Merge branch 'main' into datasets/glaciers_alps

.github/dependabot.yml

@@ -27,8 +27,6 @@ updates:
       # https://github.com/pytorch/pytorch_sphinx_theme/issues/175
       - dependency-name: 'sphinx'
         versions: '>=6'
-      # segmentation-models-pytorch pins timm, must update in unison
-      - dependency-name: 'timm'
   - package-ecosystem: 'npm'
     directory: '/'
     schedule:

docs/api/datasets/non_geo_datasets.csv

@@ -63,4 +63,4 @@ Dataset,Task,Source,License,# Samples,# Classes,Size (px),Resolution (m),Bands
 `VHR-10`_,I,"Google Earth, Vaihingen","MIT",800,10,"358--1,728",0.08--2,RGB
 `Western USA Live Fuel Moisture`_,R,"Landsat8, Sentinel-1","CC-BY-NC-ND-4.0",2615,-,-,-,-
 `xView2`_,CD,Maxar,"CC-BY-NC-SA-4.0","3,732",4,"1,024x1,024",0.8,RGB
-`ZueriCrop`_,"I, T",Sentinel-2,-,116K,48,24x24,10,MSI
+`ZueriCrop`_,"I, T",Sentinel-2,CC-BY-NC-4.0,116K,48,24x24,10,MSI

experiments/ssl4eo/flops.py

@@ -17,7 +17,7 @@
 for model in models:
     print(f'Model: {model}')
 
-    m = timm.create_model(model, num_classes=num_classes, in_chans=in_channels)
+    m = timm.create_model(model, num_classes=num_classes, in_chans=in_channels)  # type: ignore[attr-defined]
 
     # Calculate memory requirements of model
     mem_params = sum([p.nelement() * p.element_size() for p in m.parameters()])

requirements/required.txt

@@ -8,15 +8,15 @@ kornia==0.8.0
 lightly==1.5.16
 lightning[pytorch-extra]==2.5.0.post0
 matplotlib==3.10.0
-numpy==2.2.1
+numpy==2.2.2
 pandas==2.2.3
 pillow==11.1.0
 pyproj==3.7.0
 rasterio==1.4.3
 rtree==1.3.0
-segmentation-models-pytorch==0.3.4
+segmentation-models-pytorch==0.4.0
 shapely==2.0.6
-timm==0.9.7
+timm==1.0.14
 torch==2.5.1
 torchmetrics==1.6.1
 torchvision==0.20.1

tests/datasets/test_zuericrop.py

@@ -18,17 +18,11 @@
 class TestZueriCrop:
     @pytest.fixture
     def dataset(self, monkeypatch: MonkeyPatch, tmp_path: Path) -> ZueriCrop:
-        data_dir = os.path.join('tests', 'data', 'zuericrop')
-        urls = [
-            os.path.join(data_dir, 'ZueriCrop.hdf5'),
-            os.path.join(data_dir, 'labels.csv'),
-        ]
-        md5s = ['1635231df67f3d25f4f1e62c98e221a4', '5118398c7a5bbc246f5f6bb35d8d529b']
-        monkeypatch.setattr(ZueriCrop, 'urls', urls)
-        monkeypatch.setattr(ZueriCrop, 'md5s', md5s)
+        url = os.path.join('tests', 'data', 'zuericrop') + os.sep
+        monkeypatch.setattr(ZueriCrop, 'url', url)
         root = tmp_path
         transforms = nn.Identity()
-        return ZueriCrop(root=root, transforms=transforms, download=True, checksum=True)
+        return ZueriCrop(root=root, transforms=transforms, download=True)
 
     def test_getitem(self, dataset: ZueriCrop) -> None:
         x = dataset[0]

