is not fine-tuned.
+        model = model.to(self.device)
+        LOG.info("Downloading Aurora model")
+        # TODO: control location of cache
+        model.load_checkpoint("microsoft/aurora", "aurora-0.25-pretrained.ckpt")
+        LOG.info("Loading Aurora model to device %s", self.device)
+
+        model = model.to(self.device)
+        model.eval()
+
+        fields_pl = self.fields_pl
+        fields_sfc = self.fields_sfc
+
+        N, W, S, E = self.area
+        WE, NS = self.grid
+        Nj = round((N - S) / NS) + 1
+        Ni  = round((E - W) / WE) + 1
+
+        to_numpy_kwargs = dict(dtype=np.float32)
+
+
+
+
+        # Shape (Batch, Time, Lat, Lon)
+        surf_vars = {}
+
+        for k in self.surf_vars:
+            f = fields_sfc.sel(param=k).order_by(valid_datetime="ascending")
+            f = f.to_numpy(**to_numpy_kwargs)
+            f = torch.from_numpy(f)
+            f = f.unsqueeze(0) # Add batch dimension
+            print(f.shape)
+            surf_vars[k] = f
+
+        # Shape (Lat, Lon)
+        static_vars = {}
+        for k in self.static_vars:
+            f = fields_sfc.sel(param=k).order_by(valid_datetime="ascending")
+            f =f.to_numpy(**to_numpy_kwargs)[-1]
+            f = torch.from_numpy(f)
+            print(f.shape)
+            static_vars[k] = f
+
+        # Shape (Batch, Time, Level, Lat, Lon)
+        atmos_vars = {}
+        for k in self.atmos_vars:
+            f = fields_pl.sel(param=k).order_by(valid_datetime="ascending", level=self.levels)
+            f = f.to_numpy(**to_numpy_kwargs).reshape(len(self.lagged), len(self.levels), Nj, Ni)
+            f = torch.from_numpy(f)
+            f = f.unsqueeze(0) # Add batch dimension
+            print(f.shape)
+            atmos_vars[k] = f
+
+
+        # https://microsoft.github.io/aurora/batch.html
+
+        batch = Batch(
+            surf_vars=surf_vars,
+            static_vars=static_vars,
+            atmos_vars=atmos_vars,
+            metadata=Metadata(
+                lat=torch.linspace(N, S, Nj),
+                lon=torch.linspace(W, E, Ni),
+                time=self.start_datetime,
+                atmos_levels=self.levels,
+            ),
+        )
+
+        print(batch.metadata.lat.shape)
+        print(batch.metadata.lon.shape)
+
+        with torch.inference_mode():
+
+            for pred in rollout(model, batch, steps=10):
+                print(pred.metadata.time)
+
 
 model = AuroraModel