Unbreak ISPRS Potsdam example

rastervision_core/rastervision/core/data/label_source/semantic_segmentation_label_source.py

@@ -101,3 +101,10 @@ def __getitem__(self, key: Any) -> Any:
             return self.get_label_arr(key)
         else:
             return super().__getitem__(key)
+
+    def __repr__(self):
+        arg_keys = ['raster_source', 'class_config', 'bbox']
+        arg_vals = [getattr(self, k) for k in arg_keys]
+        arg_strs = [f'{k}={v!r}' for k, v in zip(arg_keys, arg_vals)]
+        arg_str = ', '.join(arg_strs)
+        return f'{type(self).__name__}({arg_str})'

rastervision_core/rastervision/core/data/raster_source/rasterio_source.py

@@ -192,3 +192,13 @@ def __getitem__(self, key: Any) -> 'np.ndarray':
 
         chip = self.get_chip(window, bands=c, out_shape=out_shape)
         return chip
+
+    def __repr__(self):
+        arg_keys = [
+            'uris', 'channel_order', 'bbox', 'raster_transformers',
+            'allow_streaming', 'tmp_dir'
+        ]
+        arg_vals = [getattr(self, k) for k in arg_keys]
+        arg_strs = [f'{k}={v!r}' for k, v in zip(arg_keys, arg_vals)]
+        arg_str = ', '.join(arg_strs)
+        return f'{type(self).__name__}({arg_str})'

rastervision_core/rastervision/core/data/utils/misc.py

@@ -103,17 +103,41 @@ def match_bboxes(raster_source: 'RasterSource',
         label_source (LabelSource | LabelStore): Source of labels for a
             scene. Can be a ``LabelStore``.
     """
+    from rastervision.core.data import (RasterioCRSTransformer,
+                                        SemanticSegmentationLabelSource)
     crs_tf_img = raster_source.crs_transformer
     crs_tf_label = label_source.crs_transformer
     bbox_img_map = crs_tf_img.pixel_to_map(raster_source.bbox)
+
+    # For SS, if a label file is not georeferenced but is the same size as
+    # the corresponding raster source (as is the case in the Potsdam dataset),
+    # implicitly assume that they are aligned.
+    if isinstance(label_source,
+                  SemanticSegmentationLabelSource) and isinstance(
+                      crs_tf_label, RasterioCRSTransformer):
+        if crs_tf_label.image_crs is None:
+            if raster_source.extent != label_source.extent:
+                raise ValueError(
+                    f'Label source ({label_source}) is not georeferenced and '
+                    f'has a different extent ({label_source.extent}) than the '
+                    f"corresponding raster source's ({raster_source}) extent "
+                    f'({raster_source.extent}).')
+            log.warning(
+                f'Label source ({label_source}) is not georeferenced but has '
+                f'the same extent ({label_source.extent}) as the '
+                f'corresponding raster source ({raster_source}). '
+                'Will assume they are aligned.')
+            return
+
     if label_source.bbox is not None:
         bbox_label_map = crs_tf_label.pixel_to_map(label_source.bbox)
         if not bbox_img_map.intersects(bbox_label_map):
             rs_cls = type(raster_source).__name__
             ls_cls = type(label_source).__name__
-            log.warning(f'{rs_cls} bbox ({bbox_img_map}) does '
-                        f'not intersect with {ls_cls} bbox '
-                        f'({bbox_label_map}).')
+            raise ValueError(f'{rs_cls} bbox ({bbox_img_map}) does '
+                             f'not intersect with {ls_cls} bbox '
+                             f'({bbox_label_map}).')
+
     # set LabelStore bbox to RasterSource bbox
     bbox_label_pixel = crs_tf_label.map_to_pixel(bbox_img_map)
     label_source.set_bbox(bbox_label_pixel)

rastervision_pytorch_backend/rastervision/pytorch_backend/examples/semantic_segmentation/isprs_potsdam.py

@@ -44,6 +44,7 @@ def get_config(runner,
                external_model: bool = True,
                augment: bool = False,
                nochip: bool = True,
+               allow_streaming: bool = False,
                num_epochs: int = 10,
                batch_sz: int = 8,
                test: bool = False) -> SemanticSegmentationConfig:
@@ -69,6 +70,8 @@ def get_config(runner,
             training instead of from pre-generated chips. The analyze and chip
             commands should not be run, if this is set to True. Defaults to
             True.
+        allow_streaming (bool): If True, read directly from remote files
+            instead of downloading them. Defaults to False.
         num_epochs (int): Number of epochs to train for.
         batch_sz (int): Batch size.
         test (bool): If True, does the following simplifications:
@@ -142,7 +145,9 @@ def make_scene(id) -> SceneConfig:
             label_uri = label_crop_uri
 
         raster_source = RasterioSourceConfig(
-            uris=[raster_uri], channel_order=channel_order)
+            uris=[raster_uri],
+            channel_order=channel_order,
+            allow_streaming=allow_streaming)
 
         # Using with_rgb_class_map because label TIFFs have classes encoded as
         # RGB colors.
@@ -151,7 +156,8 @@ def make_scene(id) -> SceneConfig:
                 uris=[label_uri],
                 transformers=[
                     RGBClassTransformerConfig(class_config=class_config)
-                ]))
+                ],
+                allow_streaming=allow_streaming))
 
         # URI will be injected by scene config.
         # Using rgb=True because we want prediction TIFFs to be in