tests/models/test_resnet.py

@@ -34,7 +34,7 @@ def mocked_weights(
         load_state_dict_from_url: None,
     ) -> WeightsEnum:
         path = tmp_path / f'{weights}.pth'
-        model = timm.create_model('resnet18', in_chans=weights.meta['in_chans'])
+        model = timm.create_model('resnet18', in_chans=weights.meta['in_chans'])  # type: ignore[attr-defined]
         torch.save(model.state_dict(), path)
         try:
             monkeypatch.setattr(weights.value, 'url', str(path))
@@ -78,7 +78,7 @@ def mocked_weights(
         load_state_dict_from_url: None,
     ) -> WeightsEnum:
         path = tmp_path / f'{weights}.pth'
-        model = timm.create_model('resnet50', in_chans=weights.meta['in_chans'])
+        model = timm.create_model('resnet50', in_chans=weights.meta['in_chans'])  # type: ignore[attr-defined]
         torch.save(model.state_dict(), path)
         try:
             monkeypatch.setattr(weights.value, 'url', str(path))
@@ -122,7 +122,7 @@ def mocked_weights(
         load_state_dict_from_url: None,
     ) -> WeightsEnum:
         path = tmp_path / f'{weights}.pth'
-        model = timm.create_model('resnet152', in_chans=weights.meta['in_chans'])
+        model = timm.create_model('resnet152', in_chans=weights.meta['in_chans'])  # type: ignore[attr-defined]
         torch.save(model.state_dict(), path)
         try:
             monkeypatch.setattr(weights.value, 'url', str(path))

tests/models/test_vit.py

@@ -27,7 +27,7 @@ def mocked_weights(
         load_state_dict_from_url: None,
     ) -> WeightsEnum:
         path = tmp_path / f'{weights}.pth'
-        model = timm.create_model(
+        model = timm.create_model(  # type: ignore[attr-defined]
             weights.meta['model'], in_chans=weights.meta['in_chans']
         )
         torch.save(model.state_dict(), path)

tests/trainers/test_byol.py

@@ -89,7 +89,7 @@ def mocked_weights(
         load_state_dict_from_url: None,
     ) -> WeightsEnum:
         path = tmp_path / f'{weights}.pth'
-        model = timm.create_model(
+        model = timm.create_model(  # type: ignore[attr-defined]
             weights.meta['model'], in_chans=weights.meta['in_chans']
         )
         torch.save(model.state_dict(), path)

tests/trainers/test_classification.py

@@ -126,7 +126,7 @@ def mocked_weights(
         load_state_dict_from_url: None,
     ) -> WeightsEnum:
         path = tmp_path / f'{weights}.pth'
-        model = timm.create_model(
+        model = timm.create_model(  # type: ignore[attr-defined]
             weights.meta['model'], in_chans=weights.meta['in_chans']
         )
         torch.save(model.state_dict(), path)

tests/trainers/test_moco.py

@@ -91,7 +91,7 @@ def mocked_weights(
         load_state_dict_from_url: None,
     ) -> WeightsEnum:
         path = tmp_path / f'{weights}.pth'
-        model = timm.create_model(
+        model = timm.create_model(  # type: ignore[attr-defined]
             weights.meta['model'], in_chans=weights.meta['in_chans']
         )
         torch.save(model.state_dict(), path)

tests/trainers/test_regression.py

@@ -115,7 +115,7 @@ def mocked_weights(
         load_state_dict_from_url: None,
     ) -> WeightsEnum:
         path = tmp_path / f'{weights}.pth'
-        model = timm.create_model(
+        model = timm.create_model(  # type: ignore[attr-defined]
             weights.meta['model'], in_chans=weights.meta['in_chans']
         )
         torch.save(model.state_dict(), path)
@@ -273,7 +273,7 @@ def mocked_weights(
         load_state_dict_from_url: None,
     ) -> WeightsEnum:
         path = tmp_path / f'{weights}.pth'
-        model = timm.create_model(
+        model = timm.create_model(  # type: ignore[attr-defined]
             weights.meta['model'], in_chans=weights.meta['in_chans']
         )
         torch.save(model.state_dict(), path)

tests/trainers/test_segmentation.py

@@ -138,7 +138,7 @@ def mocked_weights(
         load_state_dict_from_url: None,
     ) -> WeightsEnum:
         path = tmp_path / f'{weights}.pth'
-        model = timm.create_model(
+        model = timm.create_model(  # type: ignore[attr-defined]
             weights.meta['model'], in_chans=weights.meta['in_chans']
         )
         torch.save(model.state_dict(), path)

