Add timm encoder and DINOv3 config

configs/encoder/dinov3_vitl_sat.yaml

@@ -0,0 +1,10 @@
+_target_: pangaea.encoders.timm_encoder.TimmEncoder
+timm_model_name: vit_large_patch16_dinov3.sat493m
+input_size: 512
+
+input_bands:
+  optical:
+    - B2
+    - B3
+    - B4
+

pangaea/encoders/timm_encoder.py

@@ -0,0 +1,70 @@
+from logging import Logger
+
+import torch
+import numpy as np
+import timm
+
+from pangaea.encoders.base import Encoder
+
+
+class TimmEncoder(Encoder):
+    """Timm Encoder class."""
+
+    def __init__(
+        self,
+        timm_model_name: str,
+        input_bands: dict[str, list[str]],
+        input_size: int,
+    ) -> None:
+        """Initialize the TimmEncoder.
+
+        Args:
+            model_name (str): Name of the model.
+            input_bands (dict[str, list[str]]): Input bands.
+            input_size (int): Input size.
+            embed_dim (int): Embedding dimension.
+            output_layers (list[int]): Output layers.
+            logger (Logger, optional): Logger instance. Defaults to None.
+        """
+
+        super().__init__(
+            model_name=timm_model_name,
+            input_bands=input_bands,
+            input_size=input_size,
+            embed_dim=0,  # will be set after model is created
+            output_layers=[],
+            output_dim=0,  # will be set after model is created
+            multi_temporal=False,
+            multi_temporal_output=False,
+            pyramid_output=False, # will be set after model is created
+            encoder_weights="",
+            download_url="",
+        )
+
+        try:
+            self.model = timm.create_model(
+                timm_model_name,
+                pretrained=True,
+                features_only=True,
+            )
+        except RuntimeError as e:
+            raise RuntimeError(
+                f"Error loading Timm model {timm_model_name}. "
+                "Please ensure that the timm_model_name parameter is correct and that you have internet access to download the weights."
+            ) from e
+
+        self.embed_dim = self.model.model.embed_dim
+        self.output_dim = self.model.feature_info.channels()
+        self.output_layers = list(range(len(self.output_dim)))
+        self.pyramid_output = np.unique(self.output_dim).shape[0] > 1 # pyramid if multiple unique output dims
+
+    def forward(self, image):
+        x = torch.cat(
+            [image[modality].squeeze(2) for modality in self.input_bands], dim=1
+        )
+
+        pred = self.model(x)
+        return [v for v in pred]
+
+    def load_encoder_weights(self, logger: Logger) -> None:
+        pass