Seems to work. Implemented SemanticMapper in DronescapesRepresentation. Not fully generic, but good enough.

data/train_set/.task_statistics.npz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:eed73f43285542ff4d3c0781cb7ba1eedb1ca48b7f28f323bfdecd7c32ec46f6
-size 26947
+oid sha256:8f601e71980200cacfc43a1b3ed5d7d932b538b8cf94d569d74911a0d3604b54
+size 28542

dronescapes_reader/dronescapes_representations.py

@@ -94,12 +94,16 @@ class SemanticRepresentation(NpzRepresentation):
         return new_images
 
 class SemanticMapper(SemanticRepresentation):
-    """Maps one or more semantic segmentations to a final one + a merge fn. Copy-pasta from VRE"""
+    """
+    Maps one or more semantic segmentations to a final one + a merge fn. Copy-pasta from VRE.
+    TODO: allow non-semantic dependencies (must pass the correct load_fn or object instead of assuming semantic deps)
+    """
     def __init__(self, *args, original_classes: list[list[str]], mapping: list[dict[str, list[str]]],
                  color_map: list[tuple[int, int, int]],
                  merge_fn: Callable[[list[np.ndarray]], np.ndarray] | None = None, **kwargs):
         super().__init__(*args, classes=list(mapping[0].keys()), color_map=color_map, **kwargs)
-        assert len(self.dependencies) >= 1, "No dependencies provided. Need at least one semantic segmentation to map."
+        assert len(self.dependencies) >= 1 and self.dependencies[0] != self.name, \
+            "No dependencies provided. Need at least one semantic segmentation to map."
         assert isinstance(mapping, list), type(mapping)
         assert len(mapping) == (B := len(self.dependencies)), (len(mapping), B)
         assert (A := len(original_classes)) == len(self.dependencies), (A, B)
@@ -233,8 +237,10 @@ tasks = [ # some pre-baked representations
     NormalsRepresentation("normals_sfm_manual202204"),
     OpticalFlowRepresentation("opticalflow_rife"),
     SemanticRepresentation("semantic_segprop8", classes=8, color_map=color_map_8classes),
-    SemanticMapper("semantic_mask2former_swin_mapillary_converted", original_classes=[mapillary_classes],
-                   mapping=[m2f_mapillary_to_8_classes], color_map=color_map_8classes),
+    SemanticRepresentation("semantic_mask2former_swin_mapillary_converted", classes=8, color_map=color_map_8classes),
+    SemanticMapper("semantic_mask2former_swin_mapillary_converted2", original_classes=[mapillary_classes],
+                   mapping=[m2f_mapillary_to_8_classes], color_map=color_map_8classes,
+                   dependencies=["semantic_mask2former_mapillary_49189528_0"]),
     SemanticRepresentation("semantic_mask2former_coco_47429163_0", classes=coco_classes, color_map=coco_color_map),
     SemanticRepresentation("semantic_mask2former_mapillary_49189528_0", classes=mapillary_classes,
                             color_map=mapillary_color_map),

dronescapes_reader/multitask_dataset.py

@@ -107,7 +107,10 @@ class MultiTaskDataset(Dataset):
             self._statistics = None if normalization is None else self._compute_statistics()
         if self._statistics is not None:
             for task_name, task in self.name_to_task.items():
-                task.set_normalization(self.normalization[task_name], self._statistics[task_name])
+                try:
+                    task.set_normalization(self.normalization[task_name], self._statistics[task_name])
+                except:
+                    breakpoint()
 
     # Public methods and properties
 

scripts/dronescapes_viewer.py

@@ -7,6 +7,7 @@ from dronescapes_reader.dronescapes_representations import dronescapes_task_type
 from pprint import pprint
 from torch.utils.data import DataLoader
 import random
+import numpy as np
 
 def main():
     assert len(sys.argv) == 2, f"Usage ./dronescapes_viewer.py /path/to/dataset"
@@ -23,6 +24,10 @@ def main():
     data, name, repr_names = reader[rand_ix] # get a random item
     pprint({k: v for k, v in data.items()})
 
+    img_data = {}
+    for k, v in data.items():
+        img_data[k] = reader.name_to_task[k].plot_fn(v) if v is not None else np.zeros((*reader.data_shape[k][0:2], 3))
+
     print("== Random loaded batch ==")
     batch_data, name, repr_names = reader[rand_ix: min(len(reader), rand_ix + 5)] # get a random batch
     pprint({k: v for k, v in batch_data.items()}) # Nones are converted to 0s automagically