tests/trainers/test_simclr.py

@@ -89,7 +89,7 @@ def mocked_weights(
         load_state_dict_from_url: None,
     ) -> WeightsEnum:
         path = tmp_path / f'{weights}.pth'
-        model = timm.create_model(
+        model = timm.create_model(  # type: ignore[attr-defined]
             weights.meta['model'], in_chans=weights.meta['in_chans']
         )
         torch.save(model.state_dict(), path)

tests/trainers/test_utils.py

@@ -34,7 +34,7 @@ def test_extract_backbone_unsupported_model(tmp_path: Path) -> None:
 
 
 def test_get_input_layer_name_and_module() -> None:
-    key, module = _get_input_layer_name_and_module(timm.create_model('resnet18'))
+    key, module = _get_input_layer_name_and_module(timm.create_model('resnet18'))  # type: ignore[attr-defined]
     assert key == 'conv1'
     assert isinstance(module, nn.Conv2d)
     assert module.in_channels == 3

torchgeo/datasets/zuericrop.py

@@ -52,10 +52,7 @@ class ZueriCrop(NonGeoDataset):
        * `h5py <https://pypi.org/project/h5py/>`_ to load the dataset
     """
 
-    urls = (
-        'https://polybox.ethz.ch/index.php/s/uXfdr2AcXE3QNB6/download',
-        'https://raw.githubusercontent.com/0zgur0/multi-stage-convSTAR-network/fa92b5b3cb77f5171c5c3be740cd6e6395cc29b6/labels.csv',
-    )
+    url = 'https://hf.co/datasets/torchgeo/zuericrop/resolve/8ac0f416fbaab032d8670cc55f984b9f079e86b2/'
     md5s = ('1635231df67f3d25f4f1e62c98e221a4', '5118398c7a5bbc246f5f6bb35d8d529b')
     filenames = ('ZueriCrop.hdf5', 'labels.csv')
 
@@ -221,11 +218,11 @@ def _verify(self) -> None:
 
     def _download(self) -> None:
         """Download the dataset."""
-        for url, filename, md5 in zip(self.urls, self.filenames, self.md5s):
+        for filename, md5 in zip(self.filenames, self.md5s):
             filepath = os.path.join(self.root, filename)
             if not os.path.exists(filepath):
                 download_url(
-                    url,
+                    self.url + filename,
                     self.root,
                     filename=filename,
                     md5=md5 if self.checksum else None,

torchgeo/models/resnet.py

@@ -768,7 +768,7 @@ def resnet18(
     if weights:
         kwargs['in_chans'] = weights.meta['in_chans']
 
-    model: ResNet = timm.create_model('resnet18', *args, **kwargs)
+    model: ResNet = timm.create_model('resnet18', *args, **kwargs)  # type: ignore[attr-defined]
 
     if weights:
         missing_keys, unexpected_keys = model.load_state_dict(
@@ -803,7 +803,7 @@ def resnet50(
     if weights:
         kwargs['in_chans'] = weights.meta['in_chans']
 
-    model: ResNet = timm.create_model('resnet50', *args, **kwargs)
+    model: ResNet = timm.create_model('resnet50', *args, **kwargs)  # type: ignore[attr-defined]
 
     if weights:
         missing_keys, unexpected_keys = model.load_state_dict(
@@ -837,7 +837,7 @@ def resnet152(
     if weights:
         kwargs['in_chans'] = weights.meta['in_chans']
 
-    model: ResNet = timm.create_model('resnet152', *args, **kwargs)
+    model: ResNet = timm.create_model('resnet152', *args, **kwargs)  # type: ignore[attr-defined]
 
     if weights:
         missing_keys, unexpected_keys = model.load_state_dict(

torchgeo/models/scale_mae.py

@@ -92,7 +92,7 @@ def get_1d_sincos_pos_embed_from_grid_torch(embed_dim: int, pos: Tensor) -> Tens
     return emb
 
 
-class ScaleMAE(VisionTransformer):  # type: ignore[misc]
+class ScaleMAE(VisionTransformer):
     """Custom Vision Transformer for Scale-MAE with GSD positional embeddings.
 
