Autoformatting

sources/unipercept/engine/_engine.py

@@ -873,7 +873,9 @@ def _inference_loop(
         # Output memory
         if results_path is None:
             results_remove_on_exit = True
-            results_path = file_io.Path(f"//scratch/{self._params.project_name}/{self._params.session_name}/{prefix}-results.h5")
+            results_path = file_io.Path(
+                f"//scratch/{self._params.project_name}/{self._params.session_name}/{prefix}-results.h5"
+            )
         else:
             results_remove_on_exit = False
         results_mem = PersistentTensordictWriter(str(results_path), samples_total)
@@ -971,7 +973,6 @@ def _inference_loop(
             if not key.startswith(f"{prefix}/"):
                 metrics[f"{prefix}/{key}"] = metrics.pop(key)
 
-
         return metrics, samples_processed
 
     def _store_visualizations(self, visuals: dict[str, pil_image.Image], prefix: str = "") -> None:

sources/unipercept/engine/writer.py

@@ -55,7 +55,6 @@ def tensordict(self) -> TensorDictBase:
         raise NotImplementedError("Abstract property `tensordict` not implemented.")
 
 
-
 class PersistentTensordictWriter(ResultsWriter):
     """
     Writes results to a H5 file using PersistentTensorDict from multiple processes, uses a buffer to reduce the number of writes.
@@ -90,17 +89,22 @@ def __del__(self):
             self.close()
 
     @staticmethod
-    def error_when_closed(fn: T.Callable[T.Concatenate[PersistentTensordictWriter, _P], _R]) -> T.Callable[T.Concatenate[PersistentTensordictWriter, _P], _R]:
+    def error_when_closed(
+        fn: T.Callable[T.Concatenate[PersistentTensordictWriter, _P], _R]
+    ) -> T.Callable[T.Concatenate[PersistentTensordictWriter, _P], _R]:
         @functools.wraps(fn)
         def wrapper(self, *args: _P.args, **kwargs: _P.kwargs) -> _R:
             if self._is_closed:
                 raise RuntimeError(f"{self.__class__.__name__} is closed")
             return fn(self, *args, **kwargs)
+
         return wrapper
 
-    @functools.cached_property 
+    @functools.cached_property
     def path(self) -> file_io.Path:
-        return file_io.Path(self._path)
+        p = file_io.Path(self._path)
+        p.parent.mkdir(parents=True, exist_ok=True)
+        return p
 
     def __len__(self) -> int:
         return self._size
@@ -175,4 +179,3 @@ def tensordict(self) -> TensorDictBase:
         if self._td is None:
             self._td = self._td_factory(filename=self.path)
         return self._td
-

sources/unipercept/nn/layers/merge.py

@@ -34,23 +34,22 @@ class SemanticMerge(nn.Module):
 
     def __init__(
         self,
-        in_features: T.Iterable[str],
-        input_shape: T.Mapping[str, BackboneFeatureInfo],
+        in_features: T.Mapping[str, BackboneFeatureInfo],
         common_stride: int,
         out_channels: int,
         weight_method: WeightMethod | str = WeightMethod.FAST_ATTENTION,
     ):
         super().__init__()
 
-        self.in_features = list(in_features)
+        self.in_features = list(in_features.keys())
         self.common_stride = int(common_stride)
 
         if isinstance(weight_method, str):
             weight_method = WeightMethod(weight_method)
         self.weight_method = weight_method
 
-        feature_strides = {k: T.cast(int, v.stride) for k, v in input_shape.items()}
-        feature_channels = {k: T.cast(int, v.channels) for k, v in input_shape.items()}
+        feature_strides = {k: T.cast(int, v.stride) for k, v in in_features.items()}
+        feature_channels = {k: T.cast(int, v.channels) for k, v in in_features.items()}
 
         self.scale_heads = nn.ModuleList()
         for in_feature in self.in_features:

sources/unipercept/render/_plot.py

@@ -24,8 +24,7 @@
     from unipercept.data.tensors import PanopticMap
     from unipercept.model import InputData, ModelOutput
 
-__all__ = ["plot_input_data", "plot_predictions", 
-           "draw_image", "draw_image_segmentation", "draw_image_depth"]
+__all__ = ["plot_input_data", "plot_predictions", "draw_image", "draw_image_segmentation", "draw_image_depth"]
 
 from unicore.utils.missing import MissingValue