add support for treating background as an object

add option to avoid predicting all 0s for affs and lsds in background.
janelia-cellmap · Feb 14, 2024 · ee3f7cc · ee3f7cc
1 parent 39915ac
commit ee3f7cc
Show file tree

Hide file tree

Showing 3 changed files with 21 additions and 3 deletions.
diff --git a/dacapo/experiments/tasks/affinities_task.py b/dacapo/experiments/tasks/affinities_task.py
@@ -18,6 +18,7 @@ def __init__(self, task_config):
             affs_weight_clipmax=task_config.affs_weight_clipmax,
             lsd_weight_clipmin=task_config.lsd_weight_clipmin,
             lsd_weight_clipmax=task_config.lsd_weight_clipmax,
+            background_as_object=task_config.background_as_object,
         )
         self.loss = AffinitiesLoss(
             len(task_config.neighborhood), task_config.lsds_to_affs_weight_ratio

diff --git a/dacapo/experiments/tasks/affinities_task_config.py b/dacapo/experiments/tasks/affinities_task_config.py
@@ -52,3 +52,13 @@ class AffinitiesTaskConfig(TaskConfig):
         default=0.95,
         metadata={"help_text": "The maximum value for lsds weights."},
     )
+    background_as_object: bool = attr.ib(
+        default=False,
+        metadata={
+            "help_text": (
+                "Whether to treat the background as a separate object. "
+                "If set to false background should get an affinity near 0. If "
+                "set to true, the background should also have high affinity with other background."
+            )
+        },
+    )
diff --git a/dacapo/experiments/tasks/predictors/affinities_predictor.py b/dacapo/experiments/tasks/predictors/affinities_predictor.py
@@ -28,6 +28,7 @@ def __init__(
         affs_weight_clipmax: float = 0.95,
         lsd_weight_clipmin: float = 0.05,
         lsd_weight_clipmax: float = 0.95,
+        background_as_object: bool = False,
     ):
         self.neighborhood = neighborhood
         self.lsds = lsds
@@ -51,6 +52,8 @@ def __init__(
         self.lsd_weight_clipmin = lsd_weight_clipmin
         self.lsd_weight_clipmax = lsd_weight_clipmax
 
+        self.background_as_object = background_as_object
+
     def extractor(self, voxel_size):
         if self._extractor is None:
             self._extractor = LsdExtractor(
@@ -105,10 +108,12 @@ def create_target(self, gt):
             label_data = label_data[0]
         else:
             axes = ["c"] + axes
-        affinities = seg_to_affgraph(label_data, self.neighborhood).astype(np.float32)
+        affinities = seg_to_affgraph(
+            label_data + int(self.background_as_object), self.neighborhood
+        ).astype(np.float32)
         if self.lsds:
             descriptors = self.extractor(gt.voxel_size).get_descriptors(
-                segmentation=label_data,
+                segmentation=label_data + int(self.background_as_object),
                 voxel_size=gt.voxel_size,
             )
             return NumpyArray.from_np_array(
@@ -208,7 +213,9 @@ def gt_region_for_roi(self, target_spec):
                     for a, b in zip(pad_pos, self.lsd_pad(target_spec.voxel_size))
                 ]
             )
-        gt_spec.roi = gt_spec.roi.grow(pad_neg, pad_pos)
+        gt_spec.roi = gt_spec.roi.grow(pad_neg, pad_pos).snap_to_grid(
+            target_spec.voxel_size
+        )
         gt_spec.dtype = None
         return gt_spec