     This is a ViT encoder only model of the Scale-MAE architecture with GSD positional embeddings.
@@ -117,7 +117,8 @@ def __init__(self, res: float = 1.0, *args: Any, **kwargs: Any) -> None:
         self.res = res
 
         # Scale MAE uses resolution specific positional embeddings
-        self.pos_embed.requires_grad = False
+        if self.pos_embed is not None:
+            self.pos_embed.requires_grad = False
 
     def _pos_embed(self, x: Tensor) -> Tensor:
         """Apply GSD positional embeddings to the input tensor."""
@@ -133,8 +134,9 @@ def _pos_embed(self, x: Tensor) -> Tensor:
             .to(x.dtype)
             .to(x.device)
         )
-        cls_tokens = self.cls_token.expand(x.shape[0], -1, -1)
-        x = torch.cat((cls_tokens, x), dim=1)
+        if self.cls_token is not None:
+            cls_tokens = self.cls_token.expand(x.shape[0], -1, -1)
+            x = torch.cat((cls_tokens, x), dim=1)
         x = x + pos_embed
         x = self.pos_drop(x)
         return x
@@ -155,7 +157,9 @@ def interpolate_pos_embed(
     pos_embed_checkpoint = state_dict['pos_embed']
     embedding_size = pos_embed_checkpoint.shape[-1]
     num_patches = model.patch_embed.num_patches
-    num_extra_tokens = model.pos_embed.shape[-2] - num_patches
+    num_extra_tokens = 0
+    if model.pos_embed is not None:
+        num_extra_tokens = model.pos_embed.shape[-2] - num_patches
     # height (== width) for the checkpoint position embedding
     orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)
     # height (== width) for the new position embedding

torchgeo/models/vit.py

@@ -243,7 +243,7 @@ def vit_small_patch16_224(
     if weights:
         kwargs['in_chans'] = weights.meta['in_chans']
 
-    model: VisionTransformer = timm.create_model(
+    model: VisionTransformer = timm.create_model(  # type: ignore[attr-defined]
         'vit_small_patch16_224', *args, **kwargs
     )
 

torchgeo/trainers/byol.py

@@ -332,7 +332,7 @@ def configure_models(self) -> None:
         in_channels: int = self.hparams['in_channels']
 
         # Create backbone
-        backbone = timm.create_model(
+        backbone = timm.create_model(  # type: ignore[attr-defined]
             self.hparams['model'], in_chans=in_channels, pretrained=weights is True
         )
 

torchgeo/trainers/classification.py

@@ -80,7 +80,7 @@ def configure_models(self) -> None:
         weights = self.weights
 
         # Create model
-        self.model = timm.create_model(
+        self.model = timm.create_model(  # type: ignore[attr-defined]
             self.hparams['model'],
             num_classes=self.hparams['num_classes'],
             in_chans=self.hparams['in_channels'],

torchgeo/trainers/moco.py

@@ -238,10 +238,10 @@ def configure_models(self) -> None:
         output_dim: int = self.hparams['output_dim']
 
         # Create backbone
-        self.backbone = timm.create_model(
+        self.backbone = timm.create_model(  # type: ignore[attr-defined]
             model, in_chans=in_channels, num_classes=0, pretrained=weights is True
         )
-        self.backbone_momentum = timm.create_model(
+        self.backbone_momentum = timm.create_model(  # type: ignore[attr-defined]
             model, in_chans=in_channels, num_classes=0, pretrained=weights is True
         )
         deactivate_requires_grad(self.backbone_momentum)

torchgeo/trainers/regression.py

@@ -83,7 +83,7 @@ def configure_models(self) -> None:
         """Initialize the model."""
         # Create model
         weights = self.weights
-        self.model = timm.create_model(
+        self.model = timm.create_model(  # type: ignore[attr-defined]
             self.hparams['model'],
             num_classes=self.hparams['num_outputs'],
             in_chans=self.hparams['in_channels'],

torchgeo/trainers/simclr.py

@@ -153,7 +153,7 @@ def configure_models(self) -> None:
         weights = self.weights
 
         # Create backbone
-        self.backbone = timm.create_model(
+        self.backbone = timm.create_model(  # type: ignore[attr-defined]
             self.hparams['model'],
             in_chans=self.hparams['in_channels'],
             num_classes